Professional Documents
Culture Documents
Machine learning is, at its simplest level, the process whereby feeding data to a program
does not yield more data. It yields a better program. Among other implications, First-
Mover Advantage is no longer a marketing strategy. It is an R&D imperative.
Not since the mid-1990s and the Information Superhighway has the collection of a few
vague phrases and acronyms (AI, VR, AR, and deep neural learning) inspired such
conviction about a profoundly different future. We consider winners and losers.
The stocks that our analysts identify as the key beneficiaries within their coverage are:
Sunny Optical, Sony, SMIC, Fanuc, Bumrungrad, Samsung SDI, and BYD.
SEE DISCLOSURE APPENDIX OF THIS REPORT FOR IMPORTANT DISCLOSURES AND ANALYST CERTIFICATIONS
BERNSTEIN
Machine learning is at its simplest level the process whereby instead of feeding data
to a program that then generates more data, we feed data to a program that generates
another, better program. And it also works outside of the video arcade. The first two
technological breakthroughs that machine learning has facilitated are speech recognition
and image recognition at lower-than-human error rates. These triumphs along with the
falling cost of memory and the continuing improvements in computing power are
generating rapid change in a near-endless list of industries and fields of endeavor, from
autos to factory automation to healthcare.
In this Blackbook, our Asia Strategist and six of our Asia-focused technology (in the widest
sense) analysts consider the implications for their sector and assess investment
opportunity. We consider whether the flow of value over the last 20 years globally and
within Asia (from technology manufacturers to the platform) has any lessons for machine
learning and its catholic applications in the next few years. This Blackbook also reviews
the latest developments in technology around energy storage and its implications for
renewable energy, the grid, and fossil fuels. Our goal is to set out how to think about these
emerging technologies and how to pick winners and losers.
TABLE OF CONTENTS
Oh, those simple times. In the last two years, the focus has started to shift from the
extension of consumer electronics trends that began at the end of the last decade to
something that is still struggling for a name, or perhaps a name with any meaning.
"Artificial intelligence (AI)" and "deep neural learning" give the impression of technology-
driven dystopias involving a Singularity, or perhaps worse, competing Singularities. In this
Blackbook, we use the term "machine learning." However, as will become apparent in
chapter after chapter, what is relevant is the learning far more than the machine.
In 1999, the observation that the Internet doubled in size every 100 days was repeated so
often that it may have even been true. We are at a similar moment in terms of machine
learning right now. Given the rate of change, our approach in this Blackbook is necessarily
mosaic rather than exhaustive. We discuss the theoretical underpinnings of the
technology and survey its most promising applications (the "The Fourth Wave of IT
Hardware (Redux)" chapter by Alberto Moel); dive into the Cambrian explosion that
machine learning has triggered through improvements in image recognition (the "The Gift
of Vision: Deep Learning for Computer Vision" chapter by David Dai); take one of the many
applications of image recognition Automated Driving Assistance Systems (ADAS)
down to the processor level (the "ADAS: The Path to Truly Autonomous Driving" chapter
by Mark Li); and consider the raft of new technologies and applications machine learning
is enabling in healthcare (the "12 Game-Changing Technologies in the Future of
Healthcare" chapter by Laura Nelson Carney).
But it is not AI all the time. We also offer perspectives on an entirely different future of
tech with grid-connected, utility-scale energy storage (the "The Future of Energy Storage"
chapter by Mark Newman). The Chinese power grid is the world's largest machine, and
while proclamations of the rising intelligence of that particular machine may have been
overstated in recent years, the new generation of storage technologies is certainly going
to make the power grid all power grids more efficient.
Finally, we consider who wins with all of this technology and whether the implications are
simply too profound to remain either unregulated and/or in private hands (the "Does
Value Go to the Technology or the Platform" chapter by Bhavtosh Vajpayee and Michael
Parker).
The technology sector is the most profitable in Asia (see Exhibit 1). Roughly a dozen
companies account for 75% of those profits (see Exhibit 2). Advances in machine
learning, in our view, mean that this first condition is likely to continue. But a remarkably
common theme in our research at least among most of the contributors to this
Blackbook is that scale and "silicon sophistication" is less important than the years of
data that is required to "train" and optimize software/algorithms. Machines learn, but they
learn slowly, and through practice and repetition. Scale may matter less to the future of
tech than being early and being tenacious.
EXHIBIT 1: ROIC by sector and select industries (MSCI Asia EXHIBIT 2: Technology sector share of operating profits
ex-Japan)
Internet 44.2%
IT 21.6% Samsung,
Semis 20.2% 26%
Tech Hardware 16.7%
Health Care 12.0%
Electr. Equip. 11.6%
Financials 10.3%
Staples 10.1%
Telecoms 9.9% LG
Discretionary 7.1% Display,
2%
Utilities 7.0% Baidu,
Automobiles 6.7% 2%
Energy 5.8% Infosys,
3% TSMC,
Industrials 5.5% TCS, 11%
Materials 4.8% 5%
Real Estate 3.8% SK Hynix, Tencent,
5% 6%
0.0% 10.0% 20.0% 30.0% 40.0% 50.0% Alibaba, Hon Hai,
6% 6%
Source: Bloomberg L.P. and Bernstein analysis. Source: Bloomberg L.P. and Bernstein analysis.
INVERTING THE An agent (or a program, a robot, or even a human) is learning if it improves its performance
COMPUTATIONAL PARADIGM in matching the algorithm to the data after making repeated observations of samples of
(WHAT YOU MAY HAVE MISSED
the data. The key characteristic of machine learning is the capability to "learn from" in
SO FAR) (BY ALBERTO MOEL)
other words, machine-learning algorithms can take data and work their way backward to
create a new algorithm or modify the existing algorithm (themselves). Traditional
computation doesn't "learn from" anything. It takes input data, processes it using a static
program (basically a set of rules specifying how the input data is to be manipulated), and
generates output data; end of story.
The problem with traditional computation is that the real world is too complicated and
unpredictable for a static set of rules to emulate it. This rule-based approach either fails to
capture all possible outcomes (in the jargon, "had bad exception handling"), or becomes
fiendishly complex trying to capture all the possible states of the world to be modeled.
Machine learning, instead of trying to capture all the complexity and unpredictability of the
real world, emulates human behavior and embraces this complexity and unpredictability
(see Exhibit 3).
With machine learning, instead of feeding input data to a program to generate more data
as the output, we feed data to a program that generates another "program." We can think
of machine learning as the "inverse" of programming, just like the square root is the
"inverse" of the squaring function, or integration and differentiation are inverses of each
other.
In traditional programming, we can feed the number 4 to a squaring program, and get 16
as the output, or feed x2+x to a symbolic differentiator and get 2x+1 as the output. In
machine learning, we feed {16,4} to the algorithm and get back f(x) = x2, or { x2+x, 2x+1}
and the algorithm returns f(x) = df/dx.
Some of the more promising and interesting machine-learning algorithms are worth
describing in more detail:
Gesture control. Gesture control is the ability to recognize human body movements to
interact with and control a computer system without direct physical contact. Classic,
existing, widely available systems are Nintendo's Wii and Microsoft's Kinect.
DEEP LEARNING FOR MACHINE The most dramatic example of the power of machine learning in application over the last
VISION (BY DAVID DAI) five years may be image recognition. Computer vision (the "gift of vision") has made
significant progress since 2012. Each year, artificial intelligence scientists globally
participate in the ImageNet Classification competition, where computers are given a
number of pictures and asked to identify objects in them. The best program in 2010
misclassified 28% of the pictures, and the best result in the subsequent year was 26%. In
2012, deep learning was applied in the competition for the first time, and it immediately
brought down the error rate to 16%, and to 12% in the next year. By 2016, the error rate
dropped to merely 3.1%, lower even than the human error rate of 5.1% (see Exhibit 4).
25% 25.8%
Started to use deep learning
20%
16.4%
15%
11.7%
10%
7.3%
Human Error Rate: 5.1% 3.6%
5%
3.1%
0%
2010 2011 2012 2013 2014 2015 2016
The hero behind the progress in computer vision is deep learning. In the last five years,
deep learning has made great progress in image recognition, speech recognition, and
mainly to attract publicity games like Go and the old Atari 2600 classics from the early
1980s. The regulatory and ethical implications of machines that learn are important and
complex. But the investment implications are at present and for once far simpler. We
are going to need a lot more cameras.
One industry that will likely benefit from the growth of computer vision is the camera
supply chain. All smartphones today already have cameras equipped to take pictures and
selfies. Going forward, they will likely be equipped with additional cameras that enable
them to recognize faces, understand gestures, scan environments, or turn any apartment
into an augmented reality (AR) game playfield. While most of the smartphones today have
two cameras, we expect eight cameras in the future, increasing the addressable market
four times. We forecast handset camera shipment to grow 66% in the next five years,
driven by additional cameras for face recognition, iris scanning, gesture sensing, and
augmented reality/virtual reality (AR/VR).
With image recognition, vehicles today can see other vehicles, pedestrians, traffic signs,
and road markings. We expect automotive camera shipment to grow 145% in the next
five years, driven by growing ADAS penetration and the increasing number of cameras
required with heightened ADAS functionality.
VR and AR devices will likely have multiple cameras installed so that they can track the
environment and head position. This technology, called inside-out tracking, is already
used in Microsoft HoloLens, Intel's Project Alloy, and Google Tango. It cuts the
cumbersome setup process of external trackers that are currently required by VR systems
such as HTC Vive. Similar technologies also apply to drones and robots, which use
cameras for path finding and obstacle avoidance.
Surveillance cameras also benefit from AI to make our homes and organizations safer.
Surveillance systems that include video analytics analyze video footage in real time and
detect abnormal activities that could pose a security threat. Essentially, video analytics
technology helps security software to "learn" what is normal so it can identify unusual, and
potentially harmful, behavior.
In our coverage, Sunny Optical (outperform), Sony (outperform), and Largan (market-
perform) are in the camera supply chain. Among them, we like Sunny Optical for its #1
position in vehicle lens and fast-growing #2 position in handset camera lens. We also like
Sony for being #1 in CMOS image sensor.
ADAS: THE PATH TO TRULY Of course, Automated Driving Assistance Systems (ADAS) require not just cameras but
AUTONOMOUS DRIVING (BY semiconductors.
MARK LI)
The automotive market is becoming increasingly important for semiconductor companies.
ADAS is gaining broader adoption. The safety benefits of ADAS are being recognized by
safety regulators. As the focus of motor vehicle safety regulation shifts from occupant
protection and accident mitigation to active accident avoidance, we can expect to see
more government regulation mandating these ADAS features on new vehicles. As a result,
the ADAS semiconductor market is growing faster than other parts of the automotive
semiconductor market (see Exhibit 5). With an 18% CAGR, the ADAS semiconductor
revenue is expected to more than double, from US$1.6 billion in 2015 to US$3.8 billion in
2020.
ADAS semiconductors can be broadly classified into processors and sensors. The
processors are essential to process data and make decisions, and work together with
sensors, which gather information on the surrounding environment such as pedestrians
and oncoming traffic. The combination of both types of semiconductors enables the
ADAS system to sense, think, and act to aid drivers and prevent accidents. Initially, we
thought the ADAS processors would be built with cutting-edge wafer manufacturing
technologies (i.e., nodes) as the processors need to recognize objects and traffic
conditions and determine how to steer the car quickly. Then, contrary to our expectation,
the emerging market leader, Mobileye (MBLY, not covered, being acquired by Intel now),
can meet the requirement with mature nodes.
40 ADAS 18%
35 Body 8%
30
Chassis 7%
25 EV/HEV 16%
US$B
20 Infotainment 7%
15
Powertrain 4%
10
Safety 7%
5 Aftermarket 1%
Instrument 7%
0
2014 2015 2016 2017E 2018E 2019E 2020E
Instrument Cluster Aftermarket Safety Powertrain Infotainment EV/HEV Chassis Body ADAS
Further, the valuable knowhow within ADAS is mostly in software/algorithms and years of
data that is used to train and optimize the software/algorithms. The silicon actually
doesn't need to be that advanced.
Computer vision is highly important for ADAS and a very demanding task. For example, a
2-megapixel camera running at 60 frames per second will generate about 0.5 gigabyte of
data every second. Together with multiple cameras to capture images from different
angles, the amount of data is massive. Further, the data must be processed in real time to
steer cars away from dangers. Given these factors, our first reaction was that computer
vision would need expensive cutting-edge manufacturing technologies that probably only
giants like Intel and Qualcomm can afford. However, much to our surprise, our research
finds that Mobileye and other small companies are using special IC design architecture
and algorithm and are able to offer highly competitive products.
EXHIBIT 6: Mobileye's chips are smaller and more efficient than NVIDIA's
NVIDIA Mobileye
Tegra K1 Tegra X1 DRIVE PX 2 EyeQ3
(28nm) (20nm) (16nm) (40nm)
Peak Performance (GFLOPS) 326 1024 256
Actual Performance (GFLOPS) 76 287 205
2
Die Size (mm ) 132 126 481? 42
2
Actual Performance/Area (GFLOPS/mm ) 0.6 2.3 4.9
Power Efficiency (GFLOPS/W) ~9 29 ~82
Source: Luca Benini, Integrated Systems laboratory of ETH Zurich, Universit di Bologna, and Bernstein estimates and analysis.
This is because CPUs and SoCs from Intel and Qualcomm are designed to run different
software for a myriad of tasks and, hence, are "general-purpose" chips. Chips from
Mobileye and other small companies, however, are designed to accomplish a few specific
tasks with tailored software. They also exploit the repetitive nature of the data and have
many small but efficient "cores" to process data in parallel. As such, we find that
Mobileye's chips are built with more mature nodes with smaller dies. The chips are, hence,
cheaper, but the optimized architecture and algorithm enable them to deliver remarkable
performance and power efficiency (see Exhibit 6). This explains Mobileye's leading
position in the vision-based ADAS processor market.
In fact, larger companies also recognize the advantages of specialized architecture and
are moving in this direction. For example, Qualcomm's "zeroth" initiative is embedding a
"neutral process unit" (NPU) into its latest SoCs for the "cognitive computing" requirement
of automotive, wearables, mobile, and other applications. NVIDIA is also banking on its
GPUs that have many cores to process the image data in parallel. Freescale (merged with
NXPI, which is being acquired by Qualcomm) acquired CogniVue, a Canada-based IP
company, to jump-start its vision-processing capability. Intel announced plans to acquire
the computer vision chip company Movidius in September 2016, and only a few months
later, in March 2017, made another announcement of plans to acquire the ADAS
processor market leader Mobileye, in an effort to put itself ahead of competition in the
ADAS and the autonomous driving market.
THE FUTURE OF AUGMENTED Asia's healthcare systems face the same challenges that plague developed markets (e.g.,
REALITY IN MEDICINE (BY costs spiraling up, misaligned incentives across stakeholders, and unpaid debts), plus
LAURA NELSON CARNEY)
some unique and arguably more difficult emerging market obstacles to healthcare
modernization (e.g., wider income and health inequality, low health spend per capita,
inadequate infrastructure, endemic corruption, and broken public trust).
Technology can potentially solve some of these problems better than old approaches. We
think that technology may change the face of healthcare faster in Asia than elsewhere in
the world innovations in how care is administered, independent of innovations in
healthcare treatments (i.e., specific therapeutics new drugs, devices, and treatment
paradigms).
Note: Lower left-corner image is Bernstein Asia-Pacific Healthcare analyst wearing Toshiba Vital Images virtual reality goggles (development collaboration with
Microsoft HoloLens technology for surgical and medical education applications).
In this Blackbook, we discuss 12 emerging technologies (see Exhibit 7) that we think will
be transformative in healthcare in Asia and elsewhere artificial intelligence; genomics-
enabled care (precision medicine); gene editing; telemedicine; augmented reality; the
Internet of Things (IOT) monitoring and diagnostics; blockchain; smart pills; connected
communities; automation and robotics; fully digitized hospitals; and gamification of health.
We discuss why and where the pace of change toward widespread implementation of
these technologies throughout the health system may be faster in Asia (AI, genomics,
telemedicine, and connected communities) and highlight emerging companies to watch.
Artificial intelligence is the creation of learning, intelligent machines that can work and
react like humans in certain circumstances. This means programming computers to have
key human traits, including (but not limited to) knowledge, reasoning, problem-solving,
perception, learning, planning, and the ability to move and manipulate objects. The
potential implications of AI are profound across all aspects of our lives (including
healthcare) AI is arguably the most important human innovation of the century (in
general, not just in medicine). It isn't fair to categorize it as a discrete technology changing
healthcare, mutually exclusive of the other 11 discussed in this Blackbook. In fact, AI is
being incorporated into smarter uses of most of the other 11 technologies, particularly
telemedicine, genomics, IOT monitoring and diagnostics, automation and robotics, and
connected communities.
We think that disruptive new technology may be the only way to move the needle
meaningfully toward solving healthcare problems in China without dramatically increasing
healthcare expenditure or the number of doctors. Most of the problems dogging
healthcare in China relate to inefficiency, low quality of "products" and "services"
(medicines and care), asymmetry of information, corruption, and inequitable access to
care. Technology is not a panacea, but has solved similar challenges in other fields.
THE OTHER FUTURE OF TECH: Much of this Blackbook delves into a range of machine learning applications from
ENERGY STORAGE AND FLOW artificial intelligence, to autonomous driving, to augmented reality. Of course, the future of
BATTERIES (BY MARK
tech is not all about machine learning. Although the main growth market attracting
NEWMAN)
attention for batteries and energy storage is electric vehicles, there is another huge
potential market in energy storage systems (ESSs). In this Blackbook, we discuss the need
for energy storage to enable renewable energy and improve the efficiency of the world's
biggest machine, the machine that powers all machinesthe power grid.
EXHIBIT 8: Global ESS market size by technology EXHIBIT 9: of which various types of flow batteries play an
important part
18
6
16
4.9
14
Total new installations bn USD
12
4
10
3.0
GWh
8 3
6 2.1
2
4 1.4
1.0
1 0.7
2 0.5
0 0.1 0.1 0.2
0 0
Source: CairnERA, and Bernstein estimates (2017 and beyond) and analysis. Source: CairnERA, and Bernstein estimates (2017 and beyond) and analysis.
These systems allow electricity to be stored for future use upon production, and are
receiving increasing attention, given the implications for renewable energy sources, such
as wind and solar power, which are clean, but largely volatile in terms of output. ESSs can
help improve utilization, efficiency, and reliability of the electric grid and drive adoption of
renewables by enabling time control over the volatile energy supply from these sources.
Although the dominant energy storage technology for both EVs and ESSs will likely
remain lithium-ion for some time (see Exhibit 8), flow batteries (see Exhibit 9) is another
technology that has profound advantages versus alternatives. In particular, flow batteries
last longer (20 years) and are able to scale energy (duration) independent of power
requirement, thus making them cost competitive for high-energy/long-duration projects.
There are two main flow battery technologies Zinc Bromide Redox (ZBR) and Vanadium
Redox (VRB). We think VRB holds the most promise among flow batteries and we see a
very clear role for these batteries in the future. The key advantage of this technology is the
ability to scale capacity independent of power (and, thus, cheaply).
For VRB, self-discharge is not a problem because the electrolytes are stored in separate
tanks and its life is the longest among all battery ESSs. One of the largest VRB systems
that is currently under testing is the 15MW/60MWh project built by Japan's Hokkaido
Electric Power (HEPCO, ticker YTO:5802) and SEI. The two parties want to test the
performance of the system as a new way to adjust wind and solar intermittency.
For ZBR battery, the extremely corrosive nature of the bromine electrolyte will cause
degradation and potential failure, though the active materials themselves do not degrade.
Thus, the lifetime of the ZBR battery is not very dependent on the number of cycles or the
depth of discharge, but on the duration that the system has been under operation. Despite
the technology still being at an early stage, there are dozens of projects operational in
Australia, France, and the United States. Currently, the biggest operated project is in
France, conducted by ZBB and Pacific Beachcomber. The project uses ZBB's 2MWh
systems (40x 10kW/5kWh modules) for a luxury eco-resort. Apart from these completed
projects, there are an increasing number of large-scale projects that have been
announced.
WILL THE VALUE ACCRUE TO Over the last two decades, the three most compelling and far-reaching changes in
THE TECH OR THE PLATFORM? technology globally have been: the increased speed of telecommunications, the falling
(BY BHAVTOSH VAJPAYEE AND
cost of memory semiconductors, and the improved computing power of processors.
MICHAEL PARKER)
And yet, when the chapter is closed on the key technologies of the first two decades of
the 21st century, the headlines will be (roughly in order): iPhone, Facebook, iPod, WeChat,
Amazon, BABA, iPad, Instagram, Tinder, and perhaps Uberplus Samsung's Galaxy Note
7. The technology companies invented the future. The telecommunications industry built
it. And the Internet sector (the customer-facing platform) plus Apple captured most
of the value (see Exhibit 10).
The characteristics that allowed the Internet sector to capture most of the value over the
last two decades were a combination of low or zero subscriber acquisition cost and low
capital intensity. Those traits can be found in various pockets among the future of tech
companies. But there is a broader question, in our view. Blame it on karma. Technology is
the new establishment. It is front and center, the mainstream of societal change and
influence, the platform of abuse and angst, the weapon law makers and breakers wish to
possess, and the force ordinary humans are increasingly wary of. Vaulting ambition,
unrelenting innovation, ease of financing, and the rising sophistication of computing and
software have brought technology to a point where the disruption is societal in a way that
will be very different from the past.
This is no longer about PCs, smartphones, or tablets mere toys that enabled us to
connect and explore our world differently. This is about how we will live our lives, what
kinds of jobs we hope to keep and lose (see Exhibit 11 and Exhibit 12), how we will
educate and train our children for the new world, how we will fight crime, and how
geopolitics will be shaped. This is about how and how long we shall live and how we shall
approach old age. A simple ratio to define the intensity of disruption is this: number of
people better off/number of people worse off or displaced. We shall call it the net benefit
ratio. The higher this ratio, the more likely it is that the disruption will be embraced by
society, regulators, and investors. The lower this ratio, the more likely it is that new
sources of friction appear. And we are returning to a level of disruption globally that has
not been seen in the United States or Europe (outside of military conflict) since the
Industrial Revolution.
EXHIBIT 10: Market capitalization across technology and technology-adjacent sectors (2000-16)
2,000 Internet
1,800
1,600
1,400
1,200
USD B
Telcos
1,000
800
Apple
600
Processors
400 / foundries
Memory
200
Telecom Equipment
0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
The question around where value accrues from innovation rests on the premise that
ownership of its downstream benefits, powers, and influences remains unchanged. For
our generation, which grew through the Reagan-Thatcher privatization era, the end of the
Communism, and the great globalization experiment of the World Trade Organization, it is
hard to believe that the fruits and weapons of innovation have not always been in private
hands. And where technology has reached today and where it threatens to go in the future
may be leading us to revisit this debate. In part, this now looks inevitable.
This train of thought, therefore, leads us to the following conclusion. Sure, many parts of
the tech world benefited disproportionately from the infrastructure created by others
Internet giants stood on the shoulders of semiconductors, telecom equipment, and
service providers. This equation, which appears "unfair" to many as of now, could well
reverse in the future. The billions that many technology companies will likely spend in
space exploration, healthcare/gene therapy, and artificial intelligence will eventually lead
to products that will likely not exclusively belong to them.
The companies that appropriated the profits of others in the last two decades will likely
have their profits appropriated from them. In this alternate future, much of the current
R&D spend in some areas (AI, space, and genetics) is merely a hidden penalty for past
appropriations, as the benefits of these R&D outlays are unlikely to be allowed to remain
in private hands.
EXHIBIT 11: Percentage of jobs at risk from automation EXHIBIT 12: Probability an AI will replace this job
Telemarketer
UK
Accountant, auditor
Retail salesperson
US
Taxi drivers
Real estate sales agent
OECD average
Word processor, typist
Machinist
S. Africa
Commercial pilot
Economist
Argentina
Health technologist
Actor
Nigeria
Firefighter
Editor
India
Chemical engineer
Clergy
Thailand
Athletic trainer
Dentist
China
Recreational therapist
0% 20% 40% 60% 80% 100% 0.0% 40.0% 80.0% 120.0%
Source: World Bank Development Report, 2016, and Bernstein analysis. Source: "The Future of Employment", C. Frey and M Osborne (2013), and
Bernstein analysis.
VALUATION METHODOLOGIES See the "Valuation Methodology and Risks" chapter of this Blackbook.
RISKS See the "Valuation Methodology and Risks" chapter of this Blackbook.
INVESTMENT IMPLICATIONS See the "Investment Implications" sections at the end of the chapters of this Blackbook.
OVERVIEW
A couple of years ago, we published a framework to think about the next big thing in
IT hardware the Fourth Wave of IT hardware which we issued as a manifesto
(The Long View: The Fourth Wave of IT Hardware).
Our theory was (and remains) that we are now entering a Fourth Wave of IT
hardware, where computing is being transferred to the edges of the network in
smaller and smaller devices with increasing sensing, computing, storage, and
algorithmic intelligence capabilities. The possibilities, enabled by Moore's Law, are
almost infinite.
The transition into the Fourth Wave is an ongoing evolution characterized by:
(a) exponentially increasing computing power, storage density, and network
bandwidth coupled with the deflationary nature of this change; (b) the continued shift
from analog to digital; and (c) recombinant innovation in form factors, user
applications, and end markets, enabled by these performance improvement trends.
In particular, the Fourth Wave is enabling artificial intelligence (AI) and machine
learning applications to become a reality, and to bring intelligence to hardware
platforms.
We identify six major (relatively distinct) application segments of the Fourth Wave:
Human-Machine Interfaces, the Quantified Self, Smart Machines, Smart Distributed
Systems, the Connected Home, and Factory Automation.
Using our framework, we extended our coverage into factory automation (Global
Automation: From a Cyclical to a Secular Growth Story); and the bulk of our new
research has been in understanding the impact of the Fourth Wave on automation
and generating investable ideas within that sector, neglecting the broader trends.
But technology advancement continues unabated more generally in IT hardware.
In particular, since the publication of our manifesto, two trends have become more
apparent, even though they had already been flagged at the time the exponential
impact of technology in expanding the space of the possible, and the emergence of
artificial intelligence (and, in particular, machine learning) as perhaps the next
general-purpose technology (GPT). The purpose of this chapter is to address these
trends and provide a much-needed update.
As Niels Bohr is purported to have said, prediction is difficult, especially about the future.
But we are, at different time horizons, in the business of making predictions. Some of
these predictions are more short term (and not that interesting, really) as in what's the
quarter looking like? Others are more Bernstein-ish, and are at the heart of our everyday
research what's the endgame on PCs or smartphones, and who are the winners or
losers? Or when is the secular automation trend going to dominate the cyclical?
And often, we are asked to make the mother of all predictions in our industry what's the
next big thing in IT hardware? Sometimes, the question comes from frustration,
exasperation, or even boredom with a highly competitive, deflationary, and difficult-to-
invest-in sector. Other times, it emerges from a sense of curiosity as to what else could
materialize and upend the status quo in consumer and industrial electronics. And often,
it's more of an opportunistic question, wondering what we think the next big thing is going
to be (before anybody else does) and, hence, be in a position to make lots of money.
Regardless of the genesis of the question, what's the next big thing in IT hardware is fair
game for the Asian IT Hardware franchise. We came to grips early with this question as we
realized a lot of what our original coverage involved PCs, TVs, smartphones was
actually already mature and yesterday's news. So, over time, we've developed a
framework to think about the next big thing in IT hardware the Fourth Wave of IT
Hardware which we published as a manifesto over two years ago (The Long View: The
Fourth Wave of IT Hardware).
Using that framework, we've extended our coverage into factory automation, one of the
areas where we had predicted (luckily, correctly) that the next big thing in IT hardware
would have a sustained and material impact (Global Automation: From a Cyclical to a
Secular Growth Story).
The bulk of our new research has been in understanding the impact of the Fourth Wave on
automation and generating investable ideas within that sector, and so we've been
neglectful of the broader trends. But technology advancement continues unabated more
generally in IT hardware.
In particular, since the publication of our manifesto, two trends have become more
apparent, even though they had already been flagged at the time the exponential
impact of technology in expanding the space of the possible, and the emergence of
artificial intelligence (and, in particular, machine learning) as perhaps the next general-
purpose technology.
EXPANDING THE SPACE OF THE "Things only get crazy in the second half of the chessboard" Ray Kurzweil
POSSIBLE AND THE SECOND
HALF OF THE CHESSBOARD Starting with the earliest electronic computing devices, we see a transition from a highly
concentrated computing intelligence (the First Wave mainframes) to a localized,
standalone model (the Second Wave for example, PCs and electronics in everyday
items, such as automobiles), to a distributed model (the Third Wave, where we are now
the Internet and mobile devices), heading back to "mainframe" computing in a wide
network (the cloud).
This evolution continues and we are now entering the Fourth Wave, where computing is
being transferred to the edges of the network in smaller and smaller devices with
increasing sensing, computing, and storage capabilities.
The transition into the Fourth Wave isn't occurring through some structural discontinuity
in electronic hardware innovation. It's primarily an ongoing evolution characterized by
exponentially increasing computing power, storage density, and network bandwidth
coupled with the deflationary nature of this change; the continued shift from analog to
digital; and recombinant innovation in form factors, user applications, and end markets,
enabled by these performance improvement trends.
These sensing, processing, and output building blocks can be mixed and matched to
create a quasi-infinite number of combinations and end-user applications. An analog to
this exponential-combinatorial explosion is the parable of the invention of chess.
As legend has it, the inventor of chess introduced the game to an emperor in India in the
6th century CE. The emperor was so impressed with the game that he told the inventor to
ask for whatever he wanted. The inventor was very humble and only asked for a grain of
rice in the first square, two in the second square, four in the third square, and so on, all
through the 64 squares of the board. By the time they had reached the second half of the
board (the 33rd square), the grain count was already up to 232 or 4,294,976,296 grains of
rice, equivalent to 124 tons of rice, about enough to feed a family of four in modern India
for 413 years.
You get the idea; and as inventor-futurist Ray Kurzweil1 said, "things only get crazy in the
second half of the chessboard." A little more relevant than grains of rice on a chessboard,
McAfee and Brynjolfsson2 calculate that if Moore's Law started in 1958, and the doubling
period was every 18 months, we entered the second half of the chessboard with digital
progress around 2006. Where this exponential-combinatorial trend takes us remains an
open question, but we can say the impact is just beginning to be felt.
ARTIFICIAL INTELLIGENCE: THE "The impact of AI on tech is like the impact of electricity was on industry" Andrew Ng,
NEXT GENERAL-PURPOSE recently, of Baidu
TECHNOLOGY?
The most important invention of the 18th century was the steam engine, which powered
the First Industrial Revolution. The 19th century gave us two (not just one) very important
inventions electric light and power, and the internal combustion engine, which powered
the Second Industrial Revolution. What separate these three inventions from everything
1
Ray Kurzweil is one of the most interesting, creative, and infuriating people out there. Kind of an early Elon Musk, without
the car company (https://en.wikipedia.org/wiki/Ray_Kurzweil).
2
"The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies," Erik Brynjolfsson and
Andrew McAfee.
else invented since is that they can be described as general-purpose technologies (GPTs),
which themselves had multiple applications and spawned other important inventions.
For example, inventions generated by electricity as a GPT include electric machine tools,
elevators, electric streetcars, and subways; the full range of consumer appliances, from
electric irons, vacuum cleaners, to washing machines, and refrigerators to air
conditioners. Similarly, the invention of the internal combustion engine led to automobiles,
trucks, buses, and taxis; and onward to supermarkets, suburbs, motels, and of course air
travel.
We are now in what is variously called the "Information Age" or the "Post-Industrial Era,"
among other equally florid names. In the First Industrial Revolution, humans harnessed
steam power, and in the Second Industrial Revolution, it was electricity and internal
combustion. Whatever you want to call this era, it's not a stretch that this time around the
forces that are driving innovation aren't steam, electricity, or internal combustion. It is
information and its acquisition, harnessing, manipulation, and application.
And the Fourth Wave is leading us to increasingly more powerful, complex, and intelligent
modes of information processing, many of which fall under the rubric of the buzzword du
jour, artificial intelligence.3 We can imagine the continuum of intelligence from the
reflexive (feel an itch, then scratch) to the truly sublime (theoretical physics and poetry).
And the Fourth Wave is allowing machines to "possess" increasingly higher levels of
intelligence (see Exhibit 13).
3
We aren't going into details here on what AI is and how it works, as the focus of this chapter is on applications. But you can
read our first contribution to the debate: Weekend Tech Byte: Thinking Machines.
We are probably still in the early innings of this AI capability building, but the applications
of AI to an increasing number of information processing activities continue to broaden
(see Exhibit 14). Given the range of applications of AI to information acquisition,
processing, and utilization, could we not consider AI as the GPT of the 21st century?
Knowledge Autonomous
Human Interfaces Complex Analysis Personalization
Acquisition Control
Web search Speech recognition Autonomous driving Image recognition Product recommendations
Machine translation Handwriting recognition Automatic composition 3D scene reconstruction Ad placement
Information extraction Question answering Household robotics Human activity recognition Spam filtering
Document summarization Spelling correction Robotic surgery Social network analysis Fraud detection
Music information retrieval AI for video games Robot exploration Smart grid optimization Spellchecking
Medical imaging Gesture controls Delivery drones Fault diagnostics Financial planning
Augmented reality Character animation
Financial trading
Dynamic pricing
Protein folding
Medical diagnosis
Climate modeling
WHERE IS THE FOURTH WAVE More practically, we identify six major application segments where the Fourth Wave of IT
TAKING US? hardware, in conjunction with the AI GPT, is going to be an enabler of future technologies,
with our (best) estimates for market size and growth (see Exhibit 15). These are Human-
Machine Interfaces, Quantified Self, Smart Machines, Smart Distributed Systems,
Connected Home, and Factory Automation.4
Of these six, Factory Automation and Smart Distributed Systems (for example, the smart
grid) have by far the largest markets (in the tens to hundreds of billions of U.S. dollars).
However, because of their current size, they have lower growth rates (in the below 10%
range) than other more speculative applications that have captured investors'
imagination.
4
We hope these labels are relatively self-descriptive, as we delve a bit deeper into each of these later in this chapter.
EXHIBIT 15: The market opportunity in the Fourth Wave is both large and growing rapidly
Human-Machine
Interfaces
2014-2020 Market Size CAGR
50%
Connected Home
Smart Machines
Smart Distributed
Systems
Factory Automation
Quantified Self
5%
1 10 100
2016 Market Size (USD Billions)
Source: Gartner, IDC, IFR, IEK, IHS, and Bernstein estimates (2017 and beyond) and analysis.
All six of these identified applications rely on sensors and transducers to interact with their
environment, be it the physical world (physical sensors), humans (human-centered
sensing), or other machines (as in Factory Automation or Smart Distributed Systems such
as the smart grid). These sensors and transducers have their own companies and
technologies, which are part of the Fourth Wave IT hardware supply chain. The sensor and
transducer market size is hard to quantify, but we're certain it is big enough to matter.
The Second Wave, a localized, standalone model where minicomputers and then
microcomputers (PCs) providing isolated computation and storage "democratized"
the computing experience (late 1970s through late 1990s); into
The Third Wave, a distributed model where these isolated computing devices were
networked (first through wired and then through wireless means), further reduced in
size, and made highly portable (notebook PCs to mobile devices and tablets). The
introduction of cost-effective flat panel displays enabled portability and form factor
innovation to these new devices. Additionally, part of the computing power is
redistributed through the network and concentrated back into a redux mainframe in
the form of the datacenter and the cloud (late 1990s to present).5
This evolution continues and as the Third Wave enters maturity, we move to the Fourth
Wave, where computing is further devolved to the edges of the network in increasingly
smaller and novel devices with increasing sensing, computing, storage, and intelligence
capabilities.6 This computing and storage intelligence doesn't have to be purely at the
device level, it can also be distributed among other devices in the "cloud."
Effectively, the increasing computational and storage capabilities are enabling smaller
and increasingly powerful hardware combinations with exponentially increasing
application possibilities (see Exhibit 16).
EXHIBIT 16: The four long waves of IT hardware, 1950s to the present
5
Before we get all googly-eyed about the cloud, let us note that the concept of a "virtual machine" has been around since the
dinosaur computing era (for example, MULTICS, http://web.mit.edu/multics-history/), and even IBM's heavy-iron
mainframes in the 1970s and 1980s were known as the "VM" series, for, as you guessed, "virtual machine" (see
http://en.wikipedia.org/wiki/VM_(operating_system) for an interesting and lively description of VM OS). Even the idea of
"storing" content in the "cloud" is well-established: all of the web-mail services (e.g., Hotmail and Gmail) and the picture-
sharing sites (e.g., Flickr and Picasa) store the content in the cloud.
6
Some people like to refer to this Fourth Wave as the "Internet of Things," which is as badly descriptive of what it means as
the use of "cloud" to describe using somebody else's insecure computer and storage. Besides, we didn't come up with the
term, so we will avoid it.
WHAT ARE THE DRIVERS AND The transition into the Fourth Wave isn't occurring through some structural discontinuity
ENABLERS OF THIS FOURTH in electronic hardware innovation. It's primarily an ongoing evolution characterized by:
WAVE OF IT HARDWARE? WHAT
ARE THE IMPLICATIONS?
Exponentially increasing computing power, storage density, and network bandwidth
coupled with the deflationary nature of this change;
Recombinant innovation in form factors, user applications, and end markets, enabled
by these performance improvement trends.
Moore's Law still (more or less) holds. As a result, the performance of the three basic
elements of Third Wave devices computational capability, digital storage, and network
connectivity continues to improve (see Exhibit 17).
What's more, this increasing computing power and storage density is accompanied by a
deflationary spiral in their costs. The key pieces of hardware where computation and
storage "happen" solid state memories (both DRAM and Flash), HDDs, and
microprocessors have consistently exhibited very fast cost reduction rates. A fourth
one, not a processing or storage key component but a basic building block for human
input-output displays has also gotten exponentially cheaper over the last decade (not
to mention truly portable with the advent of the flat panel display).
1.00E+15
1.00E+12
Supercomputer Speed
(FLOPS)
1.00E+09
Supercomputer Energy
Efficiency (FLOPS/Watt)
1.00E+03 Microprocessor
Transistors/Chip HDD Cost Efficiency
(GB/USD)
1.00E+00
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
Source: McAfee and Brynjolfsson "The Second Machine Age," and Bernstein estimates (2014-16) and analysis.
Exponentially decreasing storage costs. Commodity DRAM and NAND Flash memory
prices per bit have dropped dramatically. Since the 1970s, commodity DRAM prices have
dropped from approximately US$80,000 per Mb to around well under 1 cent/Mb (see
Exhibit 18). Similarly, the price per bit for standard MLC NAND has dropped from over
US$1,000 per GB since introduction to below US$1.00 per GB (see Exhibit 19).
EXHIBIT 18: DRAM pricing has declined exponentially EXHIBIT 19: Flash memory pricing has also declined
exponentially
$100,000.00 $81,920 $10 ,00 0.0 0
$10,000.00
$100.00
USD/Mb
$10 0.0 0
USD/GB
$10.00
$1.00
$10 .00
$0.10
$0.01 $1.00
12,000,000 1
$0.01
$0.00
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014
2016
$0.10
200 0
200 1
200 2
200 3
200 4
200 5
200 6
200 7
200 8
200 9
201 0
201 1
201 2
201 3
201 4
201 5
201 6
Source: DramXchange and Bernstein's Global Memory team. Source: DramXchange and Bernstein's Global Memory team.
The price per Mb of HDD storage has also fallen exponentially since the advent of the first
consumer HDD in the early 1980s. In the intervening period, this price has dropped from
about US$315/Mb to less than 1/20,000th of a cent per Mb, a six order of magnitude
difference (see Exhibit 20).
The average retail price (excluding taxes) of an external HDD for PC applications has
declined from about US$3,200 in 1981 to under US$100 in 2016, which is one order of
magnitude. Of course, you felt blessed with a 20 Mb HDD in 1983, while contemporary
HDDs comfortably reach the 1+ TB (terabyte, or 1 million MB) capacities.
1000
100
10
1
USD/Mb
0.1
0.01
0.001
0.0001
1E-05
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
Source: Ivan Smith, (http://ns1758.ca/winch/winchest.html) and Bernstein analysis.
EXHIBIT 21: And, of course, the cost of processing power has also dropped exponentially
0.1
USD per Transistor
0.01
0.001
0.0001
0.00001
0.000001
0.0000001
1971
1972
1974
1978
1979
1982
1985
1989
1993
1994
1995
1997
1998
1999
2000
2000
2001
2002
2003
2004
2004
2006
2006
2006
2006
2008
2009
2010
2011
2012
2013
2014
2015
2016
Source: Intel and Bernstein analysis.
Rapidly declining display input-output costs. TFT-LCD displays are a relatively new key
component and a major driver of the growth in mobile computing and communications
products. Exhibit 22 shows the evolution of average areal (per square meter) ASPs,
COGS, and cash costs. Looking at average areal costs allows us to review the cost curve
independently of specific manufacturers or products. In aggregate, the ASP (and cost) per
unit area of a flat panel display in 2016 is less than 1/20th of that of the equivalent display
in 2001, with far superior performance specifications in every dimension.
Although not shown, ASPs of other input-output devices, such as sensors, transducers,
and actuators have also dropped quasi-exponentially, and as sensors and transducers "go
digital," they are now also following the inexorable deflation implied by Moore's Law.
EXHIBIT 22: Although not subject to Moore's Law, the cost of TFT-LCD panels has also dropped very rapidly
6,000
5,000
4,000
USD per m2
3,000
2,000
1,000
The shift from analog to digital. This computational evolution is also driving a revolution in
how we acquire, use, store, and transmit information. In 1986, only 6% of the world's data
was digital (think about those dark days vinyl records, tape decks, film, printed books!).
Today, more than 99% of the world's written words, images, music, and physical data are
captured, created, collected, and transmitted digitally.
More profoundly, sensors like microphones, cameras, and accelerometers have moved
from the analog world to the digital one. They have become computer chips, able to
generate, store, and compute digitally, and become subject to the exponential
improvement trajectories of Moore's Law.
All that information can be digitally collected, processed, and analyzed for patterns and
implications using distributed computing networks in ways not possible previously ("big
data" and "the cloud," for those so inclined to using the buzzwords).
A combinatorial explosion and multiplicity of applications. Now that the building blocks of
the Fourth Wave are all digital, these systems can be mixed and matched from a set of
components to create a quasi-infinite number of combinations and end-user applications.
This combinatorial explosion has a number of massive implications for the future of IT
hardware:
specific applications, but it will be difficult to know a priori which applications and
business models will be successful even the ultimate industry structure is
uncertain.
As we move from analog to digital electronics, products became more modular, with
tremendous implications on eventual winners and losers. For example, in the past,
Japanese companies made everything from CRTs to the phosphors of an analog TV
screen in-house; this was a source of competitive advantage. Nowadays, modern
digital sets can be assembled using a TFT-LCD display and components from a
range of suppliers. A retailer such as Amazon can now make the Kindle book reader
because it need not design the components itself and it can outsource assembly. The
fall of the storied Japanese consumer electronics concerns is probably not
coincidental to this structural change.
And lastly, a committed small team anywhere in the world with a fresh insight in a
single area of an application can mix and match relatively cheap hardware and
software (including emerging AI software development kits like TensorFlow7) to put
together a system without being deep experts in any of the underlying technologies.
Lengthy training and priestly engineering vocations will not be needed to innovate.
Even people with limited skills but with a conceptual idea that meets a consumer or
business need can give it a go and maybe even succeed innovation in IT hardware
becomes democratized.
What's the downside? Luddites, Nervous Nellies, and Cassandras will always point out
that these kinds of technological upheavals bring societal change and unintended
consequences. The Robotics Business Review provides an example that we nowadays
take automatic passenger elevators for granted: "In 1958, a wave rose and crested over
Manhattan where 200,000 people went to work each day as elevator operators. Within 10
years, nearly all had lost their jobs to automation. Today, no one scans the "help wanted"
adverts for employment as an elevator operator. Robots now perform those jobs: tirelessly
whisking millions up and down buildings everywhere, some up to 700 feet per minute, and
even warning exiting passengers to watch their step."
A number of contemporary intellectuals have warned of the pitfalls of this Fourth Wave.
McAfee and Brynjolfsson8 write of the impact of the Fourth Wave on employment and
income inequality, but remain on the optimistic side. Others, such as Evgeny Morozov,9
raise concerns about privacy, and the reliance on technology to "solve" the world's
problems. Nicholas Carr10 brings up the issue of automation and the Internet making us
(even) dumber than we already are, and the doomsday scenario of the machines getting
far too smart and doing more harm than good. But the likely positive long-term
implications of the Fourth Wave of IT hardware are mind-boggling, and we don't use that
term lightly. We'll deal with the negative implications in future research.
7
https://www.tensorflow.org/
8
"The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies," Erik Brynjolfsson and
Andrew McAfee.
9
"The Net Delusion: The Dark Side of Internet Freedom," and "To Save Everything, Click Here," both written by Evgeny
Morozov.
10
"The Glass Cage: Automation and Us" by Nicholas Carr.
Zhou Enlai was asked about the impact of the French Revolution. "Too early to say" was his
reply.
The analog to digital transition, deflationary trends, and the combinatorial explosion
generate a virtually limitless space of possibilities for the Fourth Wave. The applications
are quasi-infinite, and the state of play reflects a period of fervent innovation,
experimentation, and trial-and-error. That is to be expected as in the early stages of the
s-curve of a technology transformation, the field is wide open, industry structure is still in
flux, and barriers to entry are low.
But even now, it is possible to sketch out the potential possibilities and applications of the
Fourth Wave. At the very least, it would allow investors to classify and quantify
opportunities as they appear and evolve into developed markets. To start with, we can
identify four major classes of outcomes, to give us a sense of the range of possibilities:
Humans and machines working alongside each other for example, a mobile robot
working with a warehouse employee to move boxes, or a collaborative robot in
manufacturing; and
Machines working with each other for example, a factory where automated
processes work with each other and reconfigure automatically, or self-driving
vehicles interacting with other self-driving vehicles.
In terms of applications, we have done an in-depth analysis of what's already out there,
what's talked about, and what the industry (and investors) are excited about. The
possibilities are endless, and as Zhou Enlai would say in this day and age, probably too
early to tell what will work and what won't.
Regardless, we identify six major (relatively distinct) application segments, where the
Fourth Wave of IT hardware is going to be an enabler of future technologies (see Exhibit
23):
Quantified Self monitoring health and fitness and improving physical condition
through technology. Quantified Self is the use of self-monitoring through a wide
variety of sensors and devices. Applications use mobile apps and wearable or
portable devices to collect data about a user's activities, biometrics, environment,
and other personal experiences. Analysis of these data allows individuals to gain a
better understanding of their physical condition and improve their wellbeing.
Applications include home health monitoring, mobile sports and fitness, and wearable
user interfaces.
Smart Distributed Systems connecting small intelligent devices and harnessing the
power of the network. In contrast to Smart Machines, Smart Distributed Systems are
incorporated into an array of hardware and communication devices that interact with
each other intelligently to carry out highly sophisticated functions. Examples of Smart
Distributed Systems are telematics (in all of its forms), remote monitoring and device
management, smart grids, and smart transportation.
modes of manufacturing, such as collaborative robots are evolving rapidly, and have
been the focus of much of our research over the past two years.
EXHIBIT 23: There are six major areas where the Fourth Wave will have an impact
Human-Machine Interfaces Quantified Self Smart Machines
1 2 3
Augmented Reality
Affective Computing Home Health Monitoring UAVs (Drones)
Human Augmentation Mobile Fitness Autonomous Vehicles
Virtual Reality Mobile Health Monitoring Smart Fabrics
Tangible User Interfaces Wearables Smart Dust
Ambient Displays
HOW BIG IS EACH FOURTH Market-sizing the Fourth Wave opportunities is no easy feat. The business opportunities
WAVE PIE, AND WHERE IS THE are so broad, novel, and speculative that it's hard to estimate how big these businesses
GROWTH POTENTIAL?
currently are and how fast they'll be growing. Questions about user acceptance,
technology roadmaps, and even regulatory barriers all play a role in determining the
potential market size and growth. But we take a stab at it in Exhibit 24.
EXHIBIT 24: The market opportunity in the Fourth Wave is both large and growing rapidly
Source: Gartner, IDC, IFR, IEK, IHS, and Bernstein estimates (2017 and beyond) and analysis.
Overall, it is pretty clear to us that there is an inverse relationship between the size of the
existing market and its future growth potential. This is consistent with an s-curve model of
technology diffusion, which we have successfully applied to determine the dynamics in
the PC market (The Long View: The Future(s) of the PC (2015 Edition) What If the
Installed Base Is Actually Shrinking?) and the ongoing secular trend in automation (Global
Automation: From a Cyclical to a Secular Growth Story).
The biggest market opportunity, by far, is Factory Automation, of which we have written
extensively, and will continue to do so. But it is also the slowest growing overall, with year-
over-year revenue growth in the high-single digits. On the other hand, because of its
relative size and maturity, there are many investment opportunities, with lower technology
and customer acceptance risks.
Another market that is also relatively large and is growing at a double-digit clip is Smart
Distributed Systems, which includes RFID, telematics, and the smart grid. Although these
distributed end-markets require a systems approach and there is regulatory risk, the
applications are compelling and even disruptive, for example in auto insurance (U.S.
Insurance: The Next Generation of Mobile: A Deep Dive on Emerging Telematics
Technology).
Quantified Self, which includes wearables, has captured investor's imaginations, but
because of the consumer end-user focus, where ASP is an important driver of unit
growth, we do not expect the business to grow more than high-single-digits on revenue
terms, even if unit shipments are strong. Further, the consumer-facing nature of the
product portfolio means that product-cycle risk is very high and likely to follow the classic
downward profit slope of traditional consumer electronics (Weekend Tech Byte: The
Internet of Nothing).
The wide variations in potential of the different market opportunities means that we need
to dig a bit deeper into the specifics of each subsegment, and at least identify the players
in the space for further analysis. We also need a framework for investment, given all the
uncertainty present in the Fourth Wave.
What do the characteristics of the Fourth Wave mean for industry profitability, structure,
and investability? The following are our guiding principles that flow from this technology
transition:
The combinatorial possibilities are quasi-infinite and nobody has a "lead" in what
needs to be done. The process is very democratic. Put it another way, the hardware
and software combination thematic is whitespace it is wide open and there isn't an
incumbent to dislodge. End-market growth for many segments is very strong, which
is likely to attract many players, most of which will not succeed.
The technology roadmaps are uncertain and very broad, and the possible technology
end-states are ill-defined. Nobody really knows how things will turn out, and what will
work and what won't. As the number of building blocks and possible combinations
explodes, it will be hard to know which ones will be valuable and which ones will be a
waste of time. The likelihood of the Fourth Wave generating "excess capacity" is high,
with lots of dead ends and wasted efforts.
In the shift from analog to digital electronics, products become more modular. Modern
digital systems can be mixed and matched from a set of components. Because of this
modularity, a small (not necessarily brilliant or accomplished) team with a fresh
insight in a single area of the Fourth Wave can mix and match relatively cheap
hardware to put together an automation system without being deep experts in the
technologies involved.
Open source is the name of the game. A lot of the AI overlays used in the Fourth Wave
are all coming out of research efforts and laboratories that are open source, with the
tools and techniques widely disseminated. Replication is not easy, but is achievable
by a committed individual or group of individuals willing to spend the time with the
code and the application. Sustainable profitability of these open source business
models is always more elusive.
SUSTAINABILITY OF PROFITS: The "fuzzy" and emerging nature of the Fourth Wave as an investment opportunity is
MANEUVERABILITY, POSITION enough to give us a headache. But we propose three simple questions that quickly allow
IN THE CHAIN, AND CONTROL OF
us to determine sustainability of profits (and hence investability) is the company nimble
THE KEY COMPONENTS
and maneuverable, does it hold a privileged position in the chain, and does it have control
of the key components? If the answer is yes to all three, we likely have a winner.
Company needs to be nimble and maneuverable. Because of all the product, technology,
industry structure, and consumer acceptance uncertainty, the ability to turn on a dime,
maneuverability, and quick reaction times to respond to this uncertainty are fundamental.
In other words, we need to focus on companies with the capabilities to nimbly innovate at
the micro level, keeping ahead of the opportunities without losing the momentum or the
technology edge. Pivoting, in the parlance of Silicon Valley, is to be expected.
Evidence of this nimbleness is the high rate of product introduction, high (and effective)
R&D expenditures, and ability to consistently "see" the future better, and maybe even
shape it. Although there is evidence that R&D expenditures sometimes align with
structural overvaluations (Weekend Tech Byte: Our Two Cents About Investing in Tech,
Part I: When I Hear the Word Innovation... I Press the Short), we find that they are
predictors of excess profits (with the corresponding higher valuations).
Company needs to hold a privileged position in the IT hardware supply chain. As the IT
hardware industry evolved, a handful of segments have surfaced as clear winners that
have been able to capture and sustain high levels of economic profit, have commanded
valuation premiums above other segments, and are in general better investment
opportunities.
Company needs to hold the key to the key components. Key components are those high
value-added "pieces" in IT hardware that make up most of the value in the finished device.
In a notebook PC or tablet, these could include the CPU, GPU, HDD, Flash memory, and
DRAM. In a notebook PC, 80% of the BOM is key components.
The key components generally carry higher margin and profitability, and can capture an
overwhelming majority of the profit pool available.11 In the Fourth Wave of IT hardware, in
addition to the CPU, memory, and display key components, we can include the sensing
and transducing elements in that group, such as optical and acoustic components.
Outside of the hardware itself, we can consider key components to also include the
(ideally) proprietary algorithms (or their implementation) that provide intelligence to the
hardware platform.
Having control of these elements are likely to generate high margins as the knowledge
embedded in the algorithms can be ported to multiple hardware platforms.
"Forget the smartphone, the sensor will be the most widely shipped device" John Chen,
Blackberry CEO
Conceptually, the Fourth Wave of IT hardware is no different from that of its earlier
incarnations. All computing systems require three basic modules an input module,
which takes signals and data from humans, the environment, or other computing devices;
a processing module, which stores data, uses algorithms to analyze, and interpret the
data; and generate signals from the system through the output module.
Every computing device past, present, and future has and will have these modules and
this sensing-processing-output loop. What is changing and enabling the Fourth Wave is
the performance, complexity, and types of modules.
Input modules have evolved from punch cards to keyboards on to touch screens (and now
on to wireless radios and widely available ambient sensors); computing power and
algorithms are improving exponentially, while the output modes multiply from just
11
Among OEMs, the "key components" refer to operating systems, ecosystems, and network effects that are unique to
specific OEMs and allow them to remain ahead of hardware and operating system commoditization.
EXHIBIT 25: The building blocks of the Fourth Wave are the same as in the past Von Neumann and Turing are still
relevant
From Algorithms To
Physical signals Digital storage Physical actuation
Human input Human interfaces
Other computing systems Other computing systems
Feedback loops
As technology advances, it enables smaller and smaller units of hardware with internal or
distributed "intelligence" at the edges of the network. This intelligence requires smaller
and more responsive input-output systems, whose mode of action and functionality is
increasing, while the algorithms become increasingly sophisticated.
Input and output modules include sensors, transducers, and more complex systems. We
describe in general terms the different types of input and output modules. Simple sensors
have limited functionality, while transducers are capable of more complex signal
processing and can also serve as output modules. Complex input-output systems pile on
functionality and intelligence in self-contained processing units in their own right.
Simple sensors. Simple, but precise, analog sensors have been around for hundreds of
years.12 Ways of measuring the physical environment pressure, temperature, humidity,
position, acceleration, mass and weight, time, light, sound, or chemical composition
have been around since time immemorial. And the conversion of physical signals into
electrical impulses is nothing new. However, it has been the post-war digitization of these
electrical signals that has allowed the analog inputs to be converted to bits for
subsequent modular processing and storage.
12
To put things in context, the mercury thermometer, a temperature measuring "sensor", was invented by Daniel Gabriel
Fahrenheit in 1714.
Transducers. A transducer is a "two-way" sensor, which not only converts physical signals
into electrical (and digital) signals, but can also translate digital information back to
physical signals. Transducers include acoustic components (microphones and speakers);
optical components (sensors and cameras); and touch and motion devices. More
sophisticated and emerging forms of transduction (in rough order of mass adoption)
include:
Magnetometers measure the earth's magnetic field to determine the direction to the
magnetic poles. Magnetometers are widely used in aerospace applications. A more
mundane application is (together with accelerometers) for determining the direction
of tilt in a mobile device in order to adjust the image orientation.
Bone conduction propagates sound waves by transferring air vibrations into human
bone vibrations. Inner ear transduction with implants is used in advanced hearing
aids.
A gyroscope is a device that facilitates the sensing of position and motion in three-
dimensional space. Microelectromechanical Systems (MEMS) technology shrinks the
gyroscope into a very small size using silicon components into a MEMS gyroscope.
A light field 3D camera captures depth information about objects by gathering all the
light rays that interact with them. An image sensor records information about the
angle of light, as well as its color, intensity, and position. Light field data is stored
along with image data. Software enables the refocusing of images after shooting, the
creation of 3D images, and the use of gesture control. An in-depth analysis of 3D
sensing can be found in Global Automation: The rise of 3D vision in automation - an
overview of applications, technologies, and competition.
Algorithms are the firmware and software that processes input from sensors and
transducers, making "sense" of what is being "sensed" and generates actionable output.
"Basic" computational algorithms for capturing the physical world, vision, touch sensing,
and many other functions have been around for decades.
The "official" definition of machine learning runs along the lines of "machine learning
is a subfield of computer science that explores the study and construction of
algorithms that can learn from and make predictions on data." Within that, deep
learning is the branch of machine learning that merges artificial multi-layer neural
networks with hyperscale computing and large amounts of training data. Circuits
whose connections are inspired by the connections among neurons in a brain are
tuned to recognize and classify patterns. It is a decades old idea that has been
revived in large part because of the massive computing power and data that is now
available, compared to what existed 10 or 15 years ago. Deep learning is solving
hard "classification" problems such as image (including handwriting and pictures),
video, and sound (including speech) recognition, with much better accuracy and
faster than previously available, allowing computers to answer questions at scale,
quickly and in ways that are nonintrusive.
As interfaces become more natural and intuitive (with capabilities such as highly
accurate speech recognition and vision) and the computational capability increases,
MACHINE LEARNING: INVERTING The key here is "learn from" in other words, machine-learning algorithms can take data
THE COMPUTATIONAL and work their way backward to create or modify the algorithm that could have generated
PARADIGM
it. An agent (or a program, a robot, or even a human) is learning if it improves its
performance in matching the algorithm to the data after making repeated observations of
samples of the data.
Traditional computation doesn't "learn from" anything. It takes input data, processes it
using a static program (basically a set of rules specifying how the input data is to be
manipulated), and generates output data; end of story.
The problem with traditional computation is that the real world is too complicated and
unpredictable for a static set of rules to emulate it. This rule-based approach either fails to
capture all possible outcomes (in the jargon, "had bad exception handling"), or becomes
fiendishly complex trying to capture all the possible states of the world to be modeled.
Machine learning, instead of trying to capture all the complexity and unpredictability of the
real world, emulates human behavior and embraces this complexity and unpredictability.
In machine learning, instead of feeding input data to a program to generate more data as
the output, we feed data to a program that generates another "program."13 We can think
of machine learning as the "inverse" of programming, just like the square root is the
"inverse" of the squaring function, or integration and differentiation are inverses of each
other.
In traditional programming, we can feed the number 4 to a squaring program, and get 16
2
as the output, or feed x +x to a symbolic differentiator and get 2x+1 as the output. In
machine learning, we feed {16,4} to the algorithm and get back f(x) = x2, or { x2+x, 2x+1}
and the algorithm returns f(x) = df/dx.
The facile analogy of machine learning is that of the human brain, which can extrapolate,
analogize, synthetize, and hand-waveize its way to "learning" from both output and input
stimuli. But my favorite description of what machine learning is really "about" goes back to
Alan Turing:14
It has been said that computing machines can only carry out the purposes that they are
instructed to do. This is certainly true in the sense that if they do something other than what
they were instructed, then they have just made some mistake. It is also true that the
13
Which itself may be a program that updates or overwrites the original program with a "better" program that is more suited
to meet the objective. This "recursive" nature of machine learning led AI pioneers to one of the most colorful and entertaining
dead ends in the history of AI, the Lisp Machine (https://en.wikipedia.org/wiki/Lisp_machine). I know this useless fact
because I used to program Lisp Machines in college.
14
Turing's "Proposed electronic calculator" manuscript, also known as the ACE Report. A scan is available at the Turing
Digital Archive.
intention in constructing these machines in the first instance is to treat them as slaves,
giving them only jobs which have been thought out in detail, jobs such that the user of the
machine understands in principle what is going on all the time.
Up till the present, machines have only been used in this way. But is it necessary that they
should always be used in such a manner? Let us suppose we have set up a machine with
certain initial instruction tables, so constructed that these tables might on occasion, if good
reason arose, modify those tables. One can imagine that after the machine has been
operating for some time, the instructions would have altered out of recognition, but
nevertheless still be such that one would have to admit the machine was still doing very
worthwhile calculations.
The bolded text is worth clarifying. Turing believed the process of a machine "changing its
own instructions" to be analogous to that of human learning. He imagined a machine
altering its own instructions, as a "pupil" learning from a "master." A learning machine
he went on to explain in his 1950 Mind paper that introduced the concept of the "imitation
game:"15
might still be getting results of the type desired when the machine was first set up, but in
a much more efficient manner. In such a case, one would have to admit that the progress of
the machine had not been foreseen when its original instructions were put in. It would be
like a pupil who had learnt much from his master, but had added much more by his own
work. When this happens, I feel that one is obliged to regarding the machine as showing
intelligence. As soon as one can provide a reasonably large memory capacity, it should be
possible to begin to experiment on these lines. The memory capacity of the human brain is
of the order of 10,000 million binary digits. But most of this is probably used in
remembering visual impressions, and other comparatively wasteful ways.
Well, we're now entering the era when we can provide a "reasonably large memory
capacity," which brings us back to the present (and future) of machine learning.
MACHINE LEARNING The opportunities for material progress in algorithms, leading to truly "smart"
APPLICATIONS: A QUICK TOUR computational capability are manifold, and the possibilities almost endless. We can
imagine seeing increasingly sophisticated sensing capabilities beyond touch sensors and
hover sensing (where direct touch is not needed) and beyond that, true and accurate 3D
sensing. Natural interfaces, with reliable speech recognition, handwriting recognition,
natural language processing, and even mood recognition are getting closer by the day.
Some of the more promising and interesting machine-learning algorithms are worth
describing in more detail, as they will be the key to the development of truly intelligent
Fourth Wave hardware devices:
15
Alan Turing, Computing machinery and intelligence, Mind 59, 1950. The imitation game, is of course, the game played to
see if a machine can pass the "Turing Test", where an interrogator quizzing the (hidden) machine cannot tell if the responses
come from a machine or a human. It's also the title of a particularly fun (but terribly historically inaccurate) movie about Turing
starring the alien Benedict Cumberbatch and the lovely Keira Knightley.
Gaze control. Gaze control effects computer action by changing the direction of
one's gaze. This involves determining the angle or position of a user's visual attention,
usually through use of cameras, and responding accordingly.
Gesture control. Gesture control is the ability to recognize human body movements to
interact with and control a computer system without direct physical contact. Classic,
existing, widely available systems are Nintendo's Wii and Microsoft's Kinect.
Haptics. In a broad sense, haptics is any system that incorporates tactile feedback
and responds through the sense of touch.
In this last section, we delve a bit deeper into some specific applications from the six
major segments of the Fourth Wave: Human-Machine Interfaces, the Quantified Self,
Smart Machines, Smart Distributed Systems, the Connected Home, and Factory
Automation.
1. HUMAN-MACHINE Machines better understanding humans and the environment, humans better
INTERFACES: MAKING HUMAN- understanding machines, and machines and humans getting smarter together. As
MACHINE INTERACTION MORE
technology evolves, connecting analog humans with digitized systems through analog-
NATURAL, PRODUCTIVE, AND
PLEASANT digital man-machine interfaces becomes easier and more sophisticated. We've identified
the following types of human-machine interfaces:
Tangible user interfaces. In a tangible user interface, the user controls digital
information and processes by manipulating physical, real-world objects rather than
icons on a screen.
Ambient and glanceable displays. Ambient and glanceable devices and displays
convey minimum and specific information in a way that is designed to exploit the
"preattentive" processing ability of the human brain.
2. QUANTIFIED SELF: Quantified Self is the use of self-monitoring through a wide variety of sensors and
MONITORING HEALTH AND devices. Quantified Self uses mobile apps and wearable or portable devices to collect
FITNESS AND IMPROVING
data about a user's activities, biometrics, environment, and other personal experiences.
PHYSICAL CONDITION THROUGH
TECHNOLOGY Analysis of this data allows individuals to gain a better understanding of their physical
condition and improve their wellbeing. We've identified the following categories of
Quantified Self:
Wearable user interfaces. Wearable user interfaces are electronic systems located
on the body that help users interact with their environment and the network without
requiring specific actions, such as viewing a standalone device. There is quite a bit of
activity from start-ups and established companies in developing workable user
interfaces.
Mobile health monitoring. Related (and more expansive than home health monitoring),
in mobile health monitoring patients are provided wearable and portable monitoring
devices that capture physiological metrics, such as blood pressure, glucose level,
pulse, blood oxygen level and weight, and then transmit or stage the patient data for
analysis.
Mobile sports and fitness. Mobile ecosystems are used to track and monitor sports
and health-related efforts. The ecosystem can include anything from mobile apps to
standalone pulse readers that connect wirelessly to PCs.
3. SMART MACHINES: Smart Machines are relatively standalone (and many times autonomous) integrated
INTELLIGENT VEHICLES, systems with contextual awareness and intelligence. These capabilities allow them to
FABRICS, AND REFRIGERATORS
function independently and make "decisions" based on limited objectives defined by
FOR AN EASIER EVERYDAY LIFE
software. We've identified the following classes of Smart Machines:
Smart appliances. Smart appliances such as washing machines have the ability to
connect to a network and transmit information about maintenance and conditions.
Smart fabrics. Smart fabrics are used to make clothing, upholstery, and other textile
goods with devices that can be deployed as electronic sensors, switches,
connectors, batteries, or displays. Smart fabrics are used in applications such as
controllers for electronics, human data monitoring, and alarm systems.
Autonomous vehicles. An autonomous vehicle is one that can drive itself from one
point to another using lasers, radars, and cameras, as well as software, map data,
GPS, and wireless data communications with other vehicles or its environment.
Smart dust. Smart dust "motes" are tiny wireless MEMS, or other devices that can
detect light, temperature, and pressure to vibrations, magnetism, and chemical
compositions. They run on a computer network and are distributed over an area to
perform tasks.
4. SMART DISTRIBUTED In contrast to Smart Machines, Smart Distributed Systems are incorporated into an array
SYSTEMS: CONNECTING SMALL of hardware and communication devices that interact with each other intelligently to carry
INTELLIGENT DEVICES AND
out highly sophisticated functions. We've identified the following types of Smart
HARNESSING THE POWER OF
THE NETWORK Distributed Systems:
Ambient energy harvesting. Ambient energy harvesters gather electrical energy from
gathered light, thermal gradients, vibration energy, and radio waves. The electrical
output of these devices is extremely small, but sufficient to replace batteries or wired
electrical sources in low-power applications.
Telematics. Telematics refers to the use of in-car installed and after-factory devices
to transmit data in real time back to an organization, including vehicle use (for
example, miles driven, speed, and location using GPS), maintenance requirements,
air bag deployment, or automotive servicing. Telematics serves as the platform for
usage-based insurance (UBI) for both fleet and personal auto insurance, including
pay-per-use insurance, pay as you drive (PAYD) insurance, and pay how you drive
(PHYD) programs (U.S. Insurance: The Next Generation of Mobile: A Deep Dive on
Emerging Telematics Technology). Telematics also includes Driver Monitoring
Systems (DMSs) based on interior-facing cameras. Eye-tracking technology allowing
gaze direction and eyelid movement analysis, as well as facial recognition, allowing
personalization, security, health tracking, and distraction and fatigue detection. A
number of auto OEMs have begun to feature DMS systems in their vehicles
Mercedes-Benz's Attention Assist, Ford's Driver Alert, Volvo's Driver Alert Control,
Volkswagen's Fatigue Detection, and Toyota's and Volvo's Driver State Estimation
use a combination of technologies for driver assist. Telematics can be grouped into
three major categories: Consumer, commercial, and public:
cubicles, in-building open spaces, home offices, or mobile settings. In the smart
workplace, "objects" (whiteboards, building interfaces, large digital displays,
workstations, mobile devices, and wearable interfaces) participate in work activities
via communications features that create a network to facilitate interactions.
5. CONNECTED HOME: A Connected Home is networked to enable the interconnection of multiple devices,
INTEGRATING INTELLIGENCE IN services, and apps, ranging from communications to entertainment, and healthcare to
THE HOME IN SUPPORT OF THE
security. These services and apps are delivered over multiple interlinked devices,
LIFESTYLE OF THE OCCUPANTS
providing a connected experience for the household and enabling its inhabitants to
control and monitor it remotely. We identify three major system blocks in a connected
home, namely:
Intelligent lighting. Intelligent lighting is any application that combines the ability to
make use of highly efficient illumination technologies, such as LEDs, motion, light,
time, and other sensors that optimizes lighting conditions and minimizes energy
consumption.
6. FACTORY AUTOMATION: A The substitution of capital for labor in manufacturing has been ongoing since James
MAJOR BENEFICIARY (AND Watt's invention of the steam engine in 1781, which was the catalyst for the First
DRIVER) OF THE FOURTH WAVE
Industrial Revolution. Since then, automation has been progressing steadily, mechanizing
OF IT HARDWARE
the three "D tasks" in manufacturing dirty, dull, and dangerous (or as they say in Japan,
kitanai, kitsui, and kiken).
Defined broadly, Factory Automation is the replacement of labor for machines in the steps
of a manufacturing process. These include pick-and-place, assembly, packing,
monitoring, inspecting, and many others. These manufacturing steps involve the
production or processing of a discrete number of individual pieces or products.
Continuous flow processes (such as oil and sugar refining, or mining) are also amenable to
automation, primarily through flow control automation, which aims to limit downtime and
improve process yields.
Traditional factory automation can be broken down into four layers the Enterprise
Control System, the Plant Control System, the Shop Floor Systems, and the Component
Control Loop (see Exhibit 26). Each of these layers has its own sensing-processing-output
loop and interacts with the layers above (and below) and with the general factory
environment.
EXHIBIT 26: The four layers of factory automation interact with each other and with the factory environment
Service
Management Level
ERP
Business planning
Level 4 Enterprise
Logistic management
Production scheduling and operation management
Function
Monitoring Level
MES
Plant
Manufacturing control
Level 3 Control
Operation control
Room
Information management
Data
Control Level
Sensors
Drives Valves
Power Supply
Level 1 Component Measuring Devices Motors Regulators
Actuators
Output
Gearboxes Flow Measurement
Input
Contactors
PLC = Programmable Logic Controllers SCADA = Supervisory Control and Data Acquisition
DCS = Distributed Control Systems HMI = Human Machine Interface
CNC = Computer Numerical Controls MES = Manufacturing Execution Systems
AMC = Advanced Motion Control ERP = Enterprise Resource Planning
Automation not only allows the replacement of humans in tedious and dangerous tasks
but also augments human capability, allowing (for example) the handling of heavy loads,
large objects, or too hot or too cold substances. On the downside, current technology is
unable to automate all manufacturing steps. High initial and development costs, immature
technology, lack of flexibility, and unfavorable cost-benefit tradeoffs hamper many of the
automation efforts.
But because of the combinatorial, exponential, and deflationary nature of the building
blocks in automation (which are pretty the same of those of the Fourth Wave input,
processing, and output the only difference being that the output module usually
involves mechanical action), and the increasing digitization of information for control and
communication, it is likely we are now at a tipping point in the next wave of industrial
automation.
The main driver of replacing capital for labor in manufacturing (factory automation) has
always been cost reduction. As wages rise, labor availability worsens, and the cost of
automation declines, the tradeoff shifts toward more automation. On top of that, the
Fourth Wave is enabling increasingly capable automation, which accelerates this trend.
As the "total cost of ownership (TCO)" of human labor rises, and the TCO of automation
decreases, the tradeoff point toward automating moves closer in time. We identify three
major areas where the Fourth Wave can have an outsize impact on the acceptance and
evolution of factory automation traditional factory automation, traditional robotics, and
next-generation collaborative robotics.
integration, and worker retraining programs. It can take months to select the best robotic
automated solution and, even after installation, months to program the stationary robot
for a few highly precise, repeatable tasks.
It also very closely resembles the computer industry before the 1980s multiple
providers of vertically integrated systems offer incompatible and expensive problem-
specific solutions. Architecture and interfaces are closely guarded secrets. Any
modularization applies only typically to the present provider's portfolio. Major established
robot manufacturers, such as Fanuc, Nachi-Fujikoshi, Yaskawa, Kawasaki, Hyundai, ABB,
Staubli, KUKA, Comau, Daihen, and emerging ones such as Siasun (in China) fight it out in
order to be the preferred provider at large industrial installations.
But this static state of affairs is rapidly changing. As the Fourth Wave progresses, the cost
of robot components, such as sensors, actuators (hydraulic, pneumatic, and electric),
servo systems (mechanical systems combining a motor, drive, and a feedback device), and
high-precision reduction gears and harmonic drives (from companies such as Harmonic
Drive and Nabtesco in Japan) continues to drop.
This vastly improves robot functionality and capability, bringing forth incremental
versatility to robots as tools in the workplace, augmenting human ability, and even
allowing for multi-robot cooperation in distributed systems. This higher functionality also
helps to materially lower the fixturing and ancillary setup costs currently involved in
robotics and factory automation.
The end outcome is dropping robot and automation costs, increasing functionality, and
rising automation and robot densities in industries that traditionally have not had much
automation food and beverage, agriculture, health care, and consumer electronic
manufacturing and assembly. All incumbent and emerging factory automation providers
and robot manufacturers have been targeting new industries and smaller enterprises as
their future growth areas.
IMPLICATIONS FOR We have done more in-depth work capturing the impact of machine learning on
AUTOMATION FROM THE automation in previous research (Weekend Tech Byte: Thinking Machines and The Long
PRESENT AND FUTURE OF
View: Deep Learning in machine vision and factory automation - application and structural
MACHINE LEARNING
impact), but in summary, we're in the early stages of the transformation of automation by
machine learning:
Stage 0 now. This is where we are now, where automation continues to evolve, but
the paradigm is one of fixed automation. Automation and robotics is fast, reliable, and
tireless, but there is no "intelligence" to it. Robots are painstakingly programmed to
perform specific tasks, and when the task changes, we need to adjust the physical
setting and program it again.
Stage 1 next three to five years. At this stage, machines in factories take advantage
of large datasets and established developments in machine learning to make
automation more flexible, capable, and collaborative. This materially reduces setup
and programming efforts and upfront investments. We are now entering this stage,
and we expect it to "run" for a number of years as it diffuses across many applications
and verticals.
Stage 2 five to 10 years out. In this stage, the existing (and others yet to be
discovered) leading machine-learning algorithms that can build models and make
causal inferences are capable of one-shot learning, and are pre-trained with physical
and relational concepts that are transferred to the factory floor. The outcome is true
"human" behavior by machines.
INVESTMENT IMPLICATIONS
The earliest animals on earth appeared 1 billion years ago. For the next 450 million years,
they floated around in the oceans of ancient earth, grazing on the microbes covering the
sea floor. The organisms were simple and evolution was slow.
About 540 million years ago, the speed of evolution suddenly accelerated by an order of
magnitude. Within the next 25 million years, simple organisms became a huge number of
complex life forms, and almost all present animal groups appeared for the first time. The
emergence of a trove of complex new species was so sudden that it is sometimes called
evolution Big Bang. This period is called the Cambrian Explosion (see Exhibit 27).
Cambrian Explosion worried Darwin, since he could not find a plausible explanation to
why evolution accelerated.
EXHIBIT 27: Evolution suddenly accelerated in the Cambrian Explosion 540 million years ago
500
Number of Families
400
Cambrian Explosion
300
200
100
0
600 500 400 300
Millions of Years Before Present
Source: David Raup "Patterns and Processes in the History of Life," and Bernstein analysis.
16
See Andrew Parker's book: In The Blink Of An Eye: How Vision Sparked The Big Bang Of Evolution.
them were simple, and there was not much hunting possible. But when animals evolved
with eyes, they could see the environment and their food. Animal relationship suddenly
became complex as predators could pursue and kill the smaller ones, and small animals
had to evolve to form shells or move faster to avoid being killed. According to Andrew
Parker, vision was one of the main driving forces of natural selection and evolution. The
gift of vision changed the behaviors of animals.
Today, the gift of vision will change the behavior of machines too. Before computer vision
was developed, machines are like ancient animals 540 million years ago. Without vision,
they are really blind and can't do much by themselves. Fortunately for us, computer vision
has made significant progress over the past five years. In 2010, artificial intelligence (AI)
scientists in the world joined the ImageNet Classification competition, where computers
were given a number of pictures and were asked to identify objects in them. The best
program in 2010 misclassified 28% of the pictures, and the best result in the subsequent
year was 26%. In 2012, deep learning was applied in the competition for the first time,
and it immediately brought down the error rate to 16%, and to 12% in the next year. By
2016, the error rate dropped to merely 3.1%, which is less than the human error rate of
5.1% (see Exhibit 28).
25% 25.8%
Started to use deep learning
20%
16.4%
15%
11.7%
10%
7.3%
Human Error Rate: 5.1% 3.6%
5%
3.1%
0%
2010 2011 2012 2013 2014 2015 2016
The hero behind the progress in computer vision is deep learning. In the last five years,
deep learning has produced great success in image recognition, speech recognition, and
even the games of Go and Atari. But is deep learning the solution to everything? What are
the applications and limitations of computer vision with deep learning?
In this chapter, we will discuss the computer vision technologies, including the historical
methods and the changes brought by deep learning. We will study a few cases of
computer vision applications, namely face recognition and assisted driving. We will then
discuss the limitations of deep learning and offer a few possible solutions that are being
explored in the academia.
Computer vision is a field of science that aims at giving computers the capability to
understand images or videos. Examples of such capability include:
Tracking, where a computer tracks and follows the movement of a person or object.
This is a common task for drones or robots.
While vision is natural and intuitive to humans, that is not the case for computers. For
computers, a digital photo is nothing but a group of numbers that doesn't carry any
meaning (see Exhibit 29 and Exhibit 30). The goal of computer vision is to help computers
make sense of the numbers so that they know what the photo represents.
EXHIBIT 29: A picture of a cat to us EXHIBIT 30: is nothing but an array of numbers to the
computer
22 183 75 180 57 201 120 212 218 250 0 235 130 31 57 204 124 121 176 158 208 238
198 232 8 128 247 76 105 36 243 101 74 116 19 249 36 62 57 48 220 214 149 99
232 175 58 152 107 0 48 43 190 229 248 24 184 179 134 233 246 161 71 126 240 136
105 212 50 64 105 40 74 176 158 141 37 170 213 90 73 36 226 19 101 28 96 153
126 6 127 153 34 201 97 79 175 157 67 67 161 90 126 240 119 145 136 227 229 157
9 164 16 165 206 192 97 136 225 7 133 140 115 47 54 39 86 196 5 46 7 242
120 42 99 196 112 206 69 22 168 7 187 195 139 207 87 175 180 128 18 186 237 138
228 197 137 178 111 112 149 67 41 203 69 111 77 206 50 236 43 109 19 19 9 127
151 239 167 48 192 182 46 56 85 245 204 83 173 71 88 92 27 150 124 16 234 123
189 41 151 43 64 231 40 161 149 117 191 40 249 135 197 107 74 123 137 196 249 243
190 219 180 36 34 30 132 16 114 255 34 167 73 93 195 10 99 166 128 126 179 95
124 64 14 224 156 47 70 213 22 200 226 125 253 147 58 237 14 38 208 213 129 211
38 46 168 24 35 115 209 207 219 212 54 101 1 198 82 205 175 24 215 233 43 186
4 14 147 122 216 88 138 162 234 88 53 47 6 82 178 78 138 39 175 170 72 117
12 20 168 186 109 167 30 96 98 212 29 110 126 10 40 197 148 211 181 199 194 94
190 4 168 127 15 185 131 5 13 206 225 126 15 161 109 78 150 173 191 52 240 6
146 56 141 139 18 69 155 141 48 53 31 80 254 192 249 110 21 90 158 167 212 137
153 186 254 54 222 19 67 149 137 248 145 36 189 131 1 161 41 154 250 229 83 23
140 149 244 235 138 250 249 118 73 216 128 32 146 47 192 109 109 224 137 195 196 84
124 14 122 52 73 129 19 30 245 70 95 224 253 170 127 25 25 3 24 121 99 221
20 171 194 233 24 50 159 35 59 63 190 45 32 211 24 207 7 33 173 189 248 180
To do that, historical methods of computer vision use feature extraction and classification
(see Exhibit 31). As a high-level explanation, to detect a cat, the features may include a
head, two ears, a body, and a tail all of particular shapes and maybe colors. If enough of
such features are detected in a picture, the computer would report that it is a cat.
There are many classic algorithms for feature extraction, such as scale-invariant feature
transform (SIFT) and histogram of oriented gradients (HOG), and they are all used to
describe elementary characteristics such as the shape of a picture. Classification can be
done using either a support vector machine (SVM) or K-nearest neighbor (KNN).
EXHIBIT 31: Traditional computer vision uses fixed feature extractors and, hence, finds it impossible to cover all scenarios
The problem of the traditional computer vision method is that the feature detectors had to
be built by hand. They are hard-wired transformations built into the system. However, this
requires a lot of prior knowledge and experience of the object to be detected. For
example, there are many different kinds of cats, they can be in different colors, species,
and in different poses, and parts of the cat could be hidden, etc. This makes it almost
impossible for the computer scientists to program manual feature detectors to cover all
cases. While scientists could write detectors that work on some cats pretty well, it soon
became impossible to apply them to all cases. That is why in the ImageNet competition,
the error rate was as high as 28-26% in 2010 and 2011, as shown in Exhibit 28.
Deep learning took a different approach to the traditional methods. Instead of hand
crafting the feature detectors, deep learning uses deep neural networks (DNNs) to learn
how to extract the features and classify them automatically (see Exhibit 32). The deep
neural networks are multiple layers of connected, large-collection neural units (artificial
neurons). Each layer of the DNN is responsible for extracting information at a different
level. The lowest-level feature extractor only sees lines of different shapes and at different
angles. The mid-level layer sees local-level features such as eyes or ears. The highest-
level layer extracts objects to be identified, e.g., a cat or a face (see Exhibit 33). In real-life
applications, there can be many layers and connections of the network, hence the name
"deep" learning.
Since deep learning was applied in the ImageNet classification for the first time in 2012,
the image classification accuracy improved significantly. Today, the error rate of image
classification is merely 3.1%, less than the human error rate of 5.1% (see Exhibit 28).
EXHIBIT 32: Deep learning uses multiple layers to extract features and abstractions from the image automatically
EXHIBIT 33: Each layer of the deep neural network extracts features at different levels
The reason deep learning works so well is because the feature extraction and
classification is not hard-wired, but trained with data. With deep learning, computers learn
to tell the difference between dogs and cats by reading tens of thousands or more
pictures of dogs and cats and extract the logic on their own. Hence, the accuracy of the
result becomes a function of the examples the computer has seen. If it sees enough
examples of British Shorthair, it would know the features of British Shorthair
automatically. If it sees enough examples of cats with half a face covered, it would know
that it is still a cat and not something else. This solves the problem of traditional computer
vision, where programmers had to think of and program every scenario to classify a cat.
Deep learning is not a new concept and can be tracked back to 1965. It belongs to a
school of artificial intelligence called machine learning (see Exhibit 34). However, it was
not successful at first because the amount of computing power was not sufficient, and it
took too much time to train the deep neural networks. This changed in the late 2000s,
when almost coincidentally, scientists discovered that GPUs, which were used for video
games, were extremely suitable for deep learning. GPUs excel at the fast matrix and
vector multiplications required not only for convincing virtual realities but also for deep
neural network training, where they can speed up learning by a factor of 50 or more. Since
then, deep learning has made significant progress.
RNN CNN
Deep learning
Machine learning
Artificial Intelligence
Next, we will look at two applications of computer vision, in face recognition and
advanced driving assistance system (ADAS), and discuss the current state of computer
vision in real life.
FACE RECOGNITION In the past, the primary users of face recognition have been the government and law
enforcement agencies that use the system to capture faces and for identification. For
instance, visitors to the United States are required to be photographed upon receiving a
visa to the country. This photograph will be checked against a database of known
criminals and suspected terrorists. When they enter the country, the same photographs
will be used to verify the person. The FBI can draw from over 411 million photos across
the state and federal databases to match criminal suspects.
As the accuracy improves, face recognition is becoming more widely adopted by banks,
airports, and other organizations for identification. Baidu touts the most accurate face
recognition system in the world, with up to 99.77% accuracy.17 It has deployed the
system for employee entry at its Beijing headquarters (see Exhibit 35). Recently, its face
recognition technology has found the first commercial use to verify visitors' identities in
the Chinese tourist destination of Wuzhen. Guests can stay in hotels within Wuzhen, and
face recognition will be used instead of passes to grant them entry to various tourist
attractions. The visitors have their pictures taken when they first enter the park.
Subsequently, when they approach a gate, a camera will take a photo and compare it to
the database. This eliminates the use of photo ID and fingerprint scanning, which was the
previous system used at Wuzhen. According to the company, the new face recognition
system processes identities in merely 0.6 second, "almost as quickly as walking by."
As the cost of face recognition hardware and software drops, face recognition will find
increasingly more novel uses. In Beijing, authorities deployed face recognition technology
at some restrooms in the Temple of Heaven, one of the city's busiest tourist spots, to
prevent toilet paper theft.18
17
https://www.fastcompany.com/3065778/baidu-says-new-face-recognition-can-replace-checking-ids-or-tickets.
18
http://fortune.com/2017/03/20/facial-recognition-china-toilet-paper-thieves/.
EXHIBIT 35: Baidu uses face recognition at its headquarters building in Beijing
Source: Baidu.
Huawei Honor Magic pioneered the face recognition technology in smartphones. It uses
an infrared camera for this purpose. When a message is received, the content of the
message will be displayed on the lock screen only if the phone is held by the owner.19
Other people picking up the phone will see that there is an unread message, but not the
content. Samsung Galaxy S8 includes face recognition as part of the security system as
well (see Exhibit 36), and initial reviews indicate that it is fast and responsive.
19
https://www.forbes.com/sites/bensin/2017/03/15/honor-magic-hands-on-huaweis-face-recognizing-a-i-powered-
concept-phone/#2c7d338010f1.
20
http://www.computerworld.com/article/2897117/alibaba-uses-facial-recognition-tech-for-online-payments.html.
21
https://www.technologyreview.com/s/603494/10-breakthrough-technologies-2017-paying-with-your-face/.
Source: Samsung.
However, despite the high accuracy rates reported, face recognition has a set of
problems. For example, most of the current face recognition systems use a 2D image, and
risk being fooled by a photo. It was reported that Samsung's face recognition system in
Galaxy S8 suffers from this problem.22 Many apps that rely on regular smartphone
cameras for face recognition can also be bypassed with a photo and some software.23
Besides this, poor lighting, sunglasses, hats, scarves, makeup, or other objects can result
in poor face recognition accuracy. There are also glasses designed to target face
recognition software and trick them to believe the wearer is somebody else.24
One solution to these problems is to capture 3D face images for recognition. Instead of a
2D picture, 3D images can be captured with two stereovision cameras, or a regular
camera and a depth camera. 3D face recognition is much more accurate than 2D and
reduces the false acceptance rate (FAR) to a similar level as fingerprint verification.25
Apple is reported to be adding such 3D face recognition software to its next flagship
iPhone in the second half of the year.26
IMAGE RECOGNITION AND Computer vision is the core sensing technology that enables ADAS, and eventually,
ADAS/AUTONOMOUS DRIVING autonomous driving. The ADAS and autonomous driving architecture includes sense,
perceive, decide, and actuate, and computer vision is responsible for the sensing and
perceiving part. Other sensing technologies such as radar and ultrasonic will be used as
well, but vision provides the most information and will become more important as ADAS
features improve.
22
http://www.cnbc.com/2017/03/31/galaxy-s8-facial-recognition-can-be-tricked-with-a-photo.html.
23
http://mt.sohu.com/it/d20170316/129017757_380891.shtml (Chinese only).
24
http://www.theverge.com/2016/11/3/13507542/facial-recognition-glasses-trick-impersonate-fool.
25
https://en.wikipedia.org/wiki/Three-dimensional_face_recognition.
26
http://appleinsider.com/articles/17/03/16/apple-concept-for-biometric-facial-recognition-could-hint-at-iphone-8.
EXHIBIT 37: Computer vision is an important part of the ADAS/autonomous driving architecture
Gather
Filter, interpret &
environment Safely choose
understand sensor Initiate actions
information from actions
data
sensors
Computer Vision
Source: ARM and Bernstein analysis.
With the advancement in AI and machine vision, today, companies such as Mobileye can
use cameras to analyze the road ahead and detect vehicles, pedestrians, traffic signs, lane
marking, and free space (see Exhibit 38). In its early stage, Mobileye didn't use deep
learning as it was not available yet. But over the years, Mobileye has modernized its
algorithms to use deep learning, which significantly improved the accuracy and
robustness of its system. Notice the drivable free space that is marked in green in Exhibit
38 (see the online version for colors). The free space is similar to the walkway next to it,
and the height difference is only around 10cm. It is difficult to separate them using
traditional computer vision algorithms. However, with deep learning, Mobileye has no
problem identifying the driveway and walkway accurately.
EXHIBIT 38: Vehicles with Mobileye's ADAS systems understand what they see
Source: Mobileye.
EXHIBIT 39: NVIDIA's Drive PX 2 uses computer vision to recognize objects on the road
EXHIBIT 40: Mobileye's system is more advanced than NVIDIA's in defining the bounding box using computer vision
While vision is probably the most important sensing tool for ADAS and autonomous
driving, it is not the only one. The Tesla accident, where a Model S slammed into the side
of a tractor trailer turning across its path on a divided highway, reveals the limitation of
computer vision. The camera-based system wasn't trained to recognize the flat slab of a
truck's side, and hence didn't recognize it as a threat and stop the car. While it has been
proven by NHTSA that the accident was not Tesla's fault at all,27 it still shows that
computer vision is not perfect and has to be used in conjunction with other sensors such
as radar and LiDAR for redundancy.
DEEP LEARNING IS ONLY AS Due to the nature of deep learning, a lot of data is needed. And it doesn't just take any data
GOOD AS THE DATA the data has to be labeled. For image recognition, all the images used must be labeled
with what's in the image. For ADAS applications, the video stream captured by camera
must be labeled with pedestrians, traffic signs, etc. Those who own such labeled data
would have a significant lead over competitors, and such leadership is often sustainable
as those who have data usually are able to collect more data.
However, the data labeling process requires a lot of time and manpower. Depending on
the complexity of the data and labeling requirements, one person can only label 10-50
images a day. ImageNet, the largest image database, used almost 50,000 workers from
167 countries in two-and-a-half years to clean, sort, and label 15 million images for
training the computers.
As each company is different, individual computer vision companies need to collect and
label their own data. More importantly, the labeling workloads grow rapidly with the
amount of data collected. Mobileye had 600 employees to annotate images in early 2016,
and that number grew to 1,000 by end 2016. Smaller companies might not have the time
and resources to collect the data. There are third-party companies and crowdsourcing
platforms for data labeling, e.g., Datatang (http://www.datatang.com) and Baidu data
annotation platform (http://zhongbao.baidu.com/mark/home/mark), but the quality of
labeling varies. At the end of the day, the result of deep learning is only as good as the
quality of the data.
Some systems naturally get labeled data and, hence, do not have the problem.
reCAPTCHA is a novel system developed by Carnegie Mellon University that asks people
to solve CAPTCHAs28 to prove they are human (see Exhibit 41). The CAPTCHAs are taken
from illegible text in scanned old books. Every user that solves a CAPTCHA would supply a
labeled image data pair to the system. With more than 100 million displays every day,
the reCAPTCHA program successfully collected a vast amount of data and greatly helped
computer OCR programs to recognize previously illegible text from scanned books. It
finished serving its purpose in 2017 and was replaced by NoCAPTCHA (see Exhibit 42), a
27
https://www.wired.com/2017/01/probing-teslas-deadly-crash-feds-say-yay-self-driving/.
28
Completely Automated Public Turing test to tell Computers and Humans Apart, a type of test used on websites to allow
only human to pass and stop bots.
system that uses behavioral analysis to confirm the user is a human instead of a bot.
However, most other companies are usually not so lucky in having such clearly labeled
data to train the systems easily.
EXHIBIT 41: reCAPTCHA collects data by asking people to EXHIBIT 42: It collected a vast amount of labeled data until
solve CAPTCHAs to prove they are humans it was replaced by NoCAPTCHA in 2017
Besides the difficulty to clean label the data, sometimes it is hard to collect enough image
samples required by deep learning. Take medical diagnosis as an example. To train a
computer to read an X-ray image to diagnose a disease, like other applications in deep
learning, requires a lot of images of such a disease. But if the disease is rare, it might not
be possible to ever obtain enough images to meet the demand of deep learning. As Neil
Lawrence, a professor of machine learning at the University of Sheffield and part of
Amazons AI team, pointed out, "it is unethical to force people to become sick to acquire
data."
One way some companies are getting around that data shortage problem is through
simulation. 3D rendering is now at a point where systems can be trained in virtual
environments created through something like the Unity Engine and the results are good
enough to be applied in real-world situations. Driverless cars are even being trained using
Grand Theft Auto 5.
However, a much more promising solution is to improve the learning algorithm. Compared
with the current deep-learning methods, humans are capable of learning from one
sample, or a very small number of them. A child can learn the concept of animals such as a
fox with a few photos, or even cartoon drawings. Computers, on the other hand, still
require a large number of data.
COMPUTER VISION To work around the problem of limited labeled data, scientists are using new methods
DEVELOPMENT AREAS including unsupervised learning and transfer learning to improve the current deep-
learning algorithm in computer vision.
manually reviewed and labeled. The manual reviewing and labeling can be practical as
there are not nearly as many groups as the inputs.
The challenge with unsupervised learning is that the computer does not know the task it is
asked to do. A task to identify a cat and a person, versus another to identify the facial
expressions of a person, requires different features to be extracted, but it is hard for the
computer to know what the task is and, hence, what features to look for.
EXHIBIT 43: Unsupervised learning is a promising machine-learning method with unlabeled data
SupervisedLearning UnsupervisedLearning
Cat
Dog
Transfer learning can help in situations where the data sample size is small. Transfer
learning is about leveraging the knowledge the computer gained in one particular task
and applying that knowledge to a different task. For example, it is common to pre-train a
convolutional neural network on a large dataset such as ImageNet, and then use it as an
initialization for a different task, e.g., for ADAS applications. Depending on the size of the
new dataset and similarity to the original dataset, different levels of fine-tuning are
required. In addition to computer vision, transfer learning is useful in other deep-learning
scenarios. For example, Google uses transfer learning in Google Translate to improve the
results of translation from one language to another, by taking a pre-trained neural
network of another language pair.29
29
https://www.cnet.com/news/google-translate-uses-machine-learning-for-its-cool-new-trick/ .
EXHIBIT 44: Transfer learning reuses parameters from a pre-trained model in a different application
Both unsupervised learning and transfer learning are still evolving and have their own set
of problems to be resolved. On a positive note, computer vision technology has
progressed significantly over the past five years with deep learning, and the algorithms
have been evolving quickly. According to Christian Szegedy, Senior Research Scientist at
Google, "Today, we can design vision networks that are 5-10 times cheaper and use 15
times less parameters while outperforming their much more expensive counterparts from
one year ago, solely by the virtue of improved network architectures and better training
methodologies."30 We look forward to a future when computers are given a gift of
improved vision.
HOW FAR ARE WE FROM REAL The holy grail of computer vision is to reach human-level cognition. Even though
AI? computer vision has already surpassed human capabilities in some specific tasks such as
image recognition (see Exhibit 28), it is still far from reaching true human-level cognition.
While computer vision can be good at recognizing sophisticated objects, it can at the
same time make mistakes on simple tasks, which often surprise us. This is because
computers lack the "common sense" that humans have. For example, an image
recognition program identified an image of a baby holding a toothbrush as "a young boy is
holding a baseball bat." This is, in part, because the computer does not know the size and
weight of a baseball bat and, hence, that it is impossible to be held by the baby. We have
yet to find effective ways to teach such knowledge to the computers.
Taking it to the next level, humans are capable of not only understanding what they see
but also explaining them. Looking at the picture shown in Exhibit 45, a person would not
30
http://www.kdnuggets.com/2016/01/deep-learning-2016-beyond.html.
only see that there are soldiers and civilians, but also know that the soldiers were pointing
their guns at the civilians, that the civilian was putting flowers into the gun barrel, and feel
the antiwar emotion and be influenced by it. A computer vision program, on the other
hand, would probably never know what the man is doing, even if it accurately identifies all
the details in the picture. In this sense, we are still very, very far from real AI.
EXHIBIT 45: Computer vision has a long way to go in understanding this picture like we do
Source: Wikipedia.
INVEST IN THE CAMERA SUPPLY For investors, one industry that will likely benefit from the growth of computer vision is the
CHAIN camera supply chain. We believe that the gift of vision to computers increases camera
usage. All smartphones today already have cameras equipped to take pictures and selfies.
Going forward, they will likely be equipped with additional cameras that enable them to
recognize faces, understand gestures, scan environments, or turn any apartment into an
augmented-reality-game playfield. While most of the smartphones today have two
cameras, we expect eight cameras in the future, increasing the addressable market four
times. We forecast handset camera shipment to grow 66% in the next five years, driven by
the additional cameras for face recognition, iris scanning, gesture sensing, and AR/VR
(see Exhibit 46).
With image recognition, vehicles today can see other vehicles, pedestrians, traffic signs,
and road markings. This is one of the cornerstones of ADAS and autonomous driving,
which will not only make driving safer for everyone31 but also eventually free every driver
31
http://fortune.com/2017/01/19/teslas-autopilot-tech-safer-nhtsa-report/.
from spending 290 hours behind the wheels every year.32 We expect automotive camera
shipment to grow 145% in the next five years (see Exhibit 47), driven by growing ADAS
penetration and increasing number of cameras used with increasing ADAS functions.
VR and AR devices will likely have multiple cameras installed so that they can track the
environment and the head position. This technology, called inside-out tracking, is already
used in Microsoft HoloLens, Intel's Project Alloy, and Google Tango. It cuts the
cumbersome setup process of external trackers that are currently required by some VR
systems such as HTC Vive. Similar technologies also apply to drones and robots, which
use cameras for path finding and obstacle avoidance.
Surveillance cameras also benefit from AI to make our homes and organizations safer.
Surveillance systems that include video analytics analyze video footage in real time and
detect abnormal activities that could pose a security threat. Essentially, video analytics
technology helps security software "learn" what is normal so it can identify unusual, and
potentially harmful, behavior.
EXHIBIT 46: We forecast handset camera shipment to grow EXHIBIT 47: We expect automotive camera shipment to
66% in the next five years grow 145% in the next five years
140
5,000
120
4,000
100
Unit m n
Unit m n
3,000 80
60
2,000
40
1,000
20
0 0
Source: TSR, and Bernstein estimates and analysis. Source: Strategy Analytics, and Bernstein estimates and analysis.
In our coverage, Sunny Optical (outperform), Sony (outperform) and Largan (market-
perform) are in the camera supply chain (see Exhibit 48). Among them, we like Sunny
Optical for its #1 position in vehicle lens and fast-growing #2 position in handset camera
lens. We also like Sony for being #1 in CMOS image sensor.
32
Data for American drivers. Source: http://newsroom.aaa.com/2016/09/americans-spend-average-17600-minutes-
driving-year/ .
EXHIBIT 48: In our coverage, Sunny Optical, Sony, and Largan are in the camera supply chain
Largan,Sunny
Optical,
Genius, Sekonix,etc.
Camera Lens
Image Sensor
SunnyOptical,
LGInnotek,Semco
Ofilm,etc.
Sony, Samsung,
OmniVision,
GalaxyCore,etc.
In this section, we list the profiles of a few leading computer vision companies in China as
a reference for investors.
WISESOFT (002253-CN, NOT Wisesoft Co., Ltd. engages in air traffic management services for military and civil aviation
COVERED) users. Its products include air traffic control automation system, simulation training
system, traffic management system, airspace management system, and air traffic control
weather management system. The company was founded in August 2000 and is
headquartered in Chengdu, China.
In recent years, Wisesoft has been trying to achieve core technology innovation, especially
in image recognition in the air traffic control, urban transportation, and other industries.
The company's intelligent image recognition and synthesis business has three main
segments: automobile identification in transportation management, 3D face camera and
face automatic identification products, and intelligent, real-time interactive, virtual reality
products.
Wisesoft's core vehicle identification technology recognizes vehicle model and license
plate number based on the video footage. From 2002 to 2007, Wisesoft was the national
leader in the four tests held for the Beijing Olympic Games intelligent transportation
project. In 2015, its vehicle identification products achieved an accuracy rate of 99.60%,
more than the national standard requirement of 90%. Intelligent transportation products
can be applied to security checkpoints, electronic police, traffic video, detection system,
urban traffic control, speeding detection, etc.
In face recognition, the company has completed the development of a full-face high-
precision 3D face camera. Compared with 2D face identification, 3D identification
software can achieve more than 99.5% recognition accuracy and strong anti-
counterfeiting capabilities. The company believes that its product accuracy is several
times higher than that of competing products, with its accuracy of human face depth
measurement reaching 0.1mm. The product has been tested in the Chengdu East Railway
Station automatic inspection system for a year, and achieved 95.6% accuracy on 5 million
identifications. This product has been expanded to Chongqing and Guiyang railway
stations and will have economic benefits.
Intelligent virtual reality products can automatically identify participants' gestures and
actions and can be controlled under the command of participants' gestures. In 2016,
Wisesoft also developed interactive science classrooms, virtual reality smart cinema, etc.
The products are expected to enter the market in the second half of 2017.
HANWANG (002362-CN, NOT Hanwang Technology Co., Ltd. is traditionally a manufacturer of optical character
COVERED) recognition (OCR) devices. Its current products include portable electronic books,
documents cameras, face identification, and electronic signature devices. The company
was founded by Yingjian Liu and Dongjian Xu in 1998, and is headquartered in Beijing,
China.
Hanwang was one of the earliest Chinese companies that entered the face recognition
field with independent core technologies. It has a series of face recognition products in
the field of access control and attendance machines, which are widely used by various
companies and government agencies such as Minsheng Bank, China Merchant Bank, and
Anhui Government. The use of deep learning methods further enhanced its face
recognition technology. In 2016, Hanwang ported the deep learning model to the low-
end MIPS platform to develop the industry's first embedded infrared identification
product. This product can effectively improve the access control of enterprises and
institutions for better attendance management. The same year, the company launched a
PC-based algorithm and entered fields such as public security and hotel identity
authentication. Hanwang also took a leading position in the ARM platform based on the
deep learning algorithm. The product has been tested through the Security Alarm
Detection Product Testing Center of the Ministry of Public Security and is used in public
security police terminals.
FACE++ (PRIVATE) Face++ is a private company founded in 2011 by three young Tsinghua University
graduates. It has raised US$147 million from investors, including Foxconn Technology
Group, CCB International, and Qiming Venture Partners, with the latest round of US$100
million completed in December 2016 with an unknown valuation. The company's key
One of its most notable clients is Ant Financial, the financial subsidiary of Alibaba. Ant
Financial, which provides Chinas largest online payment platform Alipay, uses Face++'s
software to confirm user identities after they set up an online payment account. Mostly
because of this, Face++ claims to have an 85% market share in face recognition in
China's financial services industry.
Over 30,000 application developers have used Face++. Face++s API is integrated in
various photography and entertainment applications, such as Meitu and Camera360
two of the most popular photo apps in China and Flinch, a Ukrainian dating service.
INVESTMENT IMPLICATIONS
We have recently initiated coverage on these stocks (see our initiation report dated
April 25, 2017: Asian Emerging Technologies: The gift of sound and vision
initiating with a positive view).
We rate Sunny Optical (2382.HK) outperform with a target price of HK$74: Sunny
Optical is the #2 player in the growing handset lens market and has been narrowing
the gap with the leader. Share gain and margin expansion make Sunny a major
beneficiary of the increasing handset camera demand. It also continues to enjoy
dominant position in the fast-growing vehicle lens market, thanks to the high entry
barrier. Current valuation is high but justifiable, and our target price is derived from
23.5x 2018 EPS.
We rate Sony (6758.JP) outperform with a target price of JPY4,500: As the largest
image sensor maker, Sony expects image sensor shipment and utilization to increase
and profitability to grow with market growth. PlayStation VR will likely continue to
lead in tethered VR for the next three years at least, which will further strengthen the
PlayStation ecosystem. Image sensor and gaming will likely contribute 80% of the
33
FDDB: http://vis-www.cs.umass.edu/fddb/; 300-W: https://ibug.doc.ic.ac.uk/resources/300-W/; LFW: http://vis-
www.cs.umass.edu/lfw/ .
operating income growth next year and continue to drive earnings in the subsequent
years.
With the profound changes promised by fully autonomous vehicles, the topic is now
indeed headline-grabbing. Though the constant news flow about autonomous vehicles
makes many investors excited, we believe that the mass deployment of fully autonomous
vehicles will take years, given the myriad of legal and ethical issues associated with the
absence of human drivers. While the timing and degree of autonomous driving remain
hazy, advanced driver assistance systems (ADAS) as the path to fully autonomous
driving presents tangible and investable opportunities to investors in the next few years.
Automotive OEMs are increasing the ADAS content in their vehicles to enhance safety and
provide a differentiated driving experience. Government legislations are also backing
ADAS deployments to reduce accidents and fatalities. As ADAS becomes more
sophisticated, it will meet the requirements of higher levels of autonomous driving and
evolve from assistive, semi-autonomous to eventually driverless and fully autonomous
driving (see Exhibit 49).
EXHIBIT 49: Society of Automotive Engineers (SAE), an automotive standardization body, defines six levels of autonomy
for cars from Level 0 with no automation to Level 5 with full automation
Level 0 Level 1 Level 2 Level 3 Level 4 Level 5
No Automation Driver Assistance Partial Automation Conditional High Automation Full Automation
Automation
The full-time The driving mode- The driving mode- The driving mode- The driving mode- The full-time
performance by the specific execution by a specific execution by specific performance by specific performance by performance by an
human driver of all driver assistance one or more driver an automated driving an automated driving automated driving
aspects of the dynamic system of either assistance systems of system of all aspects of system of all aspects of system of all aspects of
driving task, even when steering or acceleration/ both steering and the dynamic driving task the dynamic driving the dynamic driving task
enhanced by warning or deceleration using acceleration/ with the expectation that task, even if a human under all roadway and
intervention systems information about the deceleration using the human driver will driver does not respond environmental
driving environment and information about the respond appropriately to appropriately to a conditions that can be
with the expectation that driving environment and a request to intervene request to intervene managed by a human
the human driver with the expectation that driver
performs all remaining the human driver
aspects of the dynamic performs all remaining
driving task aspects of the dynamic
driving task
34
https://www.nhtsa.gov/press-releases/us-dot-and-iihs-announce-historic-commitment-20-automakers-make-
automatic-emergency.
EXHIBIT 50: ADAS is the fastest-growing application in the automotive semiconductor market
40 ADAS 18%
35 Body 8%
30
Chassis 7%
25 EV/HEV 16%
US$B
20 Infotainment 7%
15
Powertrain 4%
10
Safety 7%
5 Aftermarket 1%
Instrument 7%
0
2014 2015 2016 2017E 2018E 2019E 2020E
Instrument Cluster Aftermarket Safety Powertrain Infotainment EV/HEV Chassis Body ADAS
PROCESSORS
Digging further, the ADAS semiconductors can be broadly classified into processors and
sensors. The processors are essential to process data and make decisions, and work
together with sensors, which gather information on the surrounding environment such as
pedestrians and oncoming traffic. The combination of both types of semiconductors
enables the ADAS system to sense, think, and act to aid drivers and prevent accidents.
ADAS PROCESSORS LAG MOBILE Initially, we thought ADAS processors would be built with cutting-edge wafer
SOCS BY A MEANINGFUL GAP IN manufacturing technologies (i.e., nodes) as the processors need to recognize objects and
WAFER MANUFACTURING
traffic conditions and determine how to steer the car quickly. Then, contrary to our
TECHNOLOGIES
expectation, we realize that the market leader, Mobileye (MBLY, not covered, being
acquired by Intel now), can meet the requirement with mature nodes. We further conclude
that the valuable knowhow is mostly in software/algorithms and years of data that is used
to train and optimize the software/algorithms. Silicon actually doesn't need to be as
advanced as we thought.
EyeQ3, the chip that powered the autopilot function of the Tesla Model S previously, is the
best example of an ADAS processor built with lagging node. It is designed by Mobileye
and is fabricated with 40nm technology node with a die size of only 42mm2. This means it
is roughly equivalent to a feature phone chip in terms of wafer manufacturing technology.
Its successor, EyeQ4, advances to a 28nm technology node but won't be commercially
found in vehicles until 2018. EyeQ4 also lags mobile SoCs by six to seven years in terms
of node. EyeQ5, Mobileye's fifth-generation SoC that the company pitched as the "central
computer performing sensor fusion for self-driving cars in 2020," is designed with a 7nm
technology node. This signals the accelerating adoption of advanced technologies, but
the vehicles powered by EyeQ5 won't hit the road until 2020. The wafer manufacturing
technology of EyeQ5, therefore, will still be about two years behind mobile SoCs (see
Exhibit 51). Competing architectures such as GPU from NVIDIA (NVDA, not covered) and
MCU from TI, ADI, and Freescale (now NXP) also lag mobile SoCs by a meaningful gap in
terms of manufacturing technology. ARM's roadmap also shows a similar gap (see Exhibit
51 to Exhibit 53).
EXHIBIT 51: In terms of manufacturing technology, ADAS chips lag mobile SoCs by a meaningful gap
Processor Manufacturing Technology Roadmap (Based on Mass Production)
Mobile SoC
Manufacturing 28nm 20nm 16/14nm 10nm 7nm 5nm
Technology
6-7 Years ~2 Years
Mobileyes Chip
Manufacturing 90nm (EyeQ2) 40nm (EyeQ3) 28nm (EyeQ4) 7nm (EyeQ5)
Technology
~2 Years
NVIDIAs Chip
Manufacturing 16nm (DRIVE PX 2)
Technology
2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
EXHIBIT 52: ARM's roadmap shows that automotive processors are also much behind mobile processors in wafer
manufacturing technologies
EXHIBIT 53: Chips for automotive applications from other vendors are also at lagging nodes
Vendor Chip Node
Analog Devices BF609 65nm
Freescale S32V 28nm
Myriad 1 65nm
Movidius
Myriad 2 28nm HPM
TDA3x 28nm
Texas Instruments
TDA2x 28nm
Source: Corporate reports and Bernstein analysis.
THERE ARE MANY COMPETING There are many competing approaches to ADAS applications, but we believe the
APPROACHES TO ADAS specialized architecture is most suitable for the vision-based segment, while conventional
APPLICATIONS
MCUs will remain mainstream in the low end.
As ADAS improves to fulfill more complex tasks, the evolving requirements open an
opportunity for different design architectures to compete. At one end, the incumbent
MCU suppliers benefit from their long industry track record and are elevating the
performance of their chips to overcome the rising computational burden of more
advanced features. At another end, heavyweights like NVIDIA, Intel (INTC), Qualcomm
(QCOM), Altera (part of Intel), and Xilinx (XLNX) are also entering this arena. Finally, led by
Mobileye (being acquired by Intel now), smaller companies are rising rapidly, with the
emphasis on the architectures specific for this market (see Exhibit 54).
EXHIBIT 54: There are many competing approaches to ADAS applications; smaller companies with design architectures
specially made for ADAS are rising quickly
ADAS Processor Vendors by Architecture Type
Specialized
GPU SoC/CPU FPGA MCU
Architecture
NVIDIA QCOM Altera NXP Mobileye
Intel Xilinx TI Movidius
Renesas Inuitive
ADI
Toshiba
Note: Altera, Movidius, and Mobileye were/are being acquired by Intel; NXP is being acquired by Qualcomm.
SPECIALIZED ARCHITECTURE IS We note that computer vision is highly important for ADAS and a very demanding task. For
MOST SUITABLE FOR THE example, a 2-megapixel camera running at 60 frames per second will generate about 0.5
VISION-BASED SEGMENT
gigabyte of data every second. With multiple cameras to capture images from different
angles, the amount of data is massive. Further, the data must be processed in real time to
steer cars away from dangers. Given these factors, our first reaction was that computer
vision would need expensive cutting-edge manufacturing technologies that probably only
giants like Intel and Qualcomm can afford. However, much to our surprise, our research,
as discussed earlier, finds that Mobileye and other small companies use special IC design
architecture and algorithm and are able to offer highly competitive products.
This is because CPUs and SoCs from Intel and Qualcomm are designed to run different
software for a myriad of tasks and, hence, are "general-purpose" chips. Chips from
Mobileye and other small companies, however, are designed to accomplish merely a few
specific tasks with tailored software. They also exploit the repetitive nature of the data
and have many small but efficient "cores" to process data in parallel. As such, we find that
Mobileye's chips are built with more mature nodes with smaller dies. The chips are, hence,
cheaper, but the optimized architecture and algorithm enable them to deliver remarkable
performance and power efficiency (see Exhibit 55). This explains Mobileye's leading
position in the vision-based ADAS processor market.
EXHIBIT 55: Mobileye's chips are smaller and more efficient than NVIDIA's
NVIDIA Mobileye
Tegra K1 Tegra X1 DRIVE PX 2 EyeQ3
(28nm) (20nm) (16nm) (40nm)
Peak Performance (GFLOPS) 326 1024 256
Actual Performance (GFLOPS) 76 287 205
2
Die Size (mm ) 132 126 481? 42
2
Actual Performance/Area (GFLOPS/mm ) 0.6 2.3 4.9
Power Efficiency (GFLOPS/W) ~9 29 ~82
Source: Luca Benini, Integrated Systems laboratory of ETH Zurich, Universit di Bologna, and Bernstein estimates and analysis.
In fact, larger companies also recognize the advantages of specialized architecture and
are moving in this direction. For example, Qualcomm's "zeroth" initiative is embedding a
"neural process unit" (NPU) into its latest SoCs for the "cognitive computing" requirement
of automotive, wearables, mobile, and other applications. NVIDIA is also banking on its
GPUs that have many cores to process image data in parallel. Freescale (merged with
NXPI, which is being acquired by Qualcomm) acquired CogniVue, a Canada-based IP
company, to jump-start its vision-processing capability. Intel announced, in September
2016, its plans to acquire computer vision chip company Movidius and only a few months
later, in March 2017, announced another plan to acquire the ADAS processor market
leader Mobileye, in an effort to put itself ahead of the competition in the ADAS and
autonomous driving market (see Exhibit 56).
THE SUPPLY CHAIN OF ADAS ADAS applications incorporate many technologies and companies in the supply chain (see
PROCESSORS INVOLVES MANY Exhibit 57). Processors sit on the key control point. With software/algorithms running on
PARTICIPANTS
them, the processors take the input information on the environment surrounding a vehicle
in real time (and mapping information for some applications) and determine how the
system should actively act/react to the situations to aid drivers or drive autonomously.
Like ADAS processors, sensors also offer an important growth opportunity for
semiconductor companies and we will discuss this in detail later in the chapter.
EXHIBIT 57: The supply chain of ADAS processors includes many participants
Representative Companies
Sensor Vendors Continental, Bosch, Delphi,
Denso, Wabco, Autoliv
SENSORS
VARIOUS TYPES OF SENSORS Different types of sensors are used in ADAS to gather information and they have different
USED IN ADAS HAVE DIFFERENT pros and cons. The most common sensors that can be found in ADAS are camera, radar,
CHARACTERISTICS 35
ultrasonic, and LIDAR (Light Detection and Ranging). Each of these offers different
characteristics in sensing distance, resolution, cost, etc. (see Exhibit 58).
35
LIDAR is a type of sensor that measures distance to a target by illuminating that target with a laser light.
EXHIBIT 58: Camera, radar, ultrasonic, and LIDAR are the most common sensors used in ADAS; given their different
characteristics, including sensing distances, they are adopted in various ADAS applications
Sensor Range Characteristics
Ultrasonic Ultrasonic 0-4m Similar to radar but with shorter detection distance
CAMERA SENSOR IS THE MOST As all sensor technologies have their inherent limitations and advantages, they are often
VERSATILE TYPE OF SENSOR used in different functions of ADAS applications. Camera is the most widely used sensor
in almost all ADAS functions, followed by radar which is used in emergency brake assist,
pedestrian detection, etc. LIDAR is used for adaptive cruise control, and ultrasonic is
typically used for park assist only (see Exhibit 59). "Sensor fusion," in which multiple types
of sensors are used together to complement each other and provide redundancy, is
gaining popularity too.
EXHIBIT 59: All sensor technologies have their inherent limitations and advantages; thus, they are used in different
functions of ADAS; camera is the most versatile and is used in almost all ADAS functions
Traffic-sign-recognition (TSR) X
CAMERA SENSOR ENJOYS THE Camera is used to sense visual light reflected or emitted from objects. High resolution and
COST REDUCTION BROUGHT BY the ability to detect color make the camera sensor ideal to identify and classify objects. It
MOORE'S LAW
is also very inexpensive and enjoys the cost reduction brought by Moore's Law (see Exhibit
60). Though the performance is affected by lighting and weather conditions, it is still one
of the most popular and fastest-growing sensors for ADAS applications. Strategy
Analytics forecasts camera demand in ADAS to reach 96 million units in 2019, up 179%
from 2014 (see Exhibit 61).
EXHIBIT 60: Camera is used to sense visual light reflected or emitted from objects; it is very inexpensive and enjoys the cost
reduction brought by Moore's Law, with the cost per one million pixels dropping at a rate of ~25% per annum
100
Cost/ MPixel (US$)
0.10
1980 1985 1990 1995 2000 2005 2010 2015
Year
EXHIBIT 61: Camera is the largest sensor technology for ADAS and it is expected to reach 96 million units in 2019, nearly
tripling from 2014
90
80
70
Million Units
60
50
40
30 2014
20 2019
10
0
LIDAR Long-range Sh/Med-Range Camera
Radar Radar
Radar emits radio waves and collects the reflection from objects to determine its range,
direction, and speed. Radar can see through fog, rain, or snow, and can measure distance
very quickly and effectively. Doppler radar has the added advantage of being able to
detect the motion of objects. However, radar has lower resolution and cannot easily
identify objects. The fusion of visible and radar data provides a solution that is much more
robust under a wide variety of conditions.
Ultrasound is similar to radar but relies on ultrasound instead of radio waves. Ultrasound is
commonly used in parking assist (PA) because of its shorter detection distance.
LIDAR is also an active sensor and employs an infrared laser (see Exhibit 62), instead of
radio waves or ultrasound. LIDAR provides very accurate distance measurement and can
provide 3D representation of objects as the collected data is processed. Emitting laser
actively, LIDAR can better deal with low-light environments than a camera sensor but can
have trouble in heavy rain, snow, and fog. In some cases, the resolution of LIDAR isn't as
good as that of a camera sensor. LIDAR's biggest bottleneck is its high cost: the one used
in the Google car costs US$75,000. Though some expect the price to drop drastically to
below US$100 in five years,36 it remains much more expensive than a camera sensor,
which costs as little as US$5 in smartphones.
EXHIBIT 62: A LIDAR from Velodyne used in the Google self-driving car
36
https://www.washingtonpost.com/news/innovations/wp/2015/12/04/the-75000-problem-for-self-driving-cars-is-
going-away/.
MORE SENSORS WILL BE USED The concept of "sensor fusion" is also important. In "sensor fusion," the data from different
TO SUPPORT THE HIGHER sensors are analyzed together to provide the increased accuracy, reliability, and
LEVELS OF AUTONOMY OF CARS
robustness under a wide variety of conditions.
ADAS is evolving quickly and so is the demand for different sensors. The Society of
Automotive Engineers37 published six levels of autonomy of cars, from Level 0 with no
automation at all, to Level 5 in which cars take full control and drivers are not expected to
be available at any time, and four other levels in between where cars are only partially
autonomous and humans should take control when needed (see Exhibit 49). As the ADAS
system become more sophisticated and move to higher levels of autonomous driving, the
number of sensors will also rise (see Exhibit 63).
EXHIBIT 63: More sensors will be used to support the rising complexity of higher levels of autonomy
AUTOMOTIVE IMAGE SENSOR IS Taking a closer look at the automotive camera sensor, we summarize its supply chain into
EXPANDING QUICKLY image sensor, foundry, outsourced assembly and test (OSAT) companies, optics, camera
module, tier-1 integration, and auto OEMs (see Exhibit 64).
37
The Society of Automotive Engineers is an automotive standardization body. The classification is published here.
EXHIBIT 64: The automotive camera sensor supply chain consists of image sensor, foundry, OAST companies, optics,
camera module, tier-1 integrator, and auto OEMs
Foundry
Image Optics Camera Integrator Auto
TSMC Sensor Module (Tier-1) OEM
(2330.TW)
SMIC ON Semi Sunny Optical Valeo Valeo Volkswagen
(981.HK) (ON.NASDAQ) (2382.HK) (FR.FP) (FR.FP) (VOW.FRA)
UMC
(2303.TW) Sony Largan Panasonic Continental Audi
(6758.JP) (3008.TW) (6752.JP) (CON.FRA) (NSU.FRA)
LFoundry
OmniVision Sekonix Sekonix Delphi BMW
(private) (053450. KOSDAQ) (053450. KOSDAQ) (DLPH.NYSE) (BMW.FRA)
Tung Thih
(3552.TW)
As a core component in cameras, the automotive image sensor market is fast expanding,
reaching US$657 million in 2015 (see Exhibit 65). Of the two common types of image
sensors for automobiles, the charge-coupled devices (CCD) market is diminishing while
the CMOS image sensor (CIS) market is expanding rapidly. Automotive CIS was a US$591
million market in 2015, having grown rapidly from being insignificant just a few years ago.
It is expected to continue to grow at a 25% CAGR to reach US$1.8 billion in 2020 (see
Exhibit 65). Automotive CIS is also playing a broader role in the overall CIS market, as the
pace of the largest application smartphones decelerates (see Exhibit 66).
EXHIBIT 65: Of the two common types of automotive image sensors, CCDs are diminishing while CISs are fast expanding
1.4
1.2
CCD
1.0
0.8 CIS
0.6
0.4
0.2
0.0
2012 2013 2014 2015 2016E 2017E 2018E 2019E 2020E
EXHIBIT 66: Automotive CIS is becoming a major driver of the overall CIS market growth
CIS Revenue
16 16%
14 14%
12 12%
US$B
10 10%
8 8%
6 6%
4 4%
2 2%
0 0%
2012 2013 2014 2015 2016E 2017E 2018E 2019E 2020E
Image sensors for automotive have some unique technological considerations, compared
to those for smartphones. Smartphone sensors now often tout 13 million or more pixels
for crisp image quality. Automotive image sensors, on the other hand, usually have a much
lower pixel count 0.3 million pixel for side and rear views to 1 million or more for front-
view cameras are common now. However, automotive image sensors demand high
sensitivity to capture images in the dark. Some use both visible light and infrared to
improve sensitivity. They often employ high dynamic range (HDR) to deliver good scene
reproduction in demanding lighting conditions. They also need to work in a wide
temperature range, from -40C to 100C. Other requirements include high-speed image
processing and the ability to reduce noise from sunlight and LED lights.
INVESTMENT IMPLICATIONS
NVIDIA, CEVA, ON Looking at both processor and sensor supply chains, we conclude that companies with
SEMICONDUCTOR, SUNNY strong domain expertise will have a competitive advantage in the ADAS market.
OPTICAL, SMIC, AND KINGPAK
ARE WORTH ATTENTION IN THE
ADAS SEMICONDUCTOR For processors, Movidius and Mobileye both have specialized designs of
MARKET software/algorithms and chip architectures for vision processing. Mobileye, in particular,
commands an absolute leadership in the vision-based ADAS processor market. After
Movidius and Mobileye were acquired by Intel, NVIDIA and CEVA (CEVA, not covered), the
digital signal processing (DSP) IP leader with a popular vision processing IP core, could be
worth attention, in our view
As for sensors, camera is the most important type of sensor for ADAS applications. Along
the automotive camera sensor supply chain, ON Semiconductor (ON, not covered), Sunny
Optical (2382.HK), Kingpak (6238.TW, not covered), and to a lesser degree SMIC
(981.HK), will likely benefit from the accelerating adoption of ADAS.
ON Semiconductor is the leader in automotive image sensors, with a 38% market share in
2015. Its CMOS image sensors are widely incorporated in rear-view cameras and
command a high market share because of their high sensitivity in low-light conditions and
wide dynamic range technology. The company also intends to develop and provide total
imaging solutions for automobiles, including image signal processors.
Sunny Optical is the #1 player in the automotive lens set market, with about 30% market
share in 2015. The automotive lens segment revenue contribution is not high (6% in
2015, up from 5% in 2014), but the company expects this segment to expand quickly and
bring further growth to the company. Its products include rear/front/side/surround-view
lens and enlist tier-1 car makers, including BMW, Benz, Audi, Toyota, and Honda as
customers. As this segment enjoys a much higher gross margin than its corporate
average, faster growth and higher contribution will likely also help lift Sunny Optical's
corporate margin.
Kingpak, a small OSAT company, is another noteworthy name, as the company provides
sensor packaging and testing services to ON Semiconductor.
ADAS IS IMPORTANT BUT As for other companies under our Asian semiconductors coverage, although ADAS will
SMALL COMPARED TO have a profound impact broadly, the upside is limited in the next three to five years
SMARTPHONE FOR OTHER
compared to that of the smartphone. TSMC (2330.TT) has the world's largest CIS foundry,
COMPANIES IN OUR COVERAGE
but the business is small compared to the company's scale. We estimate that total CIS
accounts for merely a low single digit of TSMC's revenue and automotive CIS accounts for
less than 1%. The profit contribution will be even lower, as the CIS manufacturing process
is relatively simple and carries lower margins. TSMC's exposure to ADAS processors is
rather limited as well. Similarly, for other foundry and OSAT companies in our coverage,
ADAS will likely bring relatively limited opportunity. ADAS is more meaningful for design
companies, but is not significant for either MediaTek (2454.TT) or Novatek (3034.TT).
12 GAME-CHANGING TECHNOLOGIES IN
THE FUTURE OF HEALTHCARE
Where could Asia leapfrog the West? Laura Nelson Carney, PhD
Asia's healthcare systems face the same challenges that plague developed markets (e.g.,
costs spiraling up, misaligned incentives across stakeholders, and unpaid debts), plus
some unique and arguably more difficult emerging market obstacles to healthcare
modernization (e.g., wider income and health inequality, low health spend per capita,
inadequate infrastructure, endemic corruption, and broken public trust).
Technology can potentially solve some of these problems better than old approaches. We
think that technology may change the face of healthcare faster in Asia than elsewhere in
the world innovations in how care is administered, independent of innovations in
healthcare treatments (i.e., specific therapeutics new drugs, devices, and treatment
paradigms). In this chapter, we discuss 12 emerging technologies that we think will be
transformative in healthcare in Asia and elsewhere artificial intelligence (AI); genomics-
enabled care (precision medicine); gene editing; telemedicine; augmented reality; the
Internet of Things (IOT) monitoring and diagnostics; blockchain; smart pills; connected
communities; automation and robotics; fully digitized hospitals; and gamification of health.
We discuss why and where the pace of change toward widespread implementation of
these technologies throughout the health system may be faster in Asia (AI, genomics,
telemedicine, and connected communities) and highlight emerging companies to watch.
Who will benefit? First, patients are the most important beneficiaries in the future, they
will be able to access healthcare professionals in more convenient and continuous ways,
engage with the healthcare system when they are both well and sick, enjoy greater risk
and pricing transparency, receive therapies that are personalized (and far more effective),
and benefit from treatment pathways selected with wisdom from analytics across millions
of patient records to understand outcomes and risks of similar cases. Second, private
healthcare services businesses in Asia have exploded in scale and complexity in recent
years due to attractive demographics, growing medical tourism, favorable policies, and
epidemiological trends that will likely drive strong growth in demand for the next few
decades. Over time, those healthcare businesses that have heavily invested in technology
to deliver better care to patients have performed the best financially, and we expect that
trend to continue. Third, the tech companies creating and commercializing these
innovations will likely grow dramatically.
China's healthcare system has made more progress compared to any other country over
the last decade, but it is still a long way from being ideal. Many seemingly intractable
problems remain unsolved (see Exhibit 67). The rest of emerging Asia faces similar
challenges in administering healthcare, but perhaps to a less severe degree than China.
The problems are familiar, so we highlight a few below but won't dwell on them here.
Costs are spiraling up medical inflation in Asia is among the highest in the world,
according to Aon Hewitt.38 Mercer Marsh Benefits expects medical costs to increase by
11.5% in 2017 (led by Vietnam (19.3%), Malaysia (17.3%), Indonesia (11.8%), and China
(11.5%).39 This is driven by cancer and respiratory disease incidence as well as negative
lifestyle and environmental influences such as pollution, smoking, and obesity. Low-
quality and counterfeit drugs are rampant. The general public no longer trusts "average"
doctors and believes they need to see the best specialists for care. Most public hospitals
are highly inefficient (both financially and operationally), with many stakeholders
collecting economic rent without adding much value. The incentives for administrators
and doctors are sometimes misaligned with the best interests of patients. Corruption is
endemic most drugs and devices are sold with some sort of graft attached, and
overworked physicians are lowly paid and, hence, they depend on untaxed income ("gray
area") from patients and drug companies. Finally, and perhaps most worryingly, the
incidence of cancer and chronic diseases is disproportionately high. China is home to
20% of the world's population but has 22% of new cancer cases and 27% of all cancer
deaths. Half of the world's Hepatitis B patients are in China.
We think that disruptive new technology may be the only way to move the needle
meaningfully toward solving some of these problems without dramatically increasing
healthcare expenditure or the number of doctors. Most of the problems dogging
healthcare in China relate to inefficiency, low quality of "products" and "services"
(medicines and care), asymmetry of information, corruption, and inequitable access to
care. Technology is not a panacea, but has solved similar challenges in other fields.
38
Aon Hewitt 2015 Global Medical Trend Rate Survey Report.
39
Mercer's "Trends Around the World" report (based on a survey of 171 insurers across 49 countries).
EXHIBIT 67: China's healthcare system has come a long way, but still faces many (seemingly intractable) problems
Note: Lower left-corner image is Bernstein Asia-Pacific Healthcare analyst wearing Toshiba Vital Images virtual reality goggles (development collaboration with
Microsoft HoloLens technology for surgical and medical education applications).
Vinod Khosla, founder of Sun Microsystems, famously said in 2012 that machines will
replace 80% of doctors in a future healthcare paradigm driven by technology companies,
entrepreneurs, and super-computers.40 We disagree with this vision AI can make good
doctors better (by arming them with pattern recognition across larger datasets than any
human can see across) and can handle some routine and repetitive aspects of what
doctors do, but will not be able to replace the human elements of medicine. Doctors
encounter things every day that don't conform to previously known patterns, and have to
make judgments and decisions accordingly.
It is also worth pointing out that the term "AI" is overused and misused (not just in
healthcare but everywhere); bigger computing power across bigger datasets is not the
same as AI. AI means computers that possess cognitive abilities, such as learning,
problem-solving, and prioritizing information like humans. Computers are not yet humans
(or even close), but in narrow circumstances can analyze huge amounts of data, learn
rules for how to make decisions and judgments, and improve over time with experience.
We are at the very beginning of AI in healthcare; below, we highlight a few companies that
are starting to show potential.
Imaging and diagnostics. The most crowded focus area of AI in healthcare is imaging
and diagnostics. One good example is Imagen Technologies, a NYC-based start-up
founded in 2015, which raised US$14.3 million in a new round of equity financing in
March 2017. The company develops AI-based software to reduce diagnostic errors
by detecting clinically meaningful pathologies within medical images. Another to
watch in the United States is Butterfly Networks. The initial focus of this three-year
old company is on "making medical imaging accessible to everyone in the world."
Basically, it reinvents the ultrasound machine by squeezing all of its components
onto a single silicon chip and develops new ultrasound scanners that can create 3-D
images in real time and get connected to smartphones. The system will leverage AI
technology to compile a bank of images and extract key features that will automate
diagnoses. The firm has recently financed US$100 million in a Series C round. Skin
scan assesses skin cancer risk from photographs of patients' skin lesions and moles
submitted by smartphones.
40
Khosla V 'Do we need doctors or algorighms' (2012) TechCrunch, January 10 (part of a guest series of articles authored by
legendary tech entrepreneurs).
information now being added to what it can analyze (e.g., genomics) see Exhibit 69
and Exhibit 70. In Asia, Watson has joined hands with a few leading players, including
Bumrungrad (BH.TB, market-perform), 21 public hospitals in China, and most
recently with Baheal Pharma in China. Bumrungrad was the first hospital to put
Watson For Oncology into clinical practice to generate treatment plan
recommendations (the first patient was in 4Q:15). The Watson Oncology project
costs Bumrungrad THB40 million a year (~US$1 million). In China, 21 public
hospitals are using Watson, and Baheal Pharma has bought three-year exclusive
China commercial rights to sell Watson For Oncology this March (2017). Watson For
Oncology is not without setbacks, however. MD Anderson, the early partner of
Watson, declared in a 2013 press release that the University of Texas Cancer Center
"is using the IBM Watson cognitive computing system for its mission to eradicate
cancer," but has since put the project on hold after an audit report found that despite
spending US$62 million on it, the project had fallen short of its goals. MD Anderson is
now reportedly soliciting bids from companies selling competing technologies.
Human Longevity, set up by Craig Venter (one of the fathers of the Human Genome
Project), is building an algorithm that can design patients' physical characteristics
based on their DNA. It offers patients complete genome sequencing, coupled with
body scanning, using multiple imaging modalities for an ultra-comprehensive AI-
driven body check to detect cancer and vascular diseases very early.
EXHIBIT 69: IBM Watson For Oncology helps oncologists EXHIBIT 70: It was first "trained" by Memorial Sloan
make better treatment pathway decisions by helping them Kettering, then by MD Anderson and Mayo Clinic
see patterns in patient outcomes across larger datasets than
humans can observe
Source: IBM slide from Bumrungrad's presentations about IBM Watson For Source: IBM slide from Bumrungrad's presentations about IBM Watson For
Oncology (reproduced with permission). Oncology (reproduced with permission).
Mining medical and health insurance claims records have been the focus area of
Google DeepMind Health project, which uses AI to mine huge numbers of medical
records to provide better, faster health services. Related to this, AI is being
increasingly used in the health insurance industry to detect and prevent fraudulent
claims.
Assisting repetitive tasks is another focus area of IBM, which launched another AI
healthcare product called Medical Sieve. It is a "cognitive assistant" with analytical
and reasoning capabilities and a wide range of clinical knowledge (not just oncology)
Remote patient monitoring can be another huge area to focus. U.S.-based AiCure has
raised US$12 million in a Series A round to develop an AI-enabled product to ensure
patients are taking their medications. Sentrian uses AI to analyze biosensor data and
send patient-specific medical alerts to doctors.
Medical assistant chatbots. U.K. start-up Babylon Health raised US$25 million in the
Series A round (backed by Google DeepMind) to develop an AI-based chat platform.
Babylon Health partnered with U.K.-based National Health Service (NHS) to replace
its telephone helpline with AI-driven video and text consultations. U.S.-based
Sense.ly developed a virtual nursing assistant named Molly to follow up with patients
after they are discharged from the hospital (Sense.ly claims that Molly can free up
20% of doctors' time. Other emerging companies in this space include HealthTap,
Koko, and MD.
Risk analysis. For example, Ayasdi, a U.S.-based broad AI start-up with a healthcare
vertical, is developing a platform to score patient risk and to reduce readmissions.
Drug discovery. Pharma companies, large and small, are increasingly using AI-
enabled tools to improve success rates in drug discovery, speed up the lengthy
timelines, find novel targets, and to find better ways to identify high-responding
cohorts of patients. For example, Pfizer is working with IBM Watson for Drug
Discovery for novel target identification, combination therapy development paths,
and patient selection strategies in immune-oncology. Many start-ups in the United
States have been active in this space. For example, TwoXAR has recently developed
the DUMA drug discovery platform that uses AI to predict and rank novel drug
candidates. U.S.-based Atomwise published its first findings of Ebola treatment
drugs last year, and has also partnered with Merck.
EXHIBIT 71: IBM's Medical Sieve analyzes routine radiology and cardiology images
China is gearing up in the competitive AI race, and may pull ahead in some areas. AI is
considered as an innovative tool by the government, providers, and payers to solve the
conflicts rooted in China's healthcare system and balance medical resources properly. As
early as 2003, the first project on big data analysis on medical images has been initiated
in China, though finally terminated due to data inconsistency. Now, with the more
advanced AI technology, the industry has regained confidence to manage the information
efficiently and help patients gain access to high-quality medical resources across the
nation. The central government also released a series of supportive policies to boost its
development as well as invest in big data epidemiological upgrades, Precision Medicine
Initiative, digitization of all medical records, and informatics and data mining companies,
demanding a >50% telemedicine coverage of all pilot regions on tiered medical
treatment.
In addition to the efforts by the leading Internet giants, many start-ups have emerged to
capture the tide of AI. One "unicorn" among China's medical AI service-oriented players is
iCarbonX. Established by the former CEO of genome sequencing giant Beijing Genomic
Institute (BGI) in October 2015, the Shenzhen firm hopes to provide individualized
analysis and prediction of the health index through advanced data mining and machine
analysis technologies. The goal is to create the first intelligent robot that can provide a
personalized health management system to users. iCarbonX has formed a global alliance
with seven technology companies (PatientsLikeMe, SomaLogic, HealthTell, AOBiome,
GALT, Imagu, and Robustnique) to gather different types of healthcare data. It received a
RMB1 billion (US$155 million) Series A funding led by Tencent (700 HK), and now its
valuation has increased to RMB6.5 billion (US$1 billion), making it one of the only three
healthcare start-ups in China with a US$1 billion-plus valuation. Founder and CEO Jun
Wang believes that iCarbonX can mine healthcare data faster than global rivals because
of its access to the largest population, its partners, and generous funding (more than
US$600 million raised so far). They have also created a healthcare app called Meum (see
Exhibit 72) that crunches enormous amount of genomic, physiological, and behavioral
data to offer personalized medical and lifestyle (what to eat and when to sleep) advice
directly to consumers.41
Synyi is a Chinese machine-learning start-up that uses AI to better organize and analyze
clinical history. Its solutions automatically translate accumulated clinical history reports in
hospitals into structured databases, and detect medical variables within written reports to
generate standardized data. Clients include Beijing University hospital and Fudan
University Children's Hospital. Its founder, Zhang Shaodian, PhD, a former Microsoft
engineer, is only 28 years old (and has been recently named in Forbes 30 Under 30 Asia).
There are too many start-up companies in this space to discuss all of them. Beijing-based
Yidu Cloud, founded in 2013, is developing a platform to make better use of health data
with advanced machine-learning techniques, enabling patients to benefit from expert
advices gained from massive medical data. It has recently raised RMB200 million in the
Series A round. Another player is a Hangzhou-based firm, JP Technology. It combines the
key big data analysis tool (MIBA) with the AI technology to build an intelligent diagnostics
system via deep mining on the medical images database, thus helping physicians locate
disease areas and analyze conditions more accurately. The firm was established in 2012
and has raised RMB50 million funds so far.
41
Cyranoski D "Chinese AI company plans to mine health data faster than rivals" (2017) Nature Biotechnology 541:141-142.
EXHIBIT 72: iCarbonX founder and CEO Jun Wang talks about EXHIBIT 73: Baidu's medical chatbot Melody
the Meum health app
Source: Nature Biotechnology, company website, and Bernstein analysis. Source: topbots.com.
Fifteen years later, precision medicine around the world kicked off the idea to
customize healthcare and provide tailored solutions to patients based on the context of
genomic (as well as proteomic, metabolomics, and microbiomic data). Then U.S. President
Obama launched a national Precision Medicine Initiative in 2015 as an "all of government"
effort to bring healthcare into the next level. It includes a series of plans such as collecting
data from more than 1 million of U.S. individuals and developing new regulatory
approaches for evaluating next-generation genomic sequencing technologies, and also
refining relevant policies to ensure patients' privacy and security. The 100,000 Genomes
Project, launched by the U.K. government around the same time and run by a state-owned
company called Genomics England, is aiming to sequence 100,000 whole genomes of
patients in the National Health Service's records. The goal is to match genomic and
clinical data to develop personalized therapies for cancer and rare diseases and to turn
the NHS into "the first mainstream health service in the world to offer genomic medicine
as part of routine care". The government has committed to sequencing 100,000 whole
human genomes, from 70,000 patients, by the end of 2017. The first child received a
genetic diagnosis through the 100,000 Genomes Project in January 2016.
China launched its own Precision Medicine Initiative about a year after the United States
did (in March 2016), on a grand and more ambitious scale. It is a US$9.2 billion, 15-year
project seeking to sequence the genomes of 100 million Chinese citizens and to
assemble the world's largest database of genomics and use it to improve disease
diagnosis, develop better drugs, and provide personalized scientific wellness regimes.
This is a signal to show China government's ambition to establish itself as a real
competitive force in the genomics-enabled technology field. A large number of
government institutes, technology companies, pharmaceutical companies, genomics and
informatics companies are part of this initiative. At that time, it got extensive media
coverage for being so much larger than Obama's Precision Medicine Initiative, which has
been announced as a US$215 million project. The majority of that amount was committed
as additional federal funding for the National Institutes for Health. However, this is not an
apples-to-apples comparison to China's much larger US$9.2 billion, which refers to all of
the public and private sector money that may be spent over the next 15 years. Regardless,
both projects are highly important.
Historically, the cost of whole genome sequencing has been prohibitively high. Next-
generation sequencing is high throughput DNA sequencing of millions (or even billions) of
strands of DNA in parallel, which decreases the need for fragment-cloning methods used
historically. This reduces cost and increases speed. According to the National Human
Genome Research Institute, the cost of sequencing one genome has fallen by six orders
of magnitude since 2001. The time required has fallen from months to hours. There are
six primary methods for next-generation sequencing, each with varying ability to read
DNA base pairs and different times required to perform the task:
Personalized medicine also means personalized coverage. IDC predicts that by 2021,
20% of the private and commercial health insurance payers in China will offer customers
personalized coverage schemes based on individual health data as a way to dynamically
reduce fee rates and lower patient-paid costs. Although there are many ethical issues
associated with this that remain unresolved, the window to a new era has already been
opened.
Among many players stands out BGI, the world's largest genomics organization founded
in 1999 with the vision of using genomics to benefit the human race. It develops a
systematic infrastructure to store, decode, interpret, and apply genetic information, and
currently owns around 295 sequencers in total. Headquartered in Shenzhen, BGI has
branches in over 50 countries and established 47 labs (including joint labs), with one-third
located overseas. It has also formed an alliance with leading IT giants, including Intel and
Alibaba to allow data analysis, visualization, and reporting conducted on the cloud to help
researchers gain insights quickly and easily. BGI also initiated China's first national gene
bank in 2011, aiming to collect and preserve genetic resources and building a network to
foster global communication. As of December 2015, the bank preserved thousands of
genomes, representing 80% of the finished large genome projects in the world. Currently,
the key business of BGI is to provide gene sequencing services to academic institutions,
medical facilities, and individuals, and the largest revenue contributor (~54% of total
revenue) of BGI is the Non-Invasive Prenatal Testing (NIPT), which looks at DNA from the
sample of a mother's blood to identify whether there is a risk of giving birth to a child with
a genetic disorder. BGI submitted the IPO application in Shenzhen Stock Exchange in late
2015 while still waiting for the final approval. Some people doubt whether it is a truly
innovative firm or only temporarily leads the market benefiting from its large-scale
database, while the management has strong confidence to claim its advanced technology
and consider BGI as "next Tencent" in the genomic area.
Berry Genomics is considered the strongest competitor of BGI in the market, especially in
prenatal genetic testing area. Founded in 2010, Berry Genomics started with noninvasive
prenatal DNA screening, and later on expanded to DNA sequencing and other forms of
genetic testing. The founding members all have experiences working in BGI before.
Backed by Qiming Venture Partners and Legend Capital, Berry Genomics is planning an
IPO via a reverse merger with Shenzhen-listed Chengdu Tianxing Instrument & Meter,
considered widely as an attempt to access capital markets in an easier way.
WuXi NextCODE Genomics is a subsidiary of WuXi AppTec, the leading CRO firm in China.
Wuxi App announced to acquire NextCODE Health for US$65 million in cash in 2015.
WuXi merged NextCODE and its own Genome Center (established in 2011) into a new
company and named it WuXi NextCODE Genomics. The business is headquartered in
Shanghai, with operations in Cambridge, Massachusetts, and Reykjavik (Iceland). WuXi
NextCODE offers comprehensive services, including study design, sequencing,
secondary analysis, storage, and interpretation and scalable analytics by organizing,
mining, and sharing genome sequence data. It has also been selected as the official
genomics platform by several national plans, including China's Precision Medicine
Initiative and Singapore's Precision Medicine Effort.
What is CRISPR?
CRISPR-Cas9 makes targeted genetic editing powerful, cheap (around US$30 per
sequence), precise, and relatively easy. It enables scientists to precisely "cut and paste"
DNA into almost any genetic sequence and in almost any living thing. The name refers to
the way that short, repeated DNA sequences in the genomics of bacteria and other
microorganism are organized and how they fight off attacks by viruses. Bacteria defend
themselves from viruses infecting them by "stealing" strips of the invading virus's DNA
and splice in their own using an enzyme called Cas9. The bacteria then make RNA copies
of these "stolen" bacterial genetic sequences (what science writer refers to as a
"molecular most wanted gallery"), which can then be used to recognize viral DNA and
prevent infection. Harnessing this bacterial defense mechanism and making it into a gene
editing tool is one of the most important biotechnology advances of the last decade (some
have said of the century). Two groups of scientists in parallel figured out a way to replace
the bacterial RNA system with a modified guide DNA that tells a "hunter" Cas9 enzyme
where to look. The enzyme scans the cell's genome to find a match of DNA and then slices
for the DNA in the cell's enzymes. This helps to easily make changes or add DNA within
the cell. Exhibit 74 presents a simplified schematic for how CRISPR-Cas9 gene editing
works.
The intellectual property fight that ensued from two different groups making the same
discovery in parallel has been one of the ugliest in the field of science. On one side of the
battle is Dr. Jennifer Doudna (UC Berkeley) and her collaborator Dr. Emmanuelle
Charpentier (then at Umea University in Sweden, now at Hemholtz Center for Infection
Research in Germany). In 2012, they published their first demonstration of CRISPR's
potential to precisely "cut" DNA of a microbe, at a location of the researcher's choosing.42
In 2014, they published a seminal paper demonstrating that they could cut a particular
piece of DNA in human cells and replace it with another one.43 Around the same time, a
team of scientists from Harvard University and the Broad Institute at MIT led by Dr. Feng
Zhang reported the same achievement (showing that a CRISPR gene editing tool could be
adapted to edit DNA in eukaryotic cells, such as plants, animals, and humans).44 Berkeley
filed for a patent earlier, but the Broad Institute filed for an expedited review process and
USPTO granted its patents first. The Berkeley team then asked USPTO to declare a patent
interference and filed for an investigation to establish who came up with the invention
first. The Broad Institute argued that Berkeley's initial patent filings described CRISPR-
Cas9 use in prokaryotes only (i.e., bacteria) and did not sufficiently describe the technique
in eukaryotes (an important distinction because CRISPR's most commercially valuable
potential applications are in medicine and agriculture). In February 2017, the Broad
Institute's patents were upheld. Millions (maybe billions) of dollars are at stake and the
fight is still not over Berkeley can appeal the ruling, European CRISPR patents are still
up for grabs, multiple other parties are also entering the fight claiming invention around
the same time, reports of other enzymes that can do the same thing Cas9 can are
multiplying (CasX and CasY), and CRISPR technology is already moving beyond what the
initial patents covered. The amount of CRISPR research since has exploded and
thousands of patents have been filed.
42
Jinek M et al "A programmable dual-RNA-guided DNA endonuclease in adaptive bacterial immunity" (2012) Science
337:816-821.
43
Doudna JA and Charpentier E "Genome editing. The new frontier of genome engineering with CRISPR-Cas9" (2014)
Science 346:6213.
44
Cong L et al "Multiplex genome engineering using CRISPR/Cas systems" (2013) Science 339:819-823.
Source: Adapted from Charpentier E and Doudna JA "Biotechnology: Rewriting a Genome" (2013) Nature Biotechnology 495: 50-51.
Chinese scientists were the first to report successful genetic editing in (non-viable)
human embryos (that were destroyed after three days) first, a team at Sun Yat Sen
University in 201545 and the second group at Third Affiliated Hospital of Guangzhou
46
Medical University in 2016. The latter, led by Dr. Yong Fan, was seeking to establish
proof of principle for what would be needed to generate an individual with resistance to
HIV. A firestorm of ethical debates has ensued a more robust regulatory framework for
how CRISPR technology can be used is needed; what will happen if we begin making
germline (i.e., inheritable) modifications in viable human embryos; how far away are we
from an underground market for "designer baby" services; and how do we prevent
CRISPR from getting into the wrong hands and from being used for more sinister uses
(i.e., create biological weapons selective for certain groups of people).
The potential applications of CRISPR are extremely widespread among the most
exciting therapeutic areas are agricultural and human therapeutic uses. In agriculture, it
45
Liang P et al "CRISPR/Cas9-mediated gene editing in human tripronuclear zygotes" (2015) Protein Cell 6(5):363-372.
46
Kang X et al "Introducing precise genetic modifications into human 3PN embryos by CRISPR/Cas-mediated genome
editing" (2016) J Assist Reprod Genet 33(5):581-588.
has been used to expedite crop and livestock breeding, engineer new antimicrobials, and
control disease-carrying insects with gene drives. Gene drive is a technique that
promotes the inheritance of a particular gene to increase its prevalence in a population (it
works only in sexually reproducing species, and cannot be used in viruses or bacteria). In
2015, a team from Imperial College London, University of Perugia in Italy, Cambridge
University, and University of Strasbourg reported a CRISPR gene drive that caused an
infertility mutation in female mosquitos to be passed on to all their offspring (though some
resistance also emerged),47 bringing the potential to eradicate species of mosquitos that
have long been vectors of human disease (malaria, Zika, dengue fever, etc.). In medicine,
the CRISPR's potential clinical applications are widespread and profound removing
mutations that bring higher risks of developing cancer or other diseases, slowing down
aging, and eradicating genetic diseases. The first clinical trials using CRISPR-Cas9 for
targeted cancer therapies began in 2016, both in China and the United States. More far-
fetched ideas have been discussed, including bringing woolly mammoths back to life.
Exhibit 75 summarizes the different ways CRISPR-Cas9 can accelerate drug discovery.
Many fast-growing and well-funded companies have emerged in the United States in the
last few years, trying to commercialize CRISPR uses. Notable listed ones include Intellia
Therapeutics (NTLA.US; co-founded by Dr. Jennifer Doudna), Editas Medicine (EDIT.US,
co-founded by Dr. Feng Zhang), and CRISPR Therapeutics (CRSP.US, co-founded by Dr.
Emmanuelle Charpentier). A long list of VC-funded start-ups is behind them (including
Synthego, Caribou Biosciences, eGenesis, Poseida Therapeutics, Benchling, Agenovir,
Exonics), some of them funded or co-founded by the above three CRISPR pioneers.
Source: Adapted from Fellmann et al "Cornerstones of CRISPR-Cas in drug discovery and therapy" (2017) 16:89-100.
47
Hammond A et al "A CRISPR-Cas9 gene drive system targeting female reproduction in the malaria mosquito vector
Anopheles gambiae" (2015) Nature Biotechnology 34(1):78-83.
So far, the gene editing breakthrough work in Asia has been in academic institutions in
China. We expect start-up companies to be established around these technologies, but it
hasn't happened yet.
(4) TELEMEDICINE AND ONLINE What is telemedicine and online prescription drug sales?
PRESCRIPTION DRUG SALES
A precise and widely agreed upon definition of the term "telemedicine" is difficult to pin
down. It means the remote delivery of clinical services to patients using technology
medical consultation and treatment when the healthcare provider is not in the same room
as the patient. In our report published in September 2016,48 we illustrated the whole
industry landscape in China, categorizing the content into four distinct types, and based
our analysis on these buckets remote monitoring of patients, asynchronous
consultation, real-time remote consultation, and online prescription drug sales.
Remote patient monitoring means tools and programs that enable doctors and nurses to
remotely monitor signals from patients in their daily lives, usually via wearable or
implanted devices, sensors around the home, or other technologies. Asynchronous
consultation, also known as "store and forward," means storing medical data and sending
it to a doctor in a remote location for evaluation. Real-time remote consultation is
interactions between a patient and a doctor or nurse via live video connection. The
interactions can include consultation, diagnosis, treatment, and prescription. Online
prescription drug sales refer to the service that allow patients to purchase prescribed
drugs in online stores instead of going to the hospitals/pharmacies themselves, making it
more convenient.
In China, the case may be different telemedicine in China could really take off and
widely transform how healthcare is consumed in ways that many have hoped for in other
countries (but haven't really materialized at scale). Why? The problems in China's health
system are deeper and more difficult to solve bigger constraints and infrastructure
48
See September 20, 2016 report: The Long View: Healthcare in China Part II - The Future of Telemedicine.
The business adoption of telemedicine in China can be generally broken into primary and
specialty care and include both primary care and specialty care. Among the many players,
we think that Good Doctor has the highest potential to transform how medical care is
delivered. Backed by Ping An insurance company, it serves as an Online to Offline (O2O)
healthcare service platform to provide users 24-hour one-on-one consultation with
doctors in all medical divisions. They can get a quick diagnosis and treatment and also
register for online appointments with doctors. Additionally, they can purchase other
healthcare products. Ping An Good Doctor has 130 million patient users and 300,000
doctors registered, and employs more than 1,000 full-time physicians to cope with the
250,000 online medical consultation requests it receives every day. Local media reports
put its 2015 revenue at RMB300 million (US$45 million). In May 2016, it raised US$500
million in its first round of external financing, valuing it at US$3 billion (and is fundraising
again now).
EXHIBIT 76: Ping An Good Doctor overview EXHIBIT 77: Ping An Good Doctor vision
One of the largest rivals of Good Doctor is Tencent-backed We Doctor Group (previously
known as Guahao.com). It operates an online platform for users to check doctor profiles
and ratings and make online appointment registrations. We Doctor currently has covered
physicians from 2,400+ hospitals nationally and owned more than 150 million registered
users and has raised US$520.5 million funds in previous three rounds. In December
2015, it unprecedentedly launched the country's first Internet-connected hospital in
Wuzhen, Zhejiang province, and currently owns a professional team of more than
260,000 doctors and over 7,400 medical teams across the nation. Within the innovative
Internet-based hospital, patients and physicians can communicate via various ways,
including real-time messages, phone calls, and video conversations. People in remote
areas can also enjoy high-quality services with such practices. In February 2016, We
Doctor's maternity business merged with Shanghai-based maternity-focused app
developer Beilian Information Technology to build an Internet-based maternity hospital,
backed by We Doctor's online and mobile consultation platform Wuzhen Internet Hospital.
Users of the platform can register for appointments online, and do video, image, and voice
healthcare consultations online.
Unlike in the United States, online retailing of prescription drugs is prohibited in China, but
many domestic pharmacies are trying to link online and offline services. One example is
Universal Health International (2211.HK), which had a network of more than 900
pharmacies in four provinces (which is still only a one-tenth of the size of pharmacy
network of CVS with almost 10,000 outlets). China's second-largest drug distributor
Shanghai Pharmaceuticals (2607.HK, 601607.CH) may be more successful. It has been
buying up retail pharmacies around China and incorporating them into its branded chain
(Huashi Pharmacy). In May 2015, the company announced a partnership with Internet
retailer JD.com (JD.US) to sell OTC drugs and healthcare products more broadly (and,
hopefully, prescription drugs in the future) via an online platform. In May 2016, it raised
another US$21 million (from Softbank China Venture Capital and a few other funds),
bringing the total fundraising for its online pharmacy business up to US$207 million.
China's third-largest distributor Jointown (6008998.CH) is also making large investments
in online pharmacy. Ren Bin and 111.com.cn are examples of emerging companies
entering this space with apps that can deliver non-prescription drugs to people's
doorstep after receiving orders online. Both are angling to be well positioned for
prescription drugs too.
Separate from drug sales via Tmall and TaoBao, Alibaba Health Information Technology
(241.HK) launched a pilot online prescription pharmacy pilot program that it argues is the
future of healthcare in China. The pilot was launched in late January 2016 and brought
telemedicine and online drug ordering to a remote village called Hongshiyuchang in Hubei
province. The program allows people to order drugs through an online video call with a
physician, who writes a prescription, which can be filled at an online pharmacy, and
delivered the next day (patients pay cash to the delivery person).
Augmented reality is a technology that adds value to the real world by overlaying and
displaying real-time digital information and media, such as videos and 3D models, via the
camera view of your smartphone, tablet, and PC, or via wearable technology such as
goggles or smart glasses. Augmented reality became ubiquitous because of its gaming
uses (look at the success of Pokemon Go on the day it was launched, it surpassed daily
usage time for Facebook, SnapChat, and Twitter by average iOS users on mobile phones).
However, it is not just for games it also has many potentially disruptive uses in
healthcare.
Surgery: The use of augmented reality goggles in surgery will allow surgeons a higher
level of precision and efficiency than that of the naked human eye. It will also enable
doctors to prepare for surgeries in advance and seek advice from more experienced
surgeons during a procedure. Medsights Tech is developing a software that uses
augmented reality in order to create accurate 3D reconstructions of tumors to assist
in surgeries. Vital Images (a Toshiba Medical Systems Group company) is developing
multiple medical applications for Microsoft's HoloLens technology for guided surgery
(including telemedicine surgery in remote areas).
Medical education: Vital Images and HoloAnatomy are both Microsoft's HoloLens
technology to display educational data-analytical models, highlighting one tissue
type when looking at an actual human (see Exhibit 79).
Vein scanners can help nurses find the best veins for drawing blood and IV lines
easier. AccuVein is developing such a technology, and claims that 40% of IV
injections miss on the first needle stick, and that this number is much worse for
children and the elderly.
EXHIBIT 78: Automated external defibrillator locator EXHIBIT 79: Microsoft HoloLens Augmented reality goggles
in medicine
There are many companies developing augmented reality gaming applications in Asia, but
few are focusing on medical applications (so far). We expect this to change over time.
IOT adoption in healthcare is about remote monitoring and diagnostics via next-
generation wearable devices and smart sensors in homes, cars, workplaces, and
hospitals. Third-party market research firms have estimated that the healthcare IOT
market segment is poised to hit US$117 billion by 2020. The new technology can not only
help patients to manage disease better, but also facilitate physicians to monitor a patient's
condition efficiently and improve care. Wearable devices and home health monitoring are
the two major buckets. Wearables have so far been more consumer health and fitness
oriented (steps, sleep, vital signs), but a tidal wave of wearables that are more medical in
nature is coming.
Home health monitoring connects caregivers and healthcare professionals with patients
in their own daily lives at home. The physiological data such as movement, blood pressure,
heart rate, and pulse can be recorded by sensors on peripheral devices and then
transmitted to families of patients or physicians via wireless communication devices and
sends alert when problems are detected. This is useful for the elderly and chronic disease
patients. The Commonwealth Scientific and Industrial Research Organization (CSIRO) has
developed a series of solutions to improve community care through "smarter safer homes
(SSHs)" (see Exhibit 80 and Exhibit 81). The SSH solution provides seamless connection
between the elderly, their families/relatives, and physicians by placing simple sensors
around the house. These sensors can monitor an elderly person as they go about their day
work and report the data back to family members or physicians. Caregivers can use
e-tablet at home to check activities and even make video calls when needed (see Exhibit
82).
EXHIBIT 80: What does a "smarter safer home" look like? EXHIBIT 81: The working model of the SSH solution
Other innovations in this area include at-home portable diagnostics, such as the
electrocardiogram (ECG) machine developed by AliveCOR. The device is embedded in a
smartphone that helps interpret ECG results via an app and shares data with clinicians.
There are also assistive devices, such as tremor spoons for people with Parkinson's, that
incorporate sensors and deploy analytical expertise to observe how people's disease
state changes over time.
There is no doubt about the popularity of wearable health technology, especially during
the recent days in China. From wristband step to smart glasses and heart-rate trackers,
this technology is gradually changing the way people manage their health. The iHealth Lab
a subsidiary of Tianjin-based Andon Health (002432 SH), which is one of the largest
healthcare wearable devices manufactures in China is providing a series of wireless
wearable devices to measure the key vital signs, including the previously launched
ambulatory blood pressure monitor that can be worn inside a vest. The monitor provides
continuous blood pressure monitoring and can be set to read at specific time intervals. It
can be connected to the users' smartphones or computers. Another Chinese player
Sinocare (300298 SZ), has developed a mobile glucose meter "D-Nurse." It is a small
device that can be connected to mobile phones. Users only need to insert the strip to the
device and collect the blood sample, and then the result will display on the screen.
The local information system provider, Neusoft (600718 SH), has launched a home
monitoring machine in the market and has received positive feedback so far. Their home
monitoring machine integrates the regional community medical centers and patients via
health IOT efficiently. The remote monitoring devices, including Xikang household
healthcare terminals and Xikang healthcare watches, can monitor and collect data, and
then share with regional community health centers. This system not only offers real-time
track for individual user, but also helps establish a scientific and systemic networking
management platform for hospitals and health administrative departments.
Blockchains are centralized databases that keep a record of how data is created and
changed over time. These can be trusted as authoritative records. The idea behind
creating these blockchains is to have a single, impartial authority that can guarantee
accuracy and security of many unstructured transactions. By design, blockchains are
inherently resistant to modification of data; once recorded, the data in a block cannot be
altered retroactively. The blockchain was first created in 2008 and initially used mostly for
BitCoin digital currency interactions, but has potential uses in a wide range of industries
as well.
Recently, many healthcare uses of blockchains have opened up, particularly in the United
States. Electronic health records (EHRs) and electronic medical records (EMRs) have
been widely used now and they are usually centralized, provided by a small number of
suppliers. Some argue that using the blockchain technology to better manage medical
records will bring greater resilience, interoperability, and security (physicians and patients
are given encryption keys to control who can see their data). In July 2016, the U.S.
government (Office of the National Coordinator for Health Information Technology (ONC))
launched a contest called "blockchain and its emerging role in healthcare and health-
related research," soliciting white papers on the topic of blockchain technology and its
potential use in healthcare.
documents. The blockchains can help solve the issue by setting up a system to store
medical information at the very basic level. Since the information is recorded step by step
and separately by different hospitals, it will be difficult for fraudsters to fabricate
information. Also, the patient data solutions can be easily linked with systems of other
private health insurance companies, making it difficult to use fabricated data for
fraudulent claims.
ZhongAn is the first online insurance company in China to implement blockchains into its
online health insurance platform. The firm has partnered with and will be using Ethereum's
blockchain in its platform for smart contracts benefit and fraud claims prevention. Zhong
An is setting up a cloud open blockchain platform known as Anlink and encourages other
companies to help develop and grow the market with their own ideas (healthcare and
other applications).
"Smart pills" are tablets and capsules enabled with ingestible sensors or other
technologies that help them do more than just deliver medication into the body.
Smart pills can help doctors and patients monitor adherence, measure things inside the
patient's body, and tackle counterfeit medicines. Medication adherence is a problem
everywhere in the world. A U.S.-based company Proteus Digital Health developed a new
product called "Proteus Discover," which has an innovative sensor that is so small that it
can be swallowed and combined with drugs in the form of a pill. Once the sensor reaches
the stomach, it is then activated and starts to send data to the mobile phone and also to
the Proteus cloud with patients' permissions. The system also includes a small wearable
patch that can help monitor and share patients' activities. The treatment now is
undergoing FDA NDA review, which combines the technology with a second-generation
anti-psychotic drug, Abilify (aripiprazole). The firm has secured US$50 million in Series H
financing in August 2016.
China lags the United States in this space, but stands to potentially benefit more due to
the large patient base and unprecedented ubiquity of counterfeit medicine. Wuxi Pharma
Tech announced to form an alliance with TruTag in 2014 and also included an investment
on TruTag by the Wuxi Venture Fund. They later announced a successful completion of
joint testing of TruTag's on-dose authentication solution in August 2015. TruTag is an
Asia-facing U.S. firm that specializes in providing product authentication and brand
protection solutions for a series of industries, including healthcare. It allows a digital scan
to confirm the drug's authenticity and other information, such as manufacturing location
and supply chain data. The management teams from both companies were excited about
the positive results and planned to test on additional API products as well.
EXHIBIT 83: TruTag converts high-purity silicon wafers into EXHIBIT 84: to coat the surface of pills with invisible
nano porous silica micro-particles (silicon dioxide - SiO2) encoded information, without altering the current
that are safe, inert, edible, and cheap manufacturing process
EXHIBIT 85: Si02 microparticles encode covert optical EXHIBIT 86: Massive combinations of unique codes means
patterns that can be read with a detector the surface of each pill can be encoded with batch numbers
and the entire package insert
Let's image that a diabetes patient experienced complications and went for treatment in
hospitals. He received several ideas from physicians, but not so sure which was best. He
had physicians' suggestions but was also curious about other patients' choices. It may
have been difficult for him to get the answer quickly, but now, all he needs to do is just
open the website of an online diabetes patient community and search the symptoms to
find the suggested therapies all in one minute. The online communities today serve as
efficient communication tools for physicians and patients to share ideas anytime and
anywhere, and even help track health data and facilitate the research and development
efforts.
While it is not a new idea, we still see accelerating growth in connected communities for
doctors, patients, and caregivers. In developed countries such as the United States, the
social network for patients such as MedHelp, PatientsLikeMe, and HealthUnlocked not
only serve as online support groups and information repositories, but also act as platforms
for tracking health data and contributing research. For example, PatientsLikeMe, originally
designed for ALS patients while tapping into other disease areas afterward, conducts
scientific research studies for pharmaceutical companies, leveraging the patients' data
collected online. It entered into a partnership with Novartis (NOVN US) in 2011 to study
the barriers faced by multiple sclerosis (MS) patients in being adherent to taking their
medication. Finally, it led to the development of an MS Treatment Adherence
Questionnaire (MS-TAQ) that helps patients and physicians identify and address these
issues through enhanced communication. It helps to improve understanding of disease
and generate ideas to identify new approaches to disease management.
The scale of these online communities of doctors and patients in Asia is unprecedented.
In China, the unbalanced medical resources triggers patients' desire to share experiences
online and seek others' suggestions as well. Baidu Tieba, the online community initiated by
Baidu that cover almost all topics, has been used widely among patients with certain
disease and their relatives to share ideas and seek medical suggestions. There are also
specific online patients communities targeting a certain disease, such as tnbz.com, the
biggest online diabetes patient community. It is a subsidiary of Beijing Tangjian and the
parent firm also manufactures blood glucose meters and test strips as its business lines.
Wuxi AppTec launched an online platform dedicated to rare diseases, Boshi 360 (or
Doctor 360, www.boshi360.com) in 2014. The platform is designed to be a central portal
to benefit the 16.8 million Chinese patients diagnosed with rare diseases by providing a
one-stop service for patients from medical information to diagnostics and treatment
options, and access to patient supporting communities. As a leading R&D-driven industry
organization, Wuxi AppTec also aims to leverage its global resources to accelerate rare
disease R&D, and bring more effective treatments closer to patients.
China's DXY, founded in 2000 in Hangzhou, is the most popular and the largest online
medical portal, offering a series of online services to professionals in medical,
pharmaceutical, and bioscience sectors. It has more than two million doctors on its
platform, over four million registered members, and an average of 1.8 million daily page
views. Physicians can gain insight from colleagues, discuss new clinical findings, report
usual events, and even seek recruitment information online. DXY has also opened offline
clinics to provide medical services on common diseases in collaboration with some public
hospitals on patient referral. It reached a millions of dollars Series B financing and
received a US$70 million investment led by Tencent in 2014. The two parties have set up
a series of collaboration plans, including the wide adoption of WeChat system on daily
operations. There are a number of other, smaller plans, including GBI's JiBo and MedLive.
India's Practo has a large network platform that seeks to help patients locate doctors,
healthcare information, and similar patients. It has recently closed a US$55 million Series
D financing round. Investors include China's Tencent, which led their US$90 million Series
C financing. Japan's M3 (2413.JP) has the largest online physician community in Japan
(and has been expanding to other markets including China), among other online medical
businesses.
Automation can improve quality, efficiency, and consistency in the areas of recording
medical data, order entry, medication dispensing, central lab execution, cleaning and
sanitation, imaging (self-service imaging kiosks), and even surgery. The da Vinci Surgical
System, developed by the U.S. firm Intuitive Surgical, is probably the most famous robotic
surgery system (though there are now many others). It facilitates complex surgeries using
a minimally invasive approach. The da Vinci system was first launched in China in 2006.
Currently, there are more than 50 da Vinci systems running in Chinese hospitals, and
together they have completed more than 27,000 invasive surgeries.
IDC China expects Chinese hospitals to adopt 20% more robots by 2020, mostly to
deliver medical equipment and food.49 Although currently the market is dominated by
foreign players, more and more local firms are trying to develop their own products and
seize the market. One example is Phecda Robotic System developed by Beijing Tinavi. It is
designed as an orthopedic surgical robot to facilitate physicians during surgeries. It has
been successfully tested in some leading Class III hospitals in Beijing and is now pending
for CFDA launch approval. Other Chinese companies to highlight include Harbin Boshi
Automation (002698 SZ), SIASUN Robot & Automation (300024 SH), Jinshan Science &
49
IDC FutureScape: Top10 Predictions of Healthcare IT 2017 China Implications.
Automation can also be widely applied to other parts of a hospital, such as the central lab
in IHH's (IHH.MK, outperform) recently opened Gleneagles Hong Kong hospital (IHH.MK;
see Exhibit 87) and the in-house pharmacy in Bumrungrad's hospital (see Exhibit 88).
These automated services can significantly save labor costs and improve efficiency. For
example, with the new automated lab, a common case to get the testing result can be
shortened to within two hours, which benefits both patients and physicians. Thailand's
Ramkhamhaeng (RAM.TB) has also been investing in patient-facing robotics. It utilizes a
drug-allocating robot to dispense medicine for patients, decrease the risk of human error,
and increase efficiency.
EXHIBIT 87: IHH's brand new Gleneagles Hong Kong hospital EXHIBIT 88: Bumrungrad automated its in-house pharmacy
automated many parts of its central lab in 2008 (the first in Asia)
Hospitals are complex ecosystems with many clinical and operational processes. EMR
systems, many types of medical devices and equipment, and other systems are
independently connected to the Internet. Few hospitals have yet achieved full digitization
and interconnectivity between systems in which any doctor can see all test results and
ongoing monitoring and different monitoring equipment systems interact with each other
though this is where many are trying to get to. IBM's Digital Hospital Framework
includes a full set of tools from healthcare analytics and business management, to
electronic health records and care orchestration system. The tool has already been
adopted by a wide range of hospitals globally. Other leading players such as Siemens and
Samsung have also published their own digital hospital management systems to further
enhance patient-centric care and hospital efficiency.
We see more opportunities rising in China. The digital tools such as electronic patient
records and telemedicine services have already been widely used in developed markets,
while the adoption in China is still somewhat limited. The Chinese hospitals can't wait to
transform themselves to be more efficient on day-to-day operations. The central
government also considers it as a priority to push in the future, and the digitization of
healthcare delivery is listed as an important part of the Health China 2030 strategy
released in October 2016, which includes the wide usage of robust healthcare IT systems
in medical facilities.
Although there are many leading foreign participants launching their products in China
(IBM, Siemens, etc.), the market has been dominated by local firms, including Neusoft
(600718.CH), Shanghai Kingstar, and Tianjian Tech Group. They are highly competitive on
pricing and provide customized solutions in local language to fit Chinese systems. The
government also aims to promote local companies via more collaboration.
Founded in 1991, Neusoft is one of the leading software providers in China and the first
software player listed in Shenzhen stock exchange. Its "E-Hospital Solutions" offers a full
range of application systems for different departments, including HIS (information
systems for outpatient/inpatient, medical devices, medicines, logistics, general
management, and statistics and analysis), PACS/RIS (medical image storage, transfer and
diagnosis aid, and radiology information system), EMR (electronic medical record), LIS
(laboratory information system), remote diagnosis, consultation systems, and other
relevant tools. In 2013, Neusoft formed an alliance with the leading China telecom player,
Huawei, to establish a joint innovation laboratory on developing advanced digital hospital
solutions. Huawei provides ICT devices (switches, servers, etc.), while Neusoft provides
healthcare software (HIS, LIS, etc.).
The idea that games can be a tool to improve preventive health is rooted deeply in
psychology people are motivated to change their behaviors by a few core drives that
games can tap into:
Gamification in healthcare is a relatively new area, and in the past, most of the adoptions
have been focused on wellness and health management. Here, we discuss a few
examples (by no means exhaustive) of successful gamified health apps in the United
States. For example, people can track fitness progress via mobile-based apps and
compete with others online. Healthcare insurers are also experimenting with gaming.
UnitedHealth, a U.S. insurer, has recently developed an app called "OptimizeMe," which
allows people to participate in fitness-related contests with their friends. Besides that,
games have also been used on medical training, including online surgery practice games
and knowledge quiz. It can also improve patients' self-management by rewarding patients
with points for sticking to their medicine regimens via mobile video games. Although
gamification of healthcare is still in its infancy, it is considered to have huge potential to
amplify motivation and encourage productivity with fun.
Recently, there has also been a surge in serious games that address rehabilitation,
chronic disease management, and even medical research and analysis. Ayogo, a Canadian
firm, developed an app called Empower to manage chronic illness. The users are assigned
small activities/games and rewarded for positive changes. The educational content is
delivered in a fun context to motivate people to adopt healthy new habits and reach their
health goals. It received US$2.5 million in Series A round in 2015. Another great iPhone
app, Pain Squad, encourages young patients to track pain related to cancer and get
rewards for recording. This, in turn, has allowed doctors to gather and analyze data to
improve patients' overall care. This app is developed by the iOUCH research team at The
Hospital for Sick Children in Toronto, Canada.
During the game, the player tilts the ipad/iphone to steer an alien down a river and tap the
screen every time distractions appear (a fish or a bird, etc.). This performance data is then
collected and measured against the predetermined standards to make the diagnosis on
whether the user has the disease or not. Akki has added two pharma venture funds to
Series B financing in 2016, increasing round to US$42.4 million in total. It also joined
hands with Pfizer in 2014 to use the game as a clinical tool to determine early signs of
neurodegenerative disease pathology.
Mango Health has designed an app game in which patients can earn monetary rewards for
taking their medication. Patients earn points each time they take their meds and if they
stay on schedule, they can earn larger rewards of US$5-US$10 gift carts from retailers
such as Target and the Gap. Reflexxion Health seeks to reimagine physical therapy by
making it into a game that people can do with a virtual mentor in their own homes. Pact is
a health app developed by the founder of Guitar Hero. Its users make pacts with
themselves to consistently eat healthy food, exercise, and meet other health goals, and
then compete with others in adherence to these pacts to earn real dollars. Importantly,
patients can actually lose money for falling off rack (the app is connected to users PayPal
accounts), and losses for falling off track are much larger than gains for staying on track.
Asia lags the United States in the development of apps and technologies that gamify
healthcare, but we think is likely to catch up fast. The industry is moving fast and it has
been proven repeatedly that gamification of healthcare can help improve medicine
adherence and clinical outcomes, and provide meaning, timely data to clinicians.50,51,52
We expect many start-ups and wider adoption in China (and the rest of Asia, in general).
The early Chinese players are also trying to improve medicine adherence and enhance
health in a more funny and engaging way, although still in an early stage compared to
western countries. The Shenzhen-based Perfect Doctor recently developed a mobile
video game called "Asthma baby" to encourage kids with asthma to test vital capacity on
time. The test is designed as a small game and children can get points awarded for
recording data properly and they can also compete online with other users. The data is
then collected and assessed by physicians. It is currently seeking financing to realize
further expansion.
50
Larsen M and Crutzen R 'Gamification and adherence to web-based mental health interventions' (2016) JMIR Mental
Health 3(3):39.
51
Lumsden J et al 'Gamification of cognitive assessment and cognitive training: a systematic review of applications and
efficacy' (2016) JMIR Serious Games 4(2):11.
52
Dithmer M et al '"The Heart Game": using gamification as part of a tele-rehabilitation program for heart patients' (2016)
Games Health J 5(1):27-33.
Tan Tock Seng Hospital asks trainees to navigate a 3D virtual hospital to learn to better
identify infectious disease and how to use the correct protective equipment. It also uses
an app designed to help rehabilitate survivors of brain injuries via games, allowing doctors
and caregivers to monitor the patients' performance in the game.
EXHIBIT 89: Companies at the cutting edge of technologies that are transforming how healthcare is administered in Asia
have outperformed both MSCI Asia Ex Japan and MSCI Asia Ex Japan Healthcare
40.0%
30.0%
20.0%
10.0%
0.0%
-10.0%
-20.0%
Dec-15 Jan-16 Feb-16 Mar-16 Apr-16 May-16 Jun-16 Jul-16 Aug-16 Sep-16 Oct-16 Nov-16 Dec-16 Jan-17 Feb-17
Note: "Asian health tech innovators" line represents a basket of 35 listed healthcare technology companies discussed in this chapter that are either domiciled in
Asia or derive a large part of their revenue from Asia.
HOW TO INVEST?
There are relatively few pure-play healthcare technology companies in Asia, although
many diversified technology, telecom, Internet, pharmaceutical, medical device, and drug
distributor companies have been investing heavily in this space. In this section, we share a
map of companies both listed (see Exhibit 90) and unlisted (see Exhibit 91) to watch
against the 12 transformational healthcare technology niches we think are most exciting
in Asia.
EXHIBIT 90: Which companies are leading the charge in Asia? Listed companies to watch
Telemedicine, Online Rx Connected communities Digitized hospitals Genomics enabled care
Alibaba (BABA.US) Tencent (700.HK) Siemens (SIE.DE) - NC Berry Genomics (000710 .CH) - NC
JD.com (JD;US) Acer (2353.TW) Samsung (005930.KS)
SK Chemical (006120;KS) - NC M3 (2413.JP) - NC Neusoft (600718.CN) - NC
Apollo (APHS.IN) - NC Bumrungrad (BH.TB)
Jointown (600998.CH) - NC IHH Healthcare (IHH.MK)
Ping An (2183.HK, 601318.CH) Bangkok Dusit Med Services (BDMS.TB)
Shanghai Pharma (2607.HK, 601607.CH) - Winning Health (300253.CH) - NC
NC Ramkhamhaeng Hospital (RAM.TB) - NC
Universal Health International (2211.HK) -
NC
IOT health monitoring Smart pills Cloud health, big data Blockchain
Sysmex (6869.JP) - NC . Ali Health (0241.HK) - NC .
InBody (041830.KS) - NC Hitachi Medical Corp (6910.JP) - NC
Samsung (005930.KS)
Andon Health (002432.CH) - NC
Sinocare (300298.CH) - NC
Panasonic (6752.JP) - NC
Koninklijke Philips (PHG.US) - NC
CISCO (CSCO.US)
Sinocare (300298.SZ) - NC
Neusoft Corp (600718.CH) - NC
Augmented reality Gamification of health Automation & robotics AI & machine learning
Toshiba (6502.JP) Intel (INTC.US) Intuitive Surgical (ISRG.US),- NC Baidu (BIDU.US)
Vital (3151.JP) - NC MAKO Surgical (MAKO.US) - NC IBM Watson (IBM.US)
Curexo (060280.KS) - NC Bumrungrad (BH.TB)
Meere Company (049950.KS) - NC
Panasonic (6752.JP) - NC
Bumrungrad (BH.TB)
Hyundai Heavy (009540.KS) - NC
Accuray (ARAY.US) - NC
Harbin Boshi Automation (002698.SZ) - NC
SIASUN Robot & Automation (300024.SH) - NC
EXHIBIT 91: Which companies are leading the charge in Asia? Private companies to watch
Telemedicine, Online Rx Connected communities Digitized hospitals Genomics enabled care
Good Doctor DXY Wooridul Spine Hospital (Korea) iCarbonX
We Doctor Good Doctor Asan Medical Center (Korea) Wuxi NextCODE (owned by Wuxi AppTec)
Spring Rain We Doctor Seoul National University Hospital (Korea) MyGene 23
Medistar Spring Rain Shanghai Kingstar Novogene
Karma Practo Tianjin Tech Group AsiaGenomics
RingMD Clintal iDNA
Beilian Information Technology DocDoc) BGI
Huashi Pharmacy Global Health & Travel
Ren Bin Practo
111.com.cn
IOT health monitoring Smart pills Cloud health, big data Blockchain
Respiri TruTag Synyi Factom
Sleepio Proteus Digital Health Yidu Cloud Zhong An
Xiaomi MicroCHIPS Inc JP Technology Etherium
Huami BodyCap SomaLogic Shanghai Blockchain Devt Alliance
Singapore Natl Cardiac Monitoring HealthTell ZhongAn
iHealth Lab AOBiome
Xikang GALT
Imagu
Robustinique
Augmented reality Gamification of health Automation & robotics AI & machine learning
Vital Images (Toshiba Medical Systems) Playbasis Institute of Innovative Surgical Technology MedLinker
Serious Games Asia Beijing TINAVI Horizon Robotics
PerfectDoctor Jinshan Science & Technology Yidu Cloud
Tan Tock Seng Hospital Smarobot Qingdao Baheal Pharma
Agoyo Fosun JV with Intuitive Surgical
Tan Tock Seng Hospital
INVESTMENT IMPLICATIONS
Healthcare is only 3% of MSCI Asia Ex-Japan Index (compared to 15% of S&P 500), but
the long-term drivers for huge growth in healthcare across the region are compelling.
Healthcare stocks have remained resilient relative to other sectors in economic
uncertainty and wider market re-ratings across Asia.
The majority of the healthcare technology companies discussed in this chapter are either
outside our coverage or still privately held. We see a tidal wave of technology companies
in Asia that are changing and improving how healthcare is administered. Artificial
intelligence tools, enabling physicians to make better informed clinical pathway decisions
and patients to better find the medical information they need; self-service healthcare
kiosks; remote physician consultation and patient monitoring; augmented reality-assisted
surgeries, genome editing to remove risk-associated mutations; smart pills with
authentication information invisibly encoded in their surface; and truly personalized
genomics-, proteomics-, and metabolomics-driven medical and wellness advice will likely
all be integral parts of how healthcare is consumed in Asia in the not-too-distant future.
While the United States broadly leads globally in the development of novel healthcare
technologies, we think Asia is surging ahead in pockets and will likely leapfrog the
developed West in the widespread integration of some of these technologies into daily
clinical use. Why? First, the burning platform for change runs hotter in Asia deeper,
unsolved, and seemingly intractable problems in Asia's healthcare systems mean
traditional approaches won't go far enough; technology must be a part of the solutions.
Asia's health systems face the same challenges that plague developed markets (e.g.,
costs spiraling up, misaligned incentives across stakeholders, and unpaid debts), plus
some unique and arguably more difficult emerging market obstacles to healthcare
modernization (e.g., wider income and health inequality, low health spend per capita,
inadequate infrastructure, endemic corruption, and broken public trust). Second, single-
payer systems incentivize investment in long-term preventive medicine on a more
ambitious scale than elsewhere. Third, technology- and smartphone-enabled new ways to
do things take off faster in Asia across many sectors we would argue that healthcare
won't be any different. China has higher smartphone penetration and usage than many
other markets, less "big brother" fears about where health data goes and who owns it,
capital flooding into heath technology start-ups, and high public frustration with the
status quo health system.
How to invest? Focus on companies that are genuinely using technology to solve
problems in the health system and address unmet medical need. Look for the players that
achieve the widest adoption and partner with top hospitals around the region. Watch for
IPOs of the biggest companies listed in Exhibit 91, some of which are already
extraordinarily well-funded and racing ahead (e.g., iCarbonX has raised US$600 million,
Good Doctor's raised US$500 million in a Series A round, and We Doctor raised US$400
million in a Series A round).
Three of the eight healthcare stocks that we cover are discussed in this chapter ASEAN
hospital operators that are aggressively embracing new technologies and fully digitizing
themselves Bumrungrad, IHH, and Bangkok Dusit.
We rate Bangkok Dusit Medical Services (BDMS.TB, 12-month target price of THB25.1)
and IHH (IHH.MK, 12-month target price of MYR7.0) outperform. And we rate
Bumrungrad (BH.TB, 12-month target price of THB182) market-perform. The
pharmaceutical stocks in our coverage are not discussed in this chapter.
OVERVIEW
Much of this Blackbook delves into a range of machine learning applications from
artificial intelligence, to autonomous driving, to augmented reality. In this chapter,
however, we discuss something different the need for energy storage to enable
renewable energy and improve the efficiency of the world's biggest machine, the machine
that powers all machines the power grid.
Although the main growth market for batteries and energy storage is electric vehicles,
there is a huge potential market in energy storage systems (ESSs) (see Exhibit 92 and
Exhibit 93). These systems allow electricity to be stored for future use upon production.
They are also garnering significant attention, given the implication for renewable energy
sources, such as wind and solar power, which are clean, but largely volatile in terms of
output. ESSs can help improve the utilization, efficiency, and reliability of the electric grid
and facilitate the use of renewables by enabling time control over the volatile energy
supply from these sources.
EXHIBIT 92: Total global lithium ion battery (LiB) market projection: The market size is expected to increase from US$31
billion in 2016 to US$88 billion by 2025, 16% of which will come from the ESS market
80
70
60
Bn USD
50
40
30
20
10
0
2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025
Note: Passenger vehicle includes BEV, PHEV, HEV, and start-stop vehicles. The ESS market includes flow battery.
Source: SNE Research, CairnERA, and Bernstein estimates (2017 and beyond) and analysis.
EXHIBIT 93: Breakdown of global LiB demand in GWh: By 2025, the market demand is expected to be 285GWh (epic fail),
506GWh (slow adoption), and 851GWh (rapid adoption)
900 851
800
BatteryMarketSieinGWh
700
600 BEV
506 PHEV
500
HEV
400
SSV LIB
285
300 ESS
ebus
200
88 Small
100
0
2016 2025 Epic Fail 2025 Slow Adoption 2025 Rapid Adoption
Source: SNE Research, CairnERA, and Bernstein estimates (2025) and analysis.
Although there are other needs for ESS on the grid, the main driver for the increased
demand is the adoption of renewable energy sources. Quite simply, the sun doesn't shine
24 hours a day and wind is also intermittent and unpredictable. In small quantities, this
intermittency and unpredictability isn't a huge issue as it can be largely offset by
fluctuations elsewhere and the use of gas peaker plants.
We can divide the ESS market into four main segments based on application: 1) utility
applications; 2) renewable integration; 3) UPS; and 4) residential/community storage (see
Exhibit 94). The ESS market can also be divided into three segments based on
beneficiaries: 1) independent system organization (ISO) and regional transmission
organization (TRO); 2) the utility group; and 3) customer services (see Exhibit 95). Energy
storage systems have the ability to revolutionize current methods of energy generation
and distribution by providing storage at all points throughout the grid.
The electric grid operates largely without any stored resources. Energy storage is
important for utilities, given that it can help spread out energy supply by storing up excess
energy for later use. This can help increase the efficiency of the grid, while lowering
production costs by increasing utilization rates and, in turn, matching supply with demand
better, not to mention the enablement of renewable energy sources.
One of the most basic ways by which energy storage can improve the efficiency of overall
utilities is through peak loading and frequency adjustment. Electricity demand often
varies sharply based on the time and season, with consumption peaking at work hours
and also during summer when air-conditioning needs are high. Additional energy can be
provided at such need by building more capacity, or buying electricity from other sources,
but neither option is cheap. Energy storage could, in such circumstances, allow for huge
cost savings by enabling utilities to avoid building excess capacity (T&D deferral), and
instead use stored energy generated or purchased from other utilities at non-peak hours
(see Exhibit 96 to Exhibit 99).
EXHIBIT 94: Grid energy storage solutions come in four major areas: renewable integration, frequency regulation, T&D
Deferral, peak shifting, UPS, and small-scale storage for PV systems
Source: LG Chem.
EXHIBIT 95: Energy storage systems can be broadly divided into three segments based on beneficiaries
ISO/RTO
services
Balances out fluctuations between electricity generation and load by
Frequency regulation
regulating grid frequency. Frequency drops when demand > supply.
Lowers energy cost by helping store energy produced when cheap and
Energy arbitrage
discharging in peak hours when prices are high.
Synchronizes reserve capacity to the grid that can be delivered in 10
Spinning reserve minutes to meet demand upon grid operator command. Fills in until
longer-duration assets can come online.
Voltage support Controls voltage of grid by injecting or absorbing reactive power.
Acts as a source of power/energy upon power outage. Can either
3 Stakeholders/ Black start provide reference frequency or energy to help restart the system.
Numerous Utility
Services services
Distribution & transmission Allows delay for building transmission and distribution capacity by
deferral providing energy storage that can accommodate peak demand.
Transmission congestion relief
Resource adequacy
Customer
services
Backup power Discharge during a black-out
Store solar energy generated from photovoltaic (PV) panels, allowing an
Self-consumption
energy self-consumption
Electricity bill saving EVs charge during off-peak times when electricity rate is low
Note: ISO and RTO stand for independent system operator and regional transmission organization, respectively.
EXHIBIT 96: Peak shifting: ESS is charged during non-peak EXHIBIT 97: Frequency regulation: Fluctuations in frequency
hours and utilized during peak hours when electricity is can be controlled by ESS, which acts as an online power
expensive source
EXHIBIT 98: T&D deferral: ESS can help delay building EXHIBIT 99: Renewable integration: ESS can help smooth
transmissions and distribution capacity out the intermittent supply of energy from wind and solar
energy
Energy storage technologies available today are extremely diverse, including pumped
hydro, compressed air, and various types of batteries and capacitors (see Exhibit 100). For
large-scale utility energy storage, where space and ecological concerns are not a
concern, the economics of pumped hydro are difficult to compete with. However, for
electric vehicles and residential roll-outs of solar, we find batteries especially lithium
ion batteries, with, by far, the highest efficiency and energy density the best solution
across most applications (see Exhibit 101 to Exhibit 103).
CAES
Molten salt thermal
Flow battery: VR battery, ZBR, etc.
NaS
LiBs
Lead-acid battery
Flywheels
Short
duration High-power supercapacitor
(Seconds)
Small Large
kW 100kW 1 MW 100MW GW
Power rating
Compressed Geographic
8hrs >20yrs N/A Commercial High pressure 70%
Air restrictions
Other flow
1-5hrs unknown unknown 12-17 R&D/Demo Stable 70-75%
battery
High operating
NaS 6hrs 5-15yrd yes 49 Commercial 75%
temperature
EXHIBIT 102: LiB also leads other energy storage EXHIBIT 103: Non-battery solutions are, by far, the current
technologies in terms of energy efficiency (i.e., energy leader in terms of cost, but LiB costs are also declining fast
stored/energy used)
85% 800
US$/kWh
80% 600
75%
400
70%
200
65%
60% 0
Flow
- Above Ground
LiB
Pumped
- Under Ground
NaS
Flywheel
Lead-
acid
Flow
Compressed
LiB
Pumped
NaS
Lead
Flywheel
-acid
Hydro
Hydro
CAES
CAES
Air
Non-Battery Battery
Non-Battery Battery
Source: Sandia National Laboratory and Bernstein analysis. Source: Sandia National Laboratory and Bernstein analysis.
NON-BATTERY SOLUTIONS: Pumped Hydro uses locational energy to store power during the off-peak hours and
MASSIVE INFRASTRUCTURAL generate the stored power during the peak hours. As illustrated in Exhibit 104, the
INVESTMENTS OF PUMPED
storage mechanism consists of two reservoirs of different elevations, between which lies
STORAGE AND COMPRESSED
AIR the pumping and generating mechanism. The energy storage capacity is defined by the
size of the two reservoir bodies; the designed life of these facilities often stretch over half
a century, with theoretical cycle of more than 1x per day throughout its life (weather and
other conditions permitting).
Compressed Air Energy Storage (CAES) relies on a similar concept of storing energy using
excess electricity during the off-peak hours and using that as a supplement during the
peak hours. As illustrated in Exhibit 105, the storing of energy takes the form of motor-
powered compression and storage in a reservoir, whether in naturally available cavern
(naturally created or abandoned natural gas wells in some cases) or man-made
containment vessels. During the peak hours, the compressed air is run through power
turbine and generators, which in turn supplies power to the grid. CAES facilities utilizing
natural underground capacity have been built with capacity of +1,000MWh; ones with
above-ground containment vessels have been set at ~300MWh or less.
Flywheel stores energy in the form of kinetic energy by accelerating a rotor to a very high
speed. The system could then transfer kinetic energy into AC power through the use of
control and power conversion systems. High-power flywheel systems are able to deliver
energy in seconds. Similar to other non-battery solutions, flywheels have long lifetimes of
~20 years.
EXHIBIT 104: Pumped hydro: Massive social infrastructural EXHIBIT 105: CAES: Lots of moving parts for compressing air
investment, with two reservoirs and pumps and generating electricity
Lithium-ion batteries (LiBs) are still the leading contender for most energy storage needs
due to the rapidly falling costs driven by the ramp in scale for EVs. However, for projects
where more than a few hours of power are required, other battery types such as flow
batteries are better due to their ability to scale energy independently of power. Similarly,
when guaranteeing that the system will last beyond 10 years, LiBs may not be the best
solution and flow batteries may be preferred as they last much longer.
Main technologies in the field, Zinc Bromide Redox (ZBR) and Vanadium Redox (VRB), are
named after the chemicals used for electrolytes in the process. The key advantage of
these technologies is the ability to scale capacity independently of power (and, thus,
cheaply). The disadvantage is mainly operational requirements (i.e., pumps for liquid
electrolytes). On the cost front, flow batteries are currently more expensive than LiBs, but
in many cases, due to longer lifetime, can be the most economical solution, especially for
long-duration energy (low power) requirements. The latest VRB-battery manufacturers'
claim cost of US$350/kWh, but it mostly depends on the duration and power
requirement.
EXHIBIT 106: Simplified structure of a flow battery EXHIBIT 107: Simplified structure of an NaS battery
Though the ESS market is somewhat smaller than the EV battery market, we believe the
outlook is positive with the potential to reach US$16 billion by 2025 (FY2015-25 CAGR
of 20%), thanks to the increased need for energy storage from a variety of industries as
well as homes and governments that seek to improve the cost, reliability, and quality of
energy that they are consuming (see Exhibit 108). Note that this market-size analysis
represents only the DC bus size of the system and doesn't include AC power equipment,
installation, and shipping costs.
EXHIBIT 108: We project the ESS (battery) market will grow to US$16 billion annual installations by 2025
18
16
16
14
Total new installation in bn USD
12
12
10 9
8 7
6 5
4 4
3 4
4 3 3
2
0
2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025
Ancillary Peak shifting Reserve power Load shifting Others
EXHIBIT 109: ESS ASP in US$/kW by service EXHIBIT 110: ESS ASP in US$/kWh by service
2,500 1,000
900
2,000 800
700
ASP $/kW
1,500
ASP $/kWh
600
500
1,000
400
300
500
200
- 100
Source: CairnERA estimates (2017 and beyond) and Bernstein analysis. Source: CairnERA estimates (2017 and beyond) and Bernstein analysis.
According to the U.S. Department of Energy (DOE), currently there are around 180GW
projects that are either under construction or in operation. A majority of this amount is
attributed to pumped hydro projects, which has been available in the market for decades
(see Exhibit 111). For geographic mix, China, Japan, and the United States are the three
biggest markets for ESS, with majority power capacity coming from pumped hydro (see
Exhibit 112).
EXHIBIT 111: Global ESS projects by technology: A total of 180GW projects have been completed or are currently under
construction
CAS
0.4% Lead
0.0%
Others
2.6%
LiBs
1.1%
Battery
PumpedHydro 1.3% Flow
95.7% 0.1%
Sodium
0.1%
Others
EXHIBIT 112: Global ESS projects by country: Top 7 countries account for 66% of the total 180GW projects
By analyzing the announced ESS projects since 2013, we found that there is a strong tilt
toward LiBs, though non-battery solutions still dominate the market. Top LiB ESS
technology suppliers include AES (the United States), Korean players (LG Chem and SDI),
and BYD (China) (see Exhibit 113). Currently, SDI is working on large utility projects in
North America and Europe, which are key drivers for top-line growth in 2017. The
company has recently launched a line-up of high capacity (E2)/high power (P3) ESS
modules aiming for the Europe market. The E2 doubles the energy capacity per container
to 9.1MWh.
EXHIBIT 113: Announced ESS power capacity since 2013 has shown a strong tilt toward LiB; the projects are led by
companies, including AES, LG Chem, and SDI
MW in capacity 2013 2014 2015 2016 #projects MW
Lead acid 3 11 6 0 AES 7 253
Flow 18 8 45 29 LG Chem 24 202
Li-ion 156 399 546 527 Samsung/SDI 21 154
Sodium based 63 0 1 0 BYD 10 94
Others 5,876 5,376 2,069 4,150 Kokam 9 81
Total 6,116 5,793 2,666 4,706
Note: Announced capacity. The heat map runs from Red-Yellow-Orange-Green from low to high. See the online version for colors.
EXHIBIT 114: Global forecast of total new installations EXHIBIT 115: Global forecast of total new installations (GWh)
market size (US$ billion) by technology by technology
18 45
16 40
14
Total new installations bn USD
35
12 30
10 25
8 20
6 15
4 10
2 5
0 0
LiBs Lead-acid Flow Pumped Hydro CAES LiBs Lead-acid Flow Pumped Hydro CAES
Note: The analysis includes all energy storage by DC system. Note: The analysis includes all energy storage by DC system.
Source: CairERA estimates (2017 and beyond) and Bernstein analysis. Source: CairERA estimates (2017 and beyond) and Bernstein analysis.
EXHIBIT 116: We expect LiBs to emerge as the dominant chemistry in the ESS segment by 2020
Flow Batteries
Lead Acid
LiBs
0 5 10 15 20
In GWh
2016 2020
MAJOR ESS PROJECTS While the recently announced non-battery storage projects still lead markets in terms of
the powered scale (can be as high as 330MW for CAES and over 1GW for pumped hydro),
there are increasingly more scalable battery ESS projects coming. In the LiBs sector, AES
Southland has been rewarded a 20-year power purchase agreement in California in
2014. The project will provide 100MW/400MWh of energy in the south of Los Angeles.
AES has also recently announced the world's largest LiB project (for renewables), which
will provide minimum 100MWh LiBs paired with Lyon Solar's 100MW facility in South
Australia. More recently, the company has launched an ESS with 30MW/120MWh
capacity, which utilizes LiBs from Samsung SDI. The Chinese EV giant BYD is also one of
the global-leading LiBs ESS providers. In 2015, its battery has been used in a 31.5MW
ESS (18x 1.8MW modules) that is adjacent to a 100MW wind farm in West Virginia, as well
as being involved in projects in China.
EXHIBIT 117: Top projects (in terms of rated power) announced between 2013 and 2016
Rated Energy
Duration
Project/owner ESS Supplier Country Status power capacity Service
(hr)
in MW in MWh
Pumped hydro
Revelstoke Hydro Battery n/a Canada Announced 4,000 n/a n/a Time Shift
Eagle Mountain Pumped Storage Project n/a US Contracted 1,300 n/a n/a Time Shift
Black Canyon Pumped Storage n/a US Announced 1,000 11.0 11,000 Electric Supply Capacity
CAES
Gaelectric Dresser-Rand UK Announced 330 6.0 1,980 Time Shift
Apex Compressed Air Energy Storage Dresser-Rand US Announced 317 96.0 30,432 Black Start
Flywheel
Amber Kinetics is 100% owner of Energy Nuevo Storage Farm, LLC Amber Kinetics US Contracted 20 4.0 80 Electric Energy Time Shift
Confidential VYCON US Contracted 8 n/a n/a Resiliency
Hydro One Networks Inc. Temporal Power Ltd. Canada Operational 5 0.10 1 Distribution upgrade
LiBs
AES Southland AES Energy Storage US Contracted 100 4.0 400 Electric Supply Capacity
Lyon Group AES Energy Storage Australia Under Construction 100 1.0 100 Renewables
Private Owners unknown Germany Operational 68 3.0 204 Electric Bill Management
Woojin Industrial
KEPCO South Korea Operational 48 0.25 12 Frequency Regulation
Systems / LG Chem
National Grid Corp. of the Philippines (NGCP) unknown Philippines Operational 40 n/a n/a Frequency Regulation
Tohoku Electric Power Company Toshiba Corporation Japan Operational 40 1.0 40 Renewables
KEPCO Kokam South Korea Contracted 36 0.37 13 Frequency Regulation
Southern California Edison unknown US Contracted 35 4.0 140 Electric Supply Capacity
Beech Ridge Energy Storage, LLC. BYD US Operational 32 n/a n/a Frequency Regulation
San Diego Gas & Electric (SDG&E) Samsung SDI US Contracted 30 4.0 120 Time Shift
Imperial Irrigation District Samsung SDI US Operational 30 0.67 20 Black Start
Saxony-Anhalt, Germany SK Innovation Germany Announced 30 n/a n/a Renewables
NaS
Kyushu Electric Power Co. NGK Insulators Japan Operational 50 6.0 300 Frequency Regulation
Tokyo Electric Power Company (TEPCO) NGK Insulators Japan Operational 0 7.20 3 Frequency Regulation
There are two main flow battery technologies Zinc Bromide Redox (ZBR) and Vanadium
Redox (VRB). We think VRB hold the most promise among flow batteries and we see a
very clear role for these batteries in the future. The key advantage of this technology is the
ability to scale capacity independently of power (and, thus, cheaply).
For VRB, self-discharge is not a problem because the electrolytes are stored in separate
tanks and its life is the longest among all battery ESSs. One of the largest VRB systems
that is currently under testing is the 15MW/60MWh project built by Japan's Hokkaido
Electric Power (HEPCO, ticker YTO:5802, not rated) and SEI. The two parties want to test
the performance of the system as a new way to adjust wind and solar intermittency.
For ZBR battery, the extremely corrosive nature of the bromine electrolyte will cause
degradation and potential failure, though the active materials themselves do not degrade.
Thus, the lifetime of ZBR battery is not much dependent on the number of cycles or the
depth of discharge, but on the duration that the system has been under operation. Despite
the technology still being at an early stage, there are dozens of projects operational in
Australia, France, and the United States. Currently, the biggest operated project is in
France, conducted by ZBB and Pacific Beachcomber. The project is using ZBB's 2MWh
systems (40x 10kW/5kWh modules) for a luxury eco-resort. Apart from these completed
projects, there are an increasing number of large-scale projects that have been
announced.
Another relatively less commonly seen technology in this field is hydrogen bromine (HBR)
battery. Currently, there's only one operational project located in Israel, with rated power
of 50kW in duration of two hours.
EXHIBIT 118: New capacity build up for global flow battery EXHIBIT 119: Price forecast for flow battery VRB leading
the cost declines
1,800
6
1,600
4.9
5 1,400
1,200
Price/kWh
4
1,000
3.0
GWh
3 800
2.1 600
2
400
1.4
1.0 200
1 0.7
0.5
0
0 0.1 0.1 0.2
0
VRB ZBR
VRB ZBR HBR HBR Average
Source: CairnERA estimates (2017 and beyond) and Bernstein analysis. Source: CairnERA (2017 and beyond) and Bernstein analysis
FLOW BATTERY COMPANIES In the flow battery space, there are several interesting companies that are promoting the
technology in the global ESS market. For geographic mix, the United States, Kazakhstan,
and Japan are top countries that are applying flow battery (see Exhibit 120). In terms of
the power scale, Dalian Rongke Power and UniEnergy Technologies (UTE) together rank
top. Rongke was jointly founded by Dalian Institute of Chemical Physics and Dalian
Bolong Holding in 2008. The company is focused on manufacturing VRB and has
developed its own IP in key materials, such as electrolyte and membrane. The company
has received orders of key materials of more than 100MW in Japan and Europe. UTE,
which was founded in 2012 in the United States, is a JV between Bolong Holding Group
and Pacific Northwest National Laboratory. Since then, the two companies have worked
closely on VRBs. In 2016, they planned to deploy the world's largest battery rated at
800MWh (made up of ten 20MW/80MWh VRB systems) to provide peak-shaving and
enhance grid stabilization in Dalian, China.
By contrast, Primus Power (the second-largest player) uses ZBR chemistry rather than
vanadium, which Rongke uses. The company was founded in California in 2009 and has
raised US$94 million in equity and US$20 million in government grants so far. The
company's products use a single-loop flow battery design, which plates zinc on titanium-
based electrodes to perform the key energy exchange function, rather than running
electrolyte through membranes, as most other flow batteries do. Primus claims to have
better total cost of ownership (TCO) due to less pumps, fewer other parts, higher
reliability, and high performance. In 2015, the company partnered with the principal
electricity provider in Kazakhstan to build a total of 25MW/100MWh ESS-EnergyPod in
the country. The goal is to help Kazakhstan reach its renewable energy goals of 30% by
2030 and 50% in 2050.
Sumitomo Electric Industry is another big VRB battery company. It has recently started the
demo of a VRB system (2MW/8MWh) in California. The purpose is to evaluate the
management of surplus electricity and regulate grid frequency and voltage fluctuation.
Imergy, as the fourth-largest flow battery companies in terms of power scale of projects
run in the market, unfortunately ceased operation in 2016 as investors chose not to fund
further needs. Imergy was also focused on VRB and promised to bring the costs of flow
batteries from the industry benchmark of US$500/kWh to under US$300/kWh.
Last but not the least, Puneng, as the second-largest flow battery player in China and fifth
globally, owns one of the world's largest renewable energy demonstration project
(2MW/8MWh VRB installation at the North China State Grid wind and solar demo site in
Zhangbei). Puneng has developed (and patented) its own membranes and electrolytes. It
has internally developed a low-cost composite proton-conducting polymer membrane,
comprising a blend of soluble and sulfonated polymers, saving significant cost compared
to purchasing membranes externally. Puneng claims that its own membranes cost
US$15-US$20 per square meter compared to the market alternative at US$400-
US$500 per square meter (see Exhibit 121 and Exhibit 122).
EXHIBIT 121: Global flow battery projects by company: Chinese-based Rongke and U.S.-based Primus are the top two flow
battery makers
ZBB
2%
Gildemeister Other
3% 12%
Puneng
3% UET/Rongke
29%
Imergy
6%
Sumitomo
21%
Primus
24%
EXHIBIT 122: Global top 10 flow battery ESS projects in terms of rated power
Rated power Duration Capacity
Owner ESS Supplier Technology Country Status Service
in MW (hr) in MWh
Samruk Energy Primus Power ZBR Kazakhstan Contracted 25 4.0 100 Electric Energy Time Shift
Under
China National Energy Administration UET / Rongke VR US 20 4.0 80 Black Start
Construction
Hokkaido Electric Power Sumitomo VR Japan Operational 15 4.0 60 Renewables Capacity Firming
GuoDian LongYuan Wind Power Rongke VR China Operational 5 2.0 10 Electric Supply Reserve Capacity
Sumitomo Densetsu Sumitomo VR Japan Operational 3 0.27 0.8 Electric Bill Management
Antigua and
Government of Antigua and Barbuda sun2live VR Operational 3 4.0 12 Renewables Capacity Firming
Barbuda
State Grid Corporation of China Puneng VR China Operational 2 4.0 8 Frequency Regulation
Snohomish County Public Utility District UET VR US Contracted 2 4.0 8 Electric Energy Time Shift
INVESTMENT IMPLICATIONS
We are positive on the long-term growth of electric vehicles and lithium-ion battery
suppliers. We believe that batteries are on the cusp of dramatic growth as costs fall to
levels that make EVs not just cool but also economical, and energy storage for both
utilities and renewables starts to make sense.
Samsung SDI: Although China's clear protectionist agenda favoring local battery makers
has hurt Samsung SDI (outperform, target price ofKRW 160k), we still see the company
as well positioned to benefit from long-term EV and ESS growth globally. As the purest
play in advanced batteries and one of the world's leaders, Samsung SDI is well positioned
to successfully ride the long-term EV and ESS wave. Samsung SDI, as 15% owner of
Samsung Display and OLED material supplier, also benefits significantly from the
upcoming OLED growth as iPhone adopts OLED starting 2H 2017.
BYD Company Ltd: As the largest EV maker and now the largest battery maker in the
world with the lowest battery costs in the industry, BYD is well positioned to benefit from
the long-term growth of EVs and ESSs particularly with its home court advantage in
China. We recently initiated BYD Company Ltd H-shares (1211.HK) with an outperform
rating and target price of HK$60, and A-shares (002594.CH) at market-perform and
target price of RMB55 based on our SOTP, which correlates to 20x our 2018 EPS (below
LT average of 26.5x and 45.7x for H and A-shares, respectively).
The winners from the communications and mobility revolutions over the last 20 years
have been the Internet sector, Apple, and at the margin Samsung and TSMC. The
secrets to value accretion over the period have been low capital intensity, low or zero
customer acquisition cost, and a "winner take all" market structure. The question of
whether these companies were truly creating value, or simply capturing it, has been
secondary.
Based on the lessons of the last two decades, in this chapter we consider whether looking
at companies that are doing revolutionary things with machine learning is the right
approach. Specifically, we consider whether the Internet companies could again
extract all value from this generation of technology. More fundamentally, we argue that
the stakes are too great with this round of technologies for regulators to simply allow a
small group of companies globally to once again capture all of the value.
An alternate future
Over the last two decades, the three most-compelling and far-reaching changes in
technology globally have been: the increased speed of telecommunications, the falling
cost of memory semiconductor, and the improved computing power of processors.
These technologies have put computers (in the form of smartphones) into roughly 3 billion
pockets globally, and in doing so disrupted and destroyed industries as diverse as
landline telephone service, taxis, newspapers, supermarkets, big box retailers, food
delivery, music, television, hospitality, and travel. This revolution has also changed how
people to take a few examples date, watch sport, shop, develop, and express political
allegiance, maintain friendships, and for an increasing proportion of the world's
population (call it Generation Selfie) experience their own lives.
The connected population globally can access a heretofore unimagined amount of not
just information but also tools to make them more efficient at work, more connected to
friends and family, and (one might hope) more evolved, more content, more informed, and
even happier. And, as ever, we are still early in the innings.
The technology and the infrastructure through which this revolution took place were
supplied by a remarkably small number of companies. The telecommunications
equipment companies (including Cisco, Alcatel, Lucent, Nortel, ZTE, Qualcomm, and
Huawei) increased the speed of telecommunications multiplexing and wireless
technology. The large telecommunications companies (AT&T, Verizon, Sprint, British
Telecom, Vodafone, Deutsche Telecom, Orange, China Mobile, China Telecom, China
Unicom, and others globally) accelerated the roll out of fiber optic infrastructure and
third- and fourth-generation wireless technology. Intel and TSMC (among others)
increased the speed and power of computer processing, while lowering the cost.
Samsung, Hynix, and Micron (among others) reduced the cost of memory storage.
Globally, less than three dozen companies did the lion's share of the work.
And, in combination, these three technologies enabled the mobility revolution of the last
two decades.
And yet, when the chapter is closed on the key technologies of the first two decades of
the 21st century, the headlines will be (roughly in order): iPhone, Facebook, iPod, WeChat,
Amazon, BABA, iPad, Instagram, Tinder, and perhaps Uber plus Samsung's Galaxy Note
7. The technology companies invented the future. The telecommunications industry built
it. And the Internet sector (the customer-facing platform) plus Apple captured most
of the incremental value (see Exhibit 123).
EXHIBIT 123: Market capitalization across technology and technology-adjacent sectors (2000-16)
2,000 Internet
1,800
1,600
1,400
1,200
USD B
Telcos
1,000
800
Apple
600
Processors
400 / foundries
Memory
200
Telecom Equipment
0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
THE COUNTER-FACTUAL The irony of this allocation of value between the players is that the mobility and
WHERE THE INTERNET STILL communications revolutions would still have occurred without the existence of any of the
WINS
big names in the Internet sector: Google, Facebook, Tencent, Alibaba, Baidu, Amazon, or
even Apple.
The tablet existed before the iPad. Palm rolled out touchscreen technology on its
networked Personal Digital Assistants almost a decade before the iPhone. RIMM sold
BlackBerries with color screens in 2004. Before Google, there were Ask Jeeves and
Yahoo! Before Facebook, there was MySpace. Alibaba's TaoBao and Tmall are an
adaptation of the marketplace model pioneered by eBay. Uber and Lyft figured out how to
disrupt the limousine business at roughly the same time.
In short, the winners and losers in the Internet sector are completely interchangeable: it is
a modern version of Coke versus Pepsi, but with network effects and, therefore, no stable
equilibrium other than at the very extremes. Pepsi really does taste better if all of your
friends drink it too.
Once the global fiber optic network was built, and with memory cost falling and
processors becoming more powerful, someone was going to come up with Instagram.
Success in the Internet space comes down to execution. Said differently, there are no new
ideas. In a parallel universe where none of the winners in the Internet sector exist, there
would still be an Internet and there would still be a mobility revolution.
Yet if, in still another parallel universe, the global telecommunications industry decides
as a group that it has no interest in funding wireless technologies in the 1990s or data
services in the 2000s, then stationery shops, video stores, the daily newspaper, and
network television are all still viable businesses. In short, Tinder does not exist in a dial-up
world. But in a 4G world without Barry Diller, people would still use new technology to
socialize (as humans have done since the invention of the camp fire).
Samsung and TSMC have tripled in value over the last decade, so it is not necessary to
feel too bad for all of the "losers" in this arrangement. The telecom equipment
manufacturers have had a rougher time. For a brief moment at the end of 1998, Cisco
Systems was the most valuable, publicly traded company in the world and Nokia the most
valuable company in Europe. Today, they are not (see Exhibit 124).
But if the communications and mobility revolutions of the last two decades were created
by a small number of companies, the lion's share of the value went to an even smaller
group. And the overlap between the two groups is at best haphazard.
EXHIBIT 124: Indexed price performance of technology stocks (2000/IPO-2017 year-to-date) (December 1999 = 100)
25,000 Tencent
20,000
15,000
Netflix
5,000
Apple
NVIDIA
Baidu
Google
Amazon
0
Dec-99
Jun-00
Dec-00
Jun-01
Dec-01
Jun-02
Dec-02
Jun-03
Dec-03
Jun-04
Dec-04
Jun-05
Dec-05
Jun-06
Dec-06
Jun-07
Dec-07
Jun-08
Dec-08
Jun-09
Dec-09
Jun-10
Dec-10
Jun-11
Dec-11
Jun-12
Dec-12
Jun-13
Dec-13
Jun-14
Dec-14
Jun-15
Dec-15
Jun-16
Dec-16
Note: For stocks that were not listed in December 1999, the base is the IPO month.
Meanwhile, the Internet "winners" have gone from a market capitalization virtually
indistinguishable, from zero in 2000 to close to US$2 trillion (see Exhibit 123 and Exhibit
126), even without including Apple. The important question for our purposes is: could it
happen again?
50.0%
40.0%
Internet
30.0% Processors
/ foundries
20.0% Memory
Telecom
Equipment
10.0%
Telcos
0.0%
-10.0%
2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
We address three questions in this chapter. First, why did the most critical aspects and
technology providers of the communications revolution not capture most of the value?
Second, does the future of technology (augmented reality, virtual reality, artificial
intelligence, voice recognition, energy storage, and machine learning) face the same
challenge? And, if so, from whom is that challenge emanating?
THE ALLOCATION OF VALUE The communications revolution came down to three things: faster communications,
FROM THE COMMUNICATIONS cheaper memory, and better processing power. In our view, once these three
REVOLUTION OF THE LAST 20
technologies and infrastructure capabilities were in place, the rise of the Internet was
YEARS
inevitable (see Exhibit 127 and Exhibit 129).
Faster speed in telecommunications meant two or three "data tsunami" (to borrow a
phrase from our telecom analyst Chris Lane) in the last 20 years. The first came in the late
1990s and early 2000s as a fiber optic cable backbone was built out globally through
both undersea and terrestrial networks improving the speed and lowering the cost of
data transmission across the globe. Any volume of content could suddenly move at
(something approaching) the speed of light, and for fractions of pennies.
EXHIBIT 126: Market capitalization CAGR by group (2000-16) EXHIBIT 127: Revenue CAGR by group (2000-16)
40% 40%
35%
35% 32% 35% 32%
30% 30%
25% 25% 23%
20% 20%
15% 12% 15%
10% 10% 7% 6% 6%
5% 3% 3% 5%
0% 0%
-5% -5% -2%
-10% -8% -10%
Source: Bloomberg L.P. and Bernstein analysis. Source: Bloomberg L.P. and Bernstein analysis.
The second change in speed came with the resolution of the "last mile" problem between
(roughly) 2000 and 2010 as DSL (transmitting data digitally over a 100-year twisted
copper pair infrastructure) and construction of metropolitan fiber-to-the-home and fiber-
to-the-neighborhood meant that data did not have to slow down as it approached the end
consumer.
Third, the roll-out of 3G and 4G wireless technologies has effectively meant that whatever
can be done through the wired infrastructure can now be done wirelessly, almost as fast.
EXHIBIT 128: Revenue of key companies in the memory EXHIBIT 129: Revenue trajectory across the technology
sector (2000-16) sector (2000-16)
70 Others 800
Telcos
60 700
Micron 600
50
500
40
USD B
USD B
Hynix
400
Internet
30
300
Apple
20
200 Processors
Samsung / foundries
10 100 Telco Eqpmt
Memory
0 0
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
Note: Only memory segment revenues are included for Samsung. Others Source: Bloomberg L.P. and Bernstein analysis.
include Nanya and Powerchip.
Cheaper memory arose as commodity DRAM and NAND Flash memory prices per bit
dropped dramatically (see Exhibit 128). Since the 1970s, commodity DRAM prices have
dropped from approximately US$80,000 per Mb to around well under 1 cent per Mb (see
Exhibit 130). Similarly, this century, the price per bit for standard MLC NAND has dropped
from over US$1,000/GB since introduction to below US$1.00/GB (see Exhibit 131).
EXHIBIT 130: DRAM pricing has declined exponentially EXHIBIT 131: Flash Memory pricing has also declined
exponentially
$100,000.00 $81,920 $10 ,00 0.0 0
$10,000.00
$100.00
USD/Mb
$10 0.0 0
USD/GB
$10.00
$1.00
$10 .00
$0.10
$0.01 $1.00
12,000,000 1
$0.01
$0.00
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014
2016
$0.10
200 0
200 1
200 2
200 3
200 4
200 5
200 6
200 7
200 8
200 9
201 0
201 1
201 2
201 3
201 4
201 5
201 6
Source: DramXchange and Bernstein's Global Memory team. Source: DramXchange and Bernsteins Global Memory team.
The price per Mb of HDD storage has also fallen exponentially since the advent of the first
consumer HDD in the early 1980s. In the intervening period, this price has dropped from
about US$315/Mb to less than 1/20,000th of a cent per Mb, a six order of magnitude
difference (see Exhibit 132).
The average retail price (excluding taxes) of an external HDD for PC applications has
declined from about US$3,200 in 1981 to under US$100 in 2016, which is one order of
magnitude. Of course, you felt blessed with a 20 Mb HDD in 1983, while contemporary
HDDs comfortably reach the 1+ TB (terabyte, or 1 million Mb) capacities.
1000
100
10
1
USD/Mb
0.1
0.01
0.001
0.0001
1E-05
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
Source: Ivan Smith (http://ns1758.ca/winch/winchest.html) and Bernstein analysis.
A similar phenomenon is also apparent in CPUs. Although an x86 class CPU had roughly
1.2 million transistors in 1989 (for the then newly introduced i486), compared to a
contemporary Intel Xeon processor with over 4 billion transistors (and an equivalent
multiple in the processor's performance), their nominal cost on product introduction is
roughly about the same, about US$900-US$1,000, implying an exponential decline in the
cost per transistor (see Exhibit 133).
EXHIBIT 133: And, of course, the cost of processing power has also dropped exponentially
0.1
USD per Transistor
0.01
0.001
0.0001
0.00001
0.000001
0.0000001
1971
1972
1974
1978
1979
1982
1985
1989
1993
1994
1995
1997
1998
1999
2000
2000
2001
2002
2003
2004
2004
2006
2006
2006
2006
2008
2009
2010
2011
2012
2013
2014
2015
2016
THE TRIUMPH OF THE The first lesson from this chapter is the obvious one: life is not always fair, at least over the
PLATFORM IN THE short term. The entities that created the technologies and, therefore, the value when
COMMUNICATIONS REVOLUTION
it came to the communications and mobility revolutions were not the entities that
captured most of the value.
But the more interesting question is: first, why was this the case, and second, is this a
permanent condition that we could see replicated in the machine learning/artificial
intelligence version of the future of technology?
Our underlying thesis here is that capitalism is changing. Capitalism did not used to work
like this. If you built the railway line, you can charge whatever you wanted (at least to the
extent that the charge was cheaper than using the stagecoach) for transport on that
railway system. The network won. More broadly, unless heavily regulated, the companies
that created the value used to also be the companies that captured the value. The real
threat, as we highlight later in this chapter, was regulation limiting returns.
In 1982, the largest company on the NYSE and, therefore, the largest publicly traded
company in the world was IBM. IBM made computers. The computer was Time magazine's
"Machine of the Year" for 1982. There was elegance, or at least predictability to the way
that the world worked. IBM, the company that made the best computers in the world, was
also the most valuable company in the world.
90% 18
80% 16
Five-Year Survivorship
70% 14
Prior Appearances
60% 12
50% 10
40% 8
30% 6
20% 4
10% 2
0% 0
1975 1980 1985 1990 1995 2000 2005 2010 2015
And that predictability was a near permanent condition. In the 1970s, the largest
company in the United States was IBM. In 1990, the largest company in the United States
was IBM. The happy cabal of GE, IBM, AT&T, and Exxon were the dominant publicly
trading corporations for decades (see Exhibit 135). We rank order the top 10 companies
in (what would have been) the S&P 500 beginning in 1951 (the S&P 500 was first
compiled in 1957). From 1951 to 2010, measured in five-year increments, Exxon was in
the top 10 every decade. AT&T was in the top 10 in 11 of the 14 five-year measurement
periods. IBM featured nine times. In 2000, the survivorship fell to 40% (see Exhibit 134).
The number of prior appearances of the top five companies in the list (as compared to the
top five companies 25 years ago) in 1990 was 16. The number of prior appearances in the
list by the top five in 2015 was 10; and even then, we are using five-year increments.
Our point is: winners may still be taking all, but they "take all" for a much shorter time in
2016 than in 1966 or even 1996. That is, in our view, the appropriate word of caution
when assuming that current technology victors (Apple, Alphabet, and Amazon) are
permanent.
In terms of sweeping theories for what has changed in commerce over the last two
decades, network effects are in our view the strongest. AT&T in 1951 and in 1960
had a network effectliterally. If you wanted to make a phone call in the United States
before 1984, you had to use AT&T. After 1984, if you wanted to make a long-distance
phone call in the United States, you almost certainly had to use AT&T. The original network
effect was underwhelming because it was so obvious, and it required a physical network
that was subject to regulation. Vanderbilt owned the railway system. AT&T owned the
telecommunications system. Of course, they extracted the value of the network.
EXHIBIT 136: Top 20 stocks globally by absolute market value increase, 2000-17 year-to-date
700
721
Increase in Market Cap (USD B)
600
572
500
400
405
374
300 330
273 269
200
221
200
175 172
100 147 137
121 115 105 104
93 85 84
0
- - - - - - - - - - - -
Alibaba
ICBC
Insurance
Tencent
CCB
China Life
Facebook
Novartis
Apple
Amazon
TSMC
Berkshire
Samsung
Johnson &
PetroChina
Nestle
Roche
AMBEV
Google
ExxonMobil
Hathaway
Pingan
Johnson
Note: We have assigned all pre-IPO Internet companies a value of US$1 billion in January 2000. This seems generous, given Mark Zuckerberg was 14 at that
time. For other sectors, if the company was unlisted on January 1, 2000, we have taken the earliest available value.
But something changed over the last 20 years. Specifically, two things changed. First,
owning assets was no longer necessary when building a network. Second, a business
model where there is a cost to acquiring customers is unambiguously inferior to a
business model where customers have no choice but to buy from you (AT&T, the
regulated monopoly of old) or have no choice but to use your platform today (Instagram
today).
Again, in the Internet circa 2017, the decisions between Instagram or Snap may be as
arbitrary as the decision between Coke and Pepsi 30 years ago. But Pepsi (Snap in the
modern incarnation) really does taste better if all of your friends are drinking it too.
Customer acquisition is a prosaic consideration once you achieve scale. The question is
customer engagement. Therefore, among other things: why advertise?
Since 2000, globally, the companies that have gained the most in terms of absolute
increase in market value conform to one or both of those requirements (see Exhibit 136).
500%
450%
400%
350%
300%
250%
Telco
200%
Equipment
150% Telcos
Memory
100%
Internet
50% Processors / foundries
Apple
0%
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Apple designs its products in California. They are manufactured by others elsewhere.
During the rise of the iPhone a decade ago, AT&T and Verizon in the United States
were forced to negotiate for the exclusive right to sell the iPhone when it was first
released. The currency of that negotiation was the subsidy that the carrier would provide
to every new subscriber who purchased the phone. As a consequence, Apple outsourced
not just manufacturing but customer acquisition.
The benefit of network effects without having to build a physical network is clearly in
the Internet space. Facebook managed the transition from aspirational brand (launching
only at Ivy League schools to build cache and the image of exclusivity in 2004) to utility
(500 million users by 2010) before it IPO-ed. At that point, Facebook had a network
effect.
employed. In other words, (i) do not own assets if you do not need to and (ii) try not to
waste money on things like acquiring customers.
EXHIBIT 138: Market capitalization to capex ratio (rolling five-year average) (2004-16)
180
160
140
120
100
80 Internet
Telco Equipment
60
Apple
40
Processors / foundries
20 Telcos
Memory
0
2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
If these are the generic instructions for the kind of businesses that succeed in 2017, the
sectors that can most easily comply outside of Internet include component makers
(especially those with a technology advantage over the competition TSMC and
Samsung) that do not have to "own" retail customers' relationships. Samsung's move into
a globally recognized consumer brand (with higher SG&A spending) has hurt it over the
last half decade. Telcos appear to be learning, slowly.
The second mandate is a more obvious one: do not build assets if you do not need to.
Market capitalization as a ratio of capital spending reflects whether the market is valuing a
business based on its physical assets or on its intangible qualities. Not surprisingly, the
telecommunications companies and the semiconductor producers struggle on this
criterion. The Internet excels.
Of course, top-line growth still helps. However, it does not help as much as one might
think. Globally, the telecommunications companies have tripled revenues since 2000 (see
Exhibit 139) but are not even also-rans when you consider the winners and losers from
the changes in the marketplace over this period (see Exhibit 140).
EXHIBIT 139: Revenue contribution of key companies in the EXHIBIT 140: Revenue contribution of key companies in the
telecom sector (2000-16) Internet sector (2000-16)
750 350
Others
675 Netflix
300
600 Others Baidu
525 250 Alibaba
China
Telecom Tencent
450 200
China
USD B
USD B
Facebook
375 Mobile
150
300 DT Google
225 100
Verizon
150
50
75
Amazon
AT&T
0 0
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
Note: Others include T Mobile, Sprint, BT, Orange, and China Unicom. Note: Others include Yahoo!, YY, Kakao, Netease, 58.com, Naver, SINA, and
Weibo Corp.
Source: Bloomberg L.P. and Bernstein analysis.
Source: Bloomberg L.P. and Bernstein analysis.
THE PARALLELS AND There are two contradictory conclusions that can be distilled from all of this.
DIFFERENCES WITH THE
FUTURE OF TECHNOLOGY First, the Internet sector was the extraordinarily fortunate beneficiary over the last two
decades of investment by technology and telecommunications companies that did not, for
a moment, suspect that the value they were creating could be leached away so rapidly
and so efficiently by the platform that sat on top of the true innovation (see Exhibit 141).
Second, the next generation of companies responsible for the future of technology
augmented reality, virtual reality, artificial intelligence, voice recognition, and machine
learning will not be as cavalier in their approach. Business models are going to become
more sophisticated (see Exhibit 142).
Further, given that "buying the customer" and "building the asset" are now negative
characteristics in any business, the fragility of the right to capture value is changing in a
fundamental way. The rise and fall of the companies and sectors that will dominate our
thinking and investment controversies over the next decade will be far faster than over the
last 60 years. The transition from GE and AT&T to Alphabet and Apple may not be the
passing of the baton. We are just seeing the first and the second wave of pretenders to a
throne that no longer exists.
EXHIBIT 141: Financial metrics for sector constituents: Internet, processors/foundries, memory, telecom, and telecom
equipment
Revenue (USD B) Revenue Market Cap (USD B) Market Cap ROIC
2000 2016 CAGR (%) 2000 2016 CAGR (%) 2000 2016
Internet
Alibaba - 15,912 - - 219 - - 67%
Baidu - 10,624 - - 57 - - 16%
Tencent - 22,880 - - 232 - - 24%
Google 19 90,272 70% - 539 - - 38%
Facebook - 27,638 - - 332 - - 34%
Netflix 36 8,831 41% - 53 - - 6%
Amazon 2,762 135,987 28% 6 356 30% - 19%
Yahoo 1,110 5,169 10% 17 37 5% 11% -1%
Processors / foundries
Intel 33,726 59,387 4% 202 172 -1% 37% 17%
AMD 4,644 4,272 -1% 4 11 6% 22% -8%
Qualcomm 3,197 23,554 13% 62 96 3% 14% 49%
MediaTek Inc 412 8,549 21% - 11 - - 14%
TSMC 5,324 29,415 11% 28 145 11% 24% 33%
NVIDIA 375 5,010 18% 2 58 23% - 97%
Memory
Samsung 6,645 32,603 10% 19 210 16% 22% 20%
SK Hynix 9,492 14,830 3% 2 27 19% 12% 13%
Micron 6,362 12,399 4% 20 23 1% 31% 2%
Nanya 455 1,292 7% 1 4 7% 6% 8%
Telcos
AT&T 51,476 163,786 8% 162 261 3% 12% 7%
T Mobile US - 37,242 - - 47 - - 4%
Verizon 64,707 125,980 4% 135 218 3% 7% 14%
Sprint - 32,180 - - 34 - - 4%
British Telecom 30,152 28,712 0% 56 45 -1% 7% 16%
Deutsche Telecom 37,806 80,904 5% 91 81 -1% 3% 7%
Orange 31,097 45,289 2% 100 40 -5% 3% 6%
China Mobile 7,850 106,680 18% 102 217 5% 41% 19%
China Telecom 8,579 53,050 12% - 37 - 17% 6%
China Unicom 2,862 41,291 18% 19 28 2% 10% 1%
Telecom Equipment
ZTE Corp 546 15,114 23% - 9 - 13% 0%
Nortel Networks 27,948 - - 99 - - - -
CISCO Systems 18,928 49,247 6% 275 152 -4% 13% 38%
Ericsson 29,922 26,043 -1% 89 19 -9% - -
Alcatel Lucent 29,005 - - 69 - - - -
Motorola 37,580 6,038 -11% 44 14 -7% - 35%
Nokia 28,052 26,137 0% 209 28 -12% - 10%
Note: Only memory segment revenues included for Samsung. Market capitalization as of end 2000 and 2016.
AN ALTERNATE FUTURE
Life is fair in the long term. And the technology industry may be closer to realizing this
than we can currently fathom. There are many strains of thought that lead to this view a
future that contains more disruption, but no free lunches; and a future with relentless
change, but also greater friction, higher costs to scale, and lower returns on capital. When
this alternate future plays out, our current declarations of supreme victory for today's
Internet giants may look amusing in hindsight!
Indeed, the pinnacle of a company or a segment's power has often been looked at, in
retrospect, as the moment when the seeds of its downfall were sown. In 1964, Fortune
magazine feted Sears as an "extraordinary powerhouse of a company." By 1990,
headlines told a different tale "Sears let arrogance blind it to basic changes in the
American marketplace." Closer home in tech, Digital Equipment was feted in 1986 thus:
"Taking on Digitalthese days is like standing in front of a moving train." In 1986, there
were under 12 years left in Digital's existence as an independent company. The path to
"just desserts" has been shortening through the decades. BlackBerry (founded in 1984)
peaked in revenues in 2011 at nearly US$20 billion. Four years later (in 2015), its
revenues had collapsed to US$3.3 billion.
EXHIBIT 142: Capital Intensity versus marketing expenses matrix for technology opportunities
Augmented
Reality Telcos
Virtual Telco
Unattractive
Equipment
SG&A / operating income
Reality
High
Healthcare
Memory
Social
Data Energy
Artificial Storage
Attractive
Genetics foundries
Apple Space
High Low
Old Tech (sized by
market cap)
Attractive Unattractive
New Tech
Part of the all-pervasive presence of today's Internet giants comes from their youth. As a
society, we are enamored by their spectacular rise, by their out-sized influence on our
lives, and by the exciting future that may yet lie ahead. This fascination allows us to
rationalize the power and valuation of today's tech giants, as well as explain away the less
fortunate (telecom OEMs, networking providers, memory, and chip giants). Even at the
peak of their value and influence, the leaders of Cisco or Samsung or Intel were not
elevated to the rock star status that founders such as Elon Musk (or Steve Jobs before
him) enjoy today. Yet, the Internet revolution is under three decades old. The founder of
Facebook is in his 30s, the founder of Snap Inc., in his 20s, and the founders of many
Chinese Internet giants are still in their 40s (see Exhibit 143). The eventual denouement of
this drama lies ahead.
EXHIBIT 143: Bill Gates has seen it all; Bezos and Ma are merely greying
70
61
60 53 52
48
50 45 45 44 43
40
Years old
40 32
30 26
20
10
0
Bill Gates Jeff Bezos Jack Ma Robin Li Pony Ma Elon Musk Larry Page Sergey Brin Travis Mark Evan
Kalanick Zuckerberg Spiegel
Blame it on karma. The world of tech is no longer an upstart disruptor to the status quo
the David to many a Goliath. Gone are the days when a few kids in a garage changed the
world on a diet of pizza. Technology is the new establishment. It is front and center, the
mainstream of societal change and influence, the platform of abuse and angst, the
weapon law makers and breakers wish to possess, and the force ordinary humans are
increasingly wary of. Vaulting ambition, unrelenting innovation, ease of financing, and the
rising sophistication of computing and software have brought technology to a point where
the disruption is societal in a way that will be very different from the past.
This is no longer about PCs, smartphones, or tablets mere toys that enabled us to
connect and explore our world differently. This is about how we will live our lives, what
kind of jobs we hope to keep and lose, how we will educate and train our children for the
new world, how we will fight crime, and how geopolitics will be shaped. This is about how
and how long we shall live and how we shall approach old age. How do we grapple with
the implications? There are a few methods to frame this debate on future tech.
In the 1930s, Harold Hotelling, an economist, argued that there are certain kinds of goods
that are unique, and that everyone needs to have access to them. Roads and bridges,
railroads, and electricity are some of these. This led to the question: who should pay for
such goods? Hotelling argued that the government needs to finance (and implicitly own)
these public goods and, in turn, provide these to citizens at a marginal cost in effect,
general taxation would finance goods that were needed for society as a whole.
These thoughts, which form the philosophical basis of the Collaborative Commons
movement, are being revitalized (for example, by author Jeremy Rifkin in his recent work
"The Zero Marginal Cost Society"). Creative Commons licenses are not uncommon.
YouTube boasts a library under it, Wikipedia's entire content was brought under a
Creative Commons license in 2009 and hundreds of millions of Flickr photos exist under
the same arrangement. Elinor Ostrom became the first woman ever to win the Nobel Prize
for Economics in 2009, for her work on the history of the Commons through the last
thousand years.
Compare this with where the world of technology has come thus far, and where it is
leading us. Today's tech giants own and monetize minute behavioral characteristics that
we ourselves may not be fully aware of. By owning our social data, we are "targeted" with
ads we are supposed to find relevant, news that is expected to suit our bent, videos that
we are likely to watch, and items we are likely to buy. While championing the cause of "net
neutrality", tech majors "are walling off information posted by their users from the rest of
the web," to quote Tim Berners-Lee, founder of the Internet revolution.
This train of thought, therefore, leads us to the following conclusion. Sure, many parts of
the tech world benefited disproportionately from the infrastructure created by others
Internet giants stood on the shoulders of semiconductors, telecom equipment, and
service providers. This equation, which appears "unfair" to many as of now, could well
reverse in the future. The billions of dollars that many technology companies will likely
spend in space exploration, healthcare/gene therapy, and artificial intelligence will
eventually lead to products that will likely not exclusively belong to them. The companies
that appropriated the profits of others in the last two decades will likely have their profits
appropriated from them. In this alternate future, much of the current R&D spend in some
areas (AI, space, and genetics) is merely a hidden penalty for past appropriations, as the
benefits of these R&D outlays are unlikely to be allowed to remain in private hands (see
Exhibit 144).
#2: Net benefits will determine whether adoption wins over friction: how many are better
off versus worse off?
The ideas of extracting out capital expenditure and customer acquisition costs are
considered the lynchpins of the victory of Internet-based business models over those that
created the infrastructure in the first place. But this need not be Internet's smartest
business innovation. Something of far greater consequence has been this platforms
and applications of technology (i.e., Internet) have fundamentally managed to convince us
that they exist for the greater good of society, and that the world is a better place as a
result.
Alternate thoughts have existed around the propounded benefits for long. Harold
Mendelsohn, then Director of the Center of Mass Communications Research and Policy at
the University of Denver, proposed a different thought as far back as 1979:
The notion that new technologies in communications can be nothing but functional for
both individuals and society is the peculiarly naive consequence of the marriage
between neo-technocratic ideology and a more traditional unbridled positivism. It simply
is not true that every major technological innovation has spawned nothing but benefits
to the human race. The prospective dysfunctional consequences of future technologies
are far more realistically projected in the motion picture Star Wars, for example, than in
pie-in-the-sky, oh-my-gosh, "benefits to humanity" mutterings.
Technology has proved him wrong in the last three decades. Yet, there is the simmering
discontent rising about it, and against it. Countries now accuse each other of cyber
attacks. Fake news abounds. Terrorists and other unsavory types run their hiring models
on the back of Internet's platforms. New forms of gender discrimination, privacy
violations, spying, hacking, theft, and child abuse have emerged the book Future Crimes
by Marc Goodman is, thus, a sobering read. Silicon Valley would argue, with some
justification, that technology lays bare the wounds of society that were hidden for too
long. Yet, the same power creates new ones. The cycle of technology's, and especially
Internet's much touted benefits may be about to moderate. And therein lies the rub of how
society will react to those who wield such enormous financial power and political/societal
influence from San Francisco to Shanghai.
A simple ratio to define the intensity of disruption is this: number of people better
off/number of people worse off or displaced. We shall call it the net benefit ratio. The
higher this ratio, the more likely it is that the disruption will be embraced by society,
regulators, and investors. The lower this ratio, the more likely it is that new sources of
friction appear. Let us look at some past and emerging technologies in this light.
PCs and smartphones: About 3 billion people own smartphones globally. We can
consider they are all better off. Sure, parents may fret about the screen time children
enjoy, and studies may moan about the impact of smartphones on attention spans,
sleep, and social behaviors. But we are talking cold facts here. Three billion people
benefited in some way or the other. Who were worse off? The numbers are small.
Factory workers who produced feature phones (or pagers, before phones) were re-
deployed to smartphone manufacturing. Postal workers found new work pushing e-
commerce packets (although some 200,000 U.S. postal workers did lose their jobs in
2003-13). We could count the employees of Nokia, Palm, BlackBerry, and the likes,
perhaps, among the worse off, but even they benefited in their personal lives via the
rise of smartphones. Camera manufacturing was worse off. A few other segments
were also worse off, but all small. This makes for a very high net benefit ratio.
EXHIBIT 145: Percentage of jobs at risk from automation EXHIBIT 146: Probability an AI will replace this job
Telemarketer
UK
Accountant, auditor
Retail salesperson
US
Taxi drivers
Real estate sales agent
OECD average
Word processor, typist
Machinist
S. Africa
Commercial pilot
Economist
Argentina
Health technologist
Actor
Nigeria
Firefighter
Editor
India
Chemical engineer
Clergy
Thailand
Athletic trainer
Dentist
China
Recreational therapist
0% 20% 40% 60% 80% 100% 0.0% 40.0% 80.0% 120.0%
Source: World Bank Development Report, 2016, and Bernstein analysis. Source: "The Future of Employment" C. Frey and M Osborne (2013), and
Bernstein analysis.
Augmented and virtual reality (AR and VR), and healthcare tech: In contrast to AI-
automation-driverless cars, the rise of AR/VR and healthcare tech seems benign.
AR/VR technologies are meant to benefit via entertainment, information, and e-
commerce applications. Collectively, they could be the next avenue of the "escapism"
ethos that rode the wave of radio, movies, television, and social media. Many
consumers should be happy, very few (e.g., content creators on old formats) would
be worse off. Similarly, healthcare advances that use AI and new technologies would
benefit far more people than they harm, presumably. Sure, cancer pathologists will
resist robotic reading of MRI scans (studies show that machines are already more
accurate than humans) years of medical study for some may become redundant.
But the health benefits for humanity as a whole are likely to swamp out the
naysayers.
An approach that brings the "net benefit ratio" into focus preempts future conflict of
ownership, regulation, and influence. When Facebook was a social network, it was benign.
When it is the primary mode of news access for billions of people, it is under greater
scrutiny. When Google was an algorithmic search engine, it was welcomed. When the
same algorithms randomly place ads alongside videos or content deemed inappropriate,
it is under fire. These changes are subtle at first, but inevitably build toward an inflexion
point.
Take automation for instance. During 1995-2002, 22 million manufacturing jobs were
eliminated in the global economy, even as production rose 30%. In the United States,
between 1982 and 2002, steel production rose from 75 million tons to 120 million tons,
while steel workers declined from 289,000 to 74,000. These forces are far different in
impact than the fun factors that have dominated consumer tech in recent years. And the
backlash and debate around them may have just begun.
Exhibit 147 aims to present one scenario of net benefits across emerging technologies. In
this framework, AR/VR, energy storage and space are likely to face far less
debate/friction/challenge from wider society/government/regulators than AI,
automation, healthcare/genetics, and even social data gathering platforms such as
Facebook.
EXHIBIT 147: New technologies seen through the prism of the net benefit ratio
Beneficiaries Displaced Net Benefit Ratio
Augmented reality Shoppers, gamers, perhaps all connected society High
Virtual Reality Shoppers, gamers, perhaps all connected society High
Energy Storage All society Legacy energy infrastructure High
Space Perhaps all society High
Social data Advertisers, data owners (walled gardens) Privacy seekers, governments Medium
Healthcare, genetics Perhaps all society Poor patients, those on the wrong side of IP laws Medium
Automation Shareholders; all society if workers re-deployed Workers Medium
Artificial Intelligence Perhaps all society Perhaps all society Low
Angst against the on-demand economy has been rising. If Uber were to classify its
drivers as full-time workers, its costs could rise by about US$4.1-US$4.3 billion
annually (see Exhibit 148 and Exhibit 149). Given that Uber reportedly had losses of
some US$2.8bn in 2016, an additional cost of US$4bn could upend its entire
business model. Airbnb has been similarly accused of flouting local regulations of the
cities it operates in, distorting rental markets by taking 10-15% of vacant properties
out of circulation and of casting an inflationary impact on real effective rentals. This
has let loose a swarm of economic studies, variously sponsored by divergent groups,
which aim to prove their respective viewpoints. In some cities, Airbnb has had to
backtrack and change its practices. It is now illegal in some US cities, such as New
York, to given out accommodation for short term rentals in multi-family buildings
unless the host is present as well. After threatening to take the matter to court,
Airbnb dropped its lawsuit, reportedly when regulators assured the company that
such home owners, and not Airbnb, would be punished for violating the law. The
business model lives on, for now.
EXHIBIT 148: In 2015, there were some 327,000 Uber drivers EXHIBIT 149: Another calculation yields a similar additional
in United States, who would each cost some US$13,000 cost to Uber (US$4+ billion) if it was held accountable to all
more if they were classified as full-time workers local laws in the United States
$10,000
Workers'
$8,000 Comp
12%
$5,600
$6,000
$4,000 $3,060 $3,000 Payroll tax
15%
$2,000 $1,200
$285
Reimburse-
$- 401(k)
ment for
miles (self- plans
owned car 2%
usage)
63% Vacation,
sick days
3%
Source: StaffOne, Fortune, and Bernstein analysis. Source: Paychex data, Fortune, and Bernstein analysis.
The cost of doing business could be rising elsewhere as well. Facebook has recently
had to put considerable resources to fight fake news, including employing fact
checkers as well as spending on educational campaigns. Google has had to increase
manpower that scrutinizes videos uploaded on its websites, and change rules so that
no ads are rolled out for videos that have not had at least 10,000 views a reaction
against allegations that ads were being placed against inappropriate videos.
Importantly, none of the areas of investment for major technology platforms seem to
promise anywhere close to the profitability and ROCE that their core businesses
enjoyed. The margins on search, social media advertising, e-commerce, and video
gaming stand truly head and shoulders above what can be expected, in the best
case, from cloud computing, online video, electric vehicles, voice recognition, online
payments, fintech, and on-demand businesses.
The evolution of some of the core businesses, over the longer term, also points to
lower returns. Consider omni-channel commerce, for instance. Both Amazon and
Alibaba seem to be making early investments in brick-and-mortar retail a trend in
contrast to the very revolution they pioneered. How would an equilibrium state look
like in coming decades as the way we shop settles into an online-offline blend? Can
we picture higher asset intensity, more capex, and lower returns on capital?
"Now what they would like to do is use my pipes for free, but I ain't going to let them do
that because we have spent this capital and we have to have a return for it."
In our alternate future, there will come a time when several of today's most successful
platforms will have similar angst.
#4: Ten years of solitude: If tech is not taught humility, perhaps it will at least learn patience!
It might have taken three years for 800 million Chinese users to transition to 4G-enabled
smartphones, but not all future dreams suffer from an equivalent absence of friction. It
would be foolhardy for anyone to bet against the ability of humans to overcome the odds
and achieve things considered impossible at some point or the other. Yet, overcoming the
next frontier of tech, and achieving the ambitious dreams of the industry's current leaders,
could take longer. Are we heading into 10 years of solitude for tech's innovation
juggernaut? Consider this:
Alphabet's many projects (past and ongoing) have variously included research into
longevity (Calico), urban innovation (Side Walk Labs), floating balloons as a way of
delivering Internet (Project Loon), high altitude and solar-powered Internet delivering
drones (Project Titan), and drone deliveries (Project Wing). Compared to these,
others such as self-driving cars, smart home devices, and energy/Internet access
(fiber) almost appear tame. Yet, which among these have a path to scale in the next
10 years? Which among these would show network effects as fast as the prior
business models did?
Not everything around us follows Moore's Law. The cost of chip design has risen
through the decades due to the gap between the number of transistors on a chip
(following Moore's Law) versus the ability of chip experts to design the circuitry
around them (not following Moore's Law). Indeed, while logic transistors per chip
have grown at 58% annually, productivity of designers has merely improved by 21%
annually. Similar effects are now entering other fields of tech.
Internet of Things, as a concept, has been around for over a decade some would
say longer; a Coke machine installed at the Carnegie Mellon University in 1982 was
able to report its inventory and whether drinks were cold enough. At times, the
friction comes from areas little seen. Consider 3D printers. Their industrial
applications may be well worth some of the hype, but the long anticipated boom in
consumer demand (lets print an iPhone case on our home 3D printer, shall we?) has
seemingly hit a wall. Consumers need to know proper design tools to get value out of
them, and their demands are so specific that 3D printers have simply failed to keep
pace. Like in many other parts of tech, we are waiting.
Various years in the recent past have been declared the "Year of Virtual Reality"
Ernest Cline's VR-based fiction novel, Ready Player One, was published in 2011.
Steven Spielberg has been faster on the ball (a movie based on the novel is about to
release in 2018) versus tech behemoths themselves, who have had to satisfy
themselves with meagre shipments. Tablet shipments made a smaller peak than was
expected. Wearables disappointed even sooner. And at the time of this writing, Apple
was reportedly having problems incorporating an optical fingerprint sensor in its
much awaited next iPhone.
This may sound overly cynical, but technology advancements are not a linear function, but
much rather a step function. On the up-step, things change with disproportionate pace
we move from one paradigm to another. We seem to have lived through one such up-step
in the last decade and a half. On the flat lines of the step, improvements can be
incremental a painful series of small changes that eventually do lead to magical
outcomes (at least sometimes) but this also explains why the first tablet and the first
PDA did not take over the world.
In this alternate future, technology does not learn humility (by failing, or having their profits
appropriated away, or by being rejected by society and regulators) but merely patience. It
takes longer (and costs more) to achieve the new dreams we are beginning to see. Ideas
follow hope, then hype, and time passes as investments go on. Would the returns of tech's
supremely profitable platforms look as outlandish once this long trawl ensues?
EXHIBIT 150: Ten years of solitude? Perhaps the most benign outcomes among our alternate futures
MarketPrices
ClosingPrice(April21,2017) 22,385.00 2,003.00 168.00 5.05
PotentialUpside(Downside) 21% (0%) 16% 19%
52WeekLow 15,445.00 1,198.00 136.00 4.49
52WeekHigh 23,410.00 2,294.00 179.50 6.58
BernsteinEPSAnnualChange
FY16FY17E (22%) (7%) (6%) (531%)
FY17EFY18E 45% 32% 13% (53%)
FY18EFY19E 1% 4% 6% 126%
P/EonConsensusEPS
FY17E 44.5x 20.4x 23.4x 17.6x
FY18E 36.9x 18.0x 21.0x 15.7x
FY19E 32.6x 15.2x 18.4x 12.3x
P/EonBernsteinEPS
FY17E 35.0x 25.3x 23.3x 13.0x
FY18E 24.1x 19.2x 20.6x 27.5x
FY19E 23.7x 18.5x 19.5x 12.2x
Note:
Fiscal year for Fanuc, Yaskawa, and Lenovo ends on March 31; EPS data are for FY 2016A, FY 2017E, FY 2018E, and FY 2019E.
Fanuc and Yaskawa are benchmarked against the MSCI Japan Index, which had a closing price of 890.51 on April 21, 2017; Delta Electronics, and Lenovo
benchmarked against the MSCI Asia Pacific Excl. Japan Index, which had a closing price of 478.58 on April 21, 2017.
Source: FactSet, Bloomberg L.P., corporate reports, and Bernstein estimates and analysis.
Consensus EPS
2016A 1.13 36.86 171.01
2017E 1.77 208.21 233.61
2018E 2.35 240.15 302.27
Note:
For Sony (6758.JP), the base year is 2015.
Sunny Opticals and Largan are benchmarked against MXAPJ.
Sony (6758 JP) is benchmarked against MXJP. Sony also has a secondary listing: Ticker SNE had a closing price of US$34 (as of April 24, 017), is rated
outperform, and has a target price of US$41. The stock is benchmarked against the S&P 500.
These stocks were initiated on April 25, 2017 and so were priced as of April 24, 2017 (see our initiation report: Asian Emerging Technologies: The gift of
sound and vision initiating with a positive view).
Source: Bloomberg L.P., FactSet, corporate reports, and Bernstein estimate and analysis.
Price as of Apr 21, 2017 190.00 32.08 11.90 1.90 9.30 5.96 10.68 56.20 478.58 2,348.69
Currency TWD USD TWD USD HKD USD HKD TWD
Target Price 205.00 34.00 11.90 1.90 11.20 7.20 12.00 47.00
52-Week Range 143-195 22.8-33.5 11.2-13 1.7-2.1 5.8-12.3 3.8-7.7 6.4-11.3 48.2-69.5
Market Capitalization (billion) 4,926.8 162.3 150.2 4.9 43.2 5.6 11.0 92.1
TTM Performance 21.0% 27.5% -5.2% -3.6% 38.8% 38.6% 35.9% 9.8%
TTM Relative Performance 8.5% 15.2% -17.7% -15.9% 26.3% 26.3% 23.4% -2.8%
Bernstein EPS Forecast
2016A 12.89 2.00 0.68 0.10 0.09 0.45 0.12 3.38
2017E 13.74 2.19 0.57 0.09 0.09 0.43 0.13 3.38
2018E 14.36 2.29 0.68 0.11 0.10 0.48 0.14 3.77
2019E 16.46 2.62 0.82 0.13 0.11 0.57 0.15 4.00
Shares Outstanding (mil.) 25,930 5,186 12,624 2,525 4,649 930 1,034 1,639
Dividend Yield 3.9% 3.9% 4.2% 4.2% 0.0% 0.0% 2.7% 5.3%
Dividend per Share 7.50 1.27 0.50 0.08 0.00 0.00 0.04 3.00
MediaTek Novatek ASE ASE ADR SPIL SPIL ADR TEL MXAPJ MXJP SPX
2454.TT 3034.TT 2311.TT ASX 2325.TT SPIL 8035.JP
Rating M M O O O O O
Price as of Apr 21, 2017 215.00 113.00 37.85 6.25 49.00 8.03 12,300 478.58 890.51 2,348.69
Currency TWD TWD TWD USD TWD USD JPY
Target Price 219.00 100.00 45.00 7.30 49.00 7.90 13,400
52-Week Range 192-261 102-128 28-39.9 4.1-6.6 43.2-54.5 6.6-8.4 6961-12780
Market Capitalization (billion) 340.2 68.8 312.1 10.3 152.7 5.0 2,032.1
TTM Performance -8.1% 6.1% 10.0% 14.1% -2.0% 5.4% 56.9%
TTM Relative Performance -20.6% -6.4% -2.5% 1.8% -14.5% -6.9% 51.2%
Bernstein EPS Forecast
2016A 15.16 8.22 2.83 0.44 3.19 0.50 461.10
2017E 18.54 8.11 3.13 0.51 3.63 0.59 631.18
2018E 21.11 9.13 3.44 0.56 3.96 0.64 757.06
2019E 20.63 9.33 3.75 0.60 4.34 0.70 835.86
Shares Outstanding (mil.) 1,582 609 8,245 1,649 3,116 623 165
Dividend Yield 5.1% 6.2% 4.5% 4.5% 5.5% 5.5% 2.5%
Dividend per Share 11.06 6.99 1.70 0.28 2.71 0.44 305.02
Note: Stocks trading in Taiwan and Hong Kong are benchmarked against the MXAPJ; stocks trading in Japan are benchmarked against the MXJP; and all ADRs
are benchmarked against the S&P 500.
Source: Bloomberg L.P., FactSet, corporate reports, and Bernstein estimates and analysis.
Consensus EPS
2016A 0.54 0.07 4.96
2017E 0.59 0.14 5.39
2018E 0.68 0.16 5.93
2019E 0.77 0.20 6.54
Note: The 2016 data for BYD is actual result, which we haven't reflected in our published model.
Source: Bloomberg L.P., FactSet, corporate reports, and Bernstein estimates and analysis.
Note: U.S. stocks are benchmarked against the S&P 500, which had a closing price of 2,348.69 as of April 21, 2017. Tencent is benchmarked against the
MXAPJ, which had a closing price of 478.58 as of April, 21, 2017.
Source: Bloomberg L.P., FactSet, corporate reports, and Bernstein estimates and analysis.
ASIAN IT HARDWARE
VALUATION METHODOLOGY Fanuc: Due to the exposure to the machine tool sector, Fanuc's earnings base is cyclical
by nature. Since we expect CY2017 to be a capex upcycle, we apply a forward P/E
multiple of 29x to our FY2018 EPS estimates, equivalent to one standard deviation above
its five-year historical average and also in line with its trading multiple in the last capex
upcycle.
Yaskawa: In line with our valuation methodology for Fanuc, we apply a forward P/E
multiple of 19x to our FY2018 EPS estimates, equivalent to one standard deviation above
Yaskawa's five-year historical average and also in line with its trading multiple in the last
capex upcycle. Our valuation multiple for Yaskawa (19x) implies a 34% discount to Fanuc
(29x), in line with the stocks' five-year historical trading average.
Delta: Diversified technology companies have more stable cash flows and reliably positive
earnings, hence we triangulate using historical P/E multiples and DCF model to derive our
valuation. Informed by our DCF model, we apply a P/E multiple of 24x to our CY2017 EPS
estimates.
Lenovo: Our target price for Lenovo is based on 15.0x on FY2019 (ending March 31,
2019) EPS estimates.
RISKS Fanuc: Our outperform rating and target price for Fanuc are based on our belief that
CY2017 will likely see a capex cycle upturn and that Fanuc will stand to benefit. However,
if macroeconomic conditions deteriorate significantly, the JPY appreciates notably in a
short time, or given any significant changes to the competitive landscape in the CNC
market (which we deem highly unlikely), there could be downside to our current target
price.
Yaskawa: Our market-perform rating and target price for Yaskawa are based on our belief
that Yaskawa will continue to lose share in all its key markets and that it is unable to
achieve material margin expansion due to FX headwinds. Risks to our target price include
significant JPY appreciation/depreciation and faster/slower-than-expected macro
recovery.
Delta: Our target price for Delta Electronics is based on our belief that it can execute its
strategies to grow non-power supply segments, which may not happen as fast as
expected due to market competition, macroeconomic conditions, technological change,
and product transition. End customers may allocate orders to other suppliers to diversify
the supply base.
Lenovo: Our outperform rating and target price for Lenovo are based on our belief that its
Mobile and Enterprise businesses will consistently trim their losses and expand corporate
margins, and the company can successfully turn around the emerging market business.
However, if the overall PC sector deteriorates faster than expected, or the company fails
to generate sustainable profits from emerging markets, Mobile, and Enterprise, there
could be downside from our current target price.
VALUATION METHODOLOGY We value companies in our coverage using price-to-forward earnings. The P/E is
determined by historical range, future growth rate, and as informed by our DCF model.
Sunny Optical: We value Sunny Optical on a forward P/E basis. We believe that the last
one-year average P/E ratio of 23.5x reflects the strong growth ahead and, hence, use it
as our target P/E. The EPS range used is 2Q18-1Q19, as it is the NTM range for our one-
year target price.
Sony: We value Sony on SOTP by applying forward P/E of its individual segments, and
with a 15% conglomerate discount. The EPS range used is 2Q18-1Q19, as it is the NTM
range for our one-year target price. For Sony's portion of Sony Financial Holdings, we use
its market value.
Largan: We value Largan on a forward P/E basis. The target P/E multiple of 16.4x is the
last one-year average P/E, as we believe it reflects the company's growth going forward.
The EPS range used is 2Q18-1Q19, as it is the NTM range for our one-year target price.
Technology risk: Companies in our sector rely on the evolution of technologies in the
future. If the future technologies evolve in a faster or slower fashion, it would present
upside or downside risk to our companies.
Macro risk: A weak macro environment would depress consumer IT spending (e.g.,
smartphones and tablets) and impact our companies, and vice versa.
Product risk: Our companies supply consumer products directly or indirectly. Changes in
demand of such specific products would affect the performance of our companies.
FX risk: Many of our companies sell products priced in a different currency than its
reporting currency. Significant moves of the exchange rate would impact the financials of
these companies.
Sunny Optical: Downside risks to our target price for Sunny Optical include: lower
smartphone OEM shipment, especially from Chinese OEMs; slower dual camera and other
sensing camera penetration in smartphones; slower ADAS adoption in vehicles; and
inability to gain market share from competitors.
Sony: Downside risks to our target price for Sony include: slower VR adoption in its game
consoles; share loss of game console to its key competitors, Microsoft and Nintendo;
lower demand of image sensors in consumer products, especially smartphones; lower
demand of consumer electronics, such as cameras, TVs, and speakers; lower box office of
films carried by its studios; and weaker performance of Sony Financial Holdings, a public
company in which Sony has a majority holding.
Largan: Upside and downside risks to our target price for Largan include: higher or lower
smartphone market shipment, especially iPhones; faster or slower dual camera
penetration in smartphones; gain or loss of market share; and delay of the next-gen
iPhone.
VALUATION METHODOLOGY Taiwan Semiconductor Manufacturing Co Ltd: Given the cyclical nature of the financials
and high incidence of low or negative net income periods, we follow industry practice and
value this company on a trailing P/BV basis, intangible assets and goodwill excluded. To
reflect the structural profitability challenges and mounting competition in the next few
years, we value TSMC at 3.3x NTM BV, intangible assets and goodwill excluded.
United Microelectronics Corp: Given the cyclical nature of the financials and high
incidence of low or negative net income periods, we follow industry practice and value
this company on a trailing P/BV basis, intangible assets and goodwill excluded. In light of
the faster EPS expansion, we value UMC at 0.7x NTM BV excluding intangible assets and
goodwill, which is below the average since 2011.
Semiconductor Manufacturing International Corp (SMIC): Given the cyclical nature of the
financials and high incidence of low or negative net income periods, we follow industry
practice and value this company on a trailing P/BV basis, intangible assets and goodwill
excluded. We use 1.4x NTM BV excluding intangible assets and goodwill, as our target
valuation. Since mid-2013, SMIC's P/B has been hovering around 1-1.5x. We choose to
use 1.4x, to reflect the recent improvement on return on equity.
Hua Hong Semiconductor Ltd: We value Hua Hong at 1x NTM book value excluding
intangible assets and goodwill to reflect its growth and ROE.
Siliconware Precision Industries Co (SPIL): Our target price for SPIL is set at NT$49, based
on the dividend-adjusted offer price of NT$51.8 (NT$55-3.8) after applying a 5%
discount due to uncertainties that still exist.
MediaTek Inc: We use price to one-year forward earnings as our valuation method for
MediaTek as we believe it best captures the earnings growth momentum of a fabless
company and is easier to compare among peers. To determine a relevant multiple, we
review the historical forward P/E multiples at the company level and set our multiple at
12x one-year forward earnings. Note that we exclude the gain from the transaction of
selling Autochips to NavInfo as it is one-off.
Tokyo Electron Ltd: We use forward P/E to value the company. Our forward P/E multiple
is set at 17x, slightly above the average P/E multiple since early 2015.
RISKS Asian Semiconductors and Semiconductor Equipment: The Asian semiconductors and
semiconductor equipment sector is exposed to general macroeconomic risks, which
ultimately influence end-market demand and can positively or negatively affect the entire
electronics supply chain. In addition to the effects of macroeconomic factors, the
semiconductor industry has historically been subject to seasonal and cyclical demand
fluctuations for its products or services. These industry downturns have been
characterized by diminished product demand, excess manufacturing capacity,
subsequent erosion of average selling prices, etc. Any factors affecting the
semiconductor industry can impose an upside or a downside risk to our forecasts. We
highlight some important risks below:
Supply risk. Disruption and imbalance in the electronics supply chain either at a
semiconductor company, its suppliers/distributors, or its customers will create a
chain effect and impact the company's ability to fully utilize its capacity and deliver its
financial performance as planned. This risk can be seen in the raw material supply as
it sometimes accounts for a large portion of the manufacturing costs, and any supply
shortage can lead to reduced shipments or increased costs.
Product risk. Any major delay in the company's new product launches can result in a
downside risk to our forecasts. This risk is especially remarkable for IC design
companies as the market window for them is often very narrow. Missing the market
window due to product delays can result in much smaller market share and inferior
pricing.
Taiwan Semiconductor Manufacturing Co Ltd: There are a number of risks to our TSMC
target price, including:
Share gain at major customers like Apple and Qualcomm, against Samsung;
Macroeconomic risk would change end-market demand and eventually impact the
upstream players such as TSMC;
Exchange rate fluctuation posing a substantial risk to TSMC's revenue and profit; and
Disruption and imbalance in the electronics supply chain either on TSMC, on its
suppliers/distributors, or on its customers creating a chain effect and impacting
the company's ability to fully utilize its capacity.
United Microelectronics Corp: There are a number of risks to our UMC target price,
including:
Favorable exchange rate fluctuation posing an upside risk to UMC's revenue and
profit.
Hua Hong Semiconductor Ltd: There are a number of downside risks to our Hua Hong
target price, including:
Vanguard International Semiconductor Corp: There are a number of risks to our Vanguard
target price, including:
Exchange rate fluctuation posing a substantial risk to its revenue and profit.
Advanced Semiconductor Engineering Inc: There are a number of risks to our ASE target
price, including:
Macroeconomic risk would change end-market demand and eventually impact the
upstream players such as ASE;
Exchange rate fluctuation posing a substantial risk to ASE's revenue and profit;
Disruption and imbalance in the electronics supply chain either on ASE, on its
suppliers/distributors, or on its customers will create a chain effect and impact the
company's ability to fully utilize its capacity; and
Siliconware Precision Industries Co: There are a number of risks to our SPIL target price,
including:
Macroeconomic risk would change end-market demand and eventually impact the
upstream players such as SPIL;
Exchange rate fluctuation posing a substantial risk to SPIL's revenue and profit;
Disruption and imbalance in the electronics supply chain either on SPIL, on its
suppliers/distributors, or on its customers would create a chain effect and impact
the company's ability to fully utilize its capacity; and
MediaTek Inc: There are a number of downside risks to our MediaTek target price,
including:
There are a number of upside risks to our MediaTek target price, including:
Novatek Microelectronics Corp: There are a number of risks to our Novatek target price,
including:
Tokyo Electron Ltd: There are a number of risks to our TEL target price, including:
Slowdown in global semiconductor demand, including both logic and memory, which
trigger reduced capex and capacity expansion; and
ASIA-PACIFIC HEALTHCARE
VALUATION METHODOLOGY Healthcare services: Hospital stocks in Asia trade at very high (even triple-digit multiples in
some cases) and wide-ranging valuations. We value the pure-play hospital operators in
our coverage IHH, Bangkok Dusit Medical Services, and Bumrungrad using a
blended average of three approaches that capture different aspects of value cash
flows, earnings, and returns.
Second, we apply future year 2018 P/E peer group multiples to Bernstein estimates for
2019 EPS and discount back to 12-month target prices. We do it this way because
current year P/Es are high and widely variable across the group. Tremendous macro
growth drivers (policy, demographics, and epidemiology) combined with relative scarcity
of high-quality, investable healthcare names lead to high heterogeneity and high P/E
multiples (triple digit in some cases) that can't be justified by fundamentals. In future
years, expectations converge around a lower and narrower (more reasonable) valuation
range.
RISKS Not all good health technology ideas will succeed (no different than other corners of
technology). Good ideas and unmet needs alone are not enough; many factors contribute
to the success of disruptive technologies. Gartner's famous "technology hype cycle"
(technology trigger, peak of inflated expectations, trough of disillusionment, slope of
enlightenment, and plateau of productivity) has been adapted to digital health53
investors should consider it a cautionary warning that there will be several failures as
health technology evolves. Increased health connectivity has a dark side too. IDC predicts
that Chinese medical institutions will suffer 20% more blackmail virus attacks.54 Beyond
risks specific to the types of health technology companies discussed in this chapter
(almost all of which are outside our coverage), below we also discuss risks associated
with investing in the types of Asian healthcare stocks that we do cover and mention in the
Blackbook.
Healthcare services
Downside risks. Worsening economic outlook for any of the countries we cover
China, Thailand, Malaysia, or Singapore would create downward pressure on
volumes by reducing the number of people who can afford out-of-pocket healthcare
expenditure. Civil unrest or outbreak of MERS or other infectious disease in the
region would negatively affect medical tourism from overseas, and deter uninfected
local people from visiting hospitals.
53
See http://bionic.ly/digital-health-hype-cycle/
54
IDC FutureScape: Top10 Predictions of Healthcare IT 2017 China Implications
VALUATION METHODOLOGY Samsung SDI Co Ltd: The W160k target price for Samsung SDI is based on 1.0x our
2017E BVPS, or a 22% discount to our SOTP fair value.
BYD Co Ltd: We derive our target prices of HK$60 for 1211.HK and RMB55 for
002594.CH using SOTP, correlating to 20x our 2018 EPS (well below historical average
of 26.5x for H and 45.7x for A shares)
RISK Samsung SDI Co Ltd: Samsung SDI's earnings growth depends on the adoption of electric
vehicles and energy storage systems to boost battery revenues and profits. Any change in
strategy by automakers, or lack of cost declines would reduce this upside. In addition,
display still plays a large role on the equity income line. Small battery profit recovery
depends on the utilization of its polymer lines improving, which in turn depends on orders
from customers, including parent Samsung Electronics. Risks to display (driving equity
income) include supply/demand balance pressuring pricing and, hence, margins
BYD Co Ltd: Key risks to our target price (on H-shares) include the execution on its EV
business, which generates the majority of profits, growth, and 99% of the value of the
company. Rising competition from domestic OEMs and falling subsidies in particular
could lead to falling margins and reduced profit growth worse than our estimates. In
addition, execution on the battery business drives much of the cost decline in EVs, which
is necessary to ensure profitability. Given the importance of retail investors to the A-share
markets, A-share listed stocks may be relatively more volatile than their H-share listed
counterparts. Upside or downside risks could come from Chinese government policies as
China looks to control the rate of growth of its economy in general, or capital markets in
particular. These policies may manifest in market rules that affect A- and H- shares
differently.
ASIA INTERNET
VALUATION METHODOLOGY We derive the target prices for our coverage of China's Internet stocks using discounted
cash flows over 10 fiscal periods into the future. These target prices are also triangulated
with price-to-adjusted EPS and EV-to-adjusted EBITDA multiples for 2018, with an eye
on growth prospects beyond 2018.
RISKS 58.com Inc (WUBA): Having disappointed for much of 2016, WUBA headed into 2017
with significantly lowered expectations and valuations. Any pricing improvement will be a
major reversal from recent trends, while benefits could show up stronger from the
ongoing rationalization of sales teams between 58.com and acquired entities Ganji and
Anjuke.
Alibaba Group Holding Ltd (BABA): Video losses could be steeper than our projections, if
competitive intensity worsens, posing downside to company-wide margins. We have
assumed that Tencent's challenge in payments and cloud will not involve blood-letting,
but it could delay monetization nonetheless. Among these spaces, Tencent has reportedly
gained considerable share in payments this could put Ant Financial on the defensive,
with some of the impact accruing to Alibaba as well. GMV growth in e-commerce has
moderated, but ad monetization has ensured revenues grow faster than GMV. Our
assumptions that Alibaba will defend share in advertising, while Baidu and others finance
Tencent's gains, need to be tested. JD's challenge continues to expand across more
categories while flow share trends have stabilized for several quarters now, a steady
picture for e-commerce market share is yet to emerge. There are also pending SEC
investigations into GMV reporting and loss recognition from some affiliates these could
continue to be in the news. Further, Alibaba was recently added back to the U.S. blacklist
of "notorious marketplaces," an indication that controversies around alleged counterfeits
on the platform will linger.
Baidu.com (BIDU): A retreat from O2O is a binary event for Baidu stock. Profits could
double if this initiative is abandoned, creating a significant "pop" to the upside for the
stock. While the current guidance is for a gradual reduction in spending, an outright
transaction risk always exists to the upside. Shorting this stock is, therefore, inadvisable.
The recovery in search advertising, after the controversies of 2016, could be faster than
our projections (in the teens). In video, Baidu seems cognizant of balancing premium with
free user-generated content a favorable equilibrium could surprise on the upside for
video margins. Baidu has also done well on video curation, and there are examples of TV
series bought relatively cheaply that have gone on to be very popular with viewers. These
are hidden skills that are often ignored by the markets, but could help differentiate the
Baidu video platform in coming years. Baidu has also invested in artificial intelligence, and
its efforts on both natural language processing and autonomous driving are well
advanced even by global standards. What it seems to lack is a path to commercialization
some of which could be addressed though synergistic alliances with hardware OEMs.
To the downside, its AI team is at risk of splintering post the exit of Chief Scientist Dr.
Andrew Ng competitors will take note of the talent resident in Baidu's AI labs in
California and Beijing. A new management team is in place, including the President and
COO, and the contours of future strategy and execution are still being defined.
Ctrip.com International Ltd (CTRP): Amid steady top-line growth, the quantum of margin
improvements is a debate about intensity, not direction. Ctrip's stock could perform
better, if margins move up faster. Downside risks come from M&A activity, related dilution
risks, fund-raising, and still extant discounting in some parts of the market (e.g., budget
hotels).
JD.com Inc (JD): Having added users, warehouses, and delivery centers faster than GMV in
recent quarters, JD could look forward to some scale benefits, which could surprise on the
upside. GMV has been somewhat depressed by JD's self-initiated culling of some
merchants to reduce undesirable practices a bounce back could elate investors. JD has
also seemed more welcoming to hire temporary workers lately, especially after the
acquisition of the Walmart business, which operated on a temporary worker model. This
could aid fulfillment cost control. On the flip side, expansion of new categories such as
FMCG will continue to challenge logistics for the business.
Netease.com (NTES): Onmyoji, a hit game, was launched in late September 2016. This,
alongside the successful licensed game, Overwatch, are yet to annualize in the financials.
While recent data seems to indicate that the growth of Onmyoji is slowing, these new
games will cast an upward bias to YoY growth numbers for at least two more quarters.
Cross-border e-commerce targets the affluent urban user, and its economics may
surprise on the upside, reducing the margin dilution we fear. NetEase also owns excellent
portal properties it could press harder for more advertising revenues to make up for any
gaming slowdown. On the downside, a private label domestic ecommerce business has
begun to grow, and will dilute gross margins.
Tencent Holdings Ltd (700.HK): To the upside, Tencent could press harder for more
advertising revenue than our projections (a 45% CAGR). Its ad load is low enough for such
scenarios. In gaming, it enjoys superior distribution, and our projected slowdown contains
timing risks: Committed gamers could prove more generous with their time and wallets
than we suspect, at least in the near term. Share gains in video, cloud, and payments could
show up sooner than we estimate, especially in payments, where inroads are apparent
this could excite markets. To the downside, the extent of losses in video, cloud, and
payments need watching. Tencent has also become more vocal on investments in artificial
intelligence costs of the new AI expansion are not yet factored into our forecasts.
Vipshop Holdings Ltd (VIPS): The company could find a better balance among user
acquisition, category expansion, and margin defense than we expect. Any improvement in
per-customer order flow and GMV will be greeted positively. Any substantial changes in
the discovery process on the website and the mobile app via a better search function
could help improve customer acquisition and retention, creating upside possibilities.
RISKS
See the "Valuation Methodology and Risks" chapter of this Blackbook.
Bernstein analysts are compensated based on aggregate contributions to the research franchise as measured by account penetration,
productivity and proactivity of investment ideas. No analysts are compensated based on performance in, or contributions to, generating
investment banking revenues.
Bernstein rates stocks based on forecasts of relative performance for the next 6-12 months versus the S&P 500 for stocks listed on the U.S.
and Canadian exchanges, versus the MSCI Pan Europe Index for stocks listed on the European exchanges (except for Russian companies),
versus the MSCI Emerging Markets Index for Russian companies and stocks listed on emerging markets exchanges outside of the Asia
Pacific region, and versus the MSCI Asia Pacific ex-Japan Index for stocks listed on the Asian (ex-Japan) exchanges - unless otherwise
specified. We have three categories of ratings:
Outperform: Stock will outpace the market index by more than 15 pp in the year ahead.
Market-Perform: Stock will perform in line with the market index to within +/-15 pp in the year ahead.
Underperform: Stock will trail the performance of the market index by more than 15 pp in the year ahead.
Not Rated: The stock Rating, Target Price and/or estimates (if any) have been suspended temporarily.
As of 04/25/2017, Bernstein's ratings were distributed as follows: Outperform - 45.5% (0.3% banking clients) ; Market-Perform - 40.1%
(0.0% banking clients); Underperform - 14.3% (0.0% banking clients); Not Rated - 0.2% (0.0% banking clients). The numbers in parentheses
represent the percentage of companies in each category to whom Bernstein provided investment banking services within the last twelve (12)
months.
Bernstein currently makes a market in the following companies BIDU / Baidu.com, NTES / Netease.com.
This research publication covers six or more companies. For price chart disclosures, please visit www.bernsteinresearch.com, you can also
write to either: Sanford C. Bernstein & Co. LLC, Director of Compliance, 1345 Avenue of the Americas, New York, N.Y. 10105 or Sanford C.
Bernstein Limited, Director of Compliance, 50 Berkeley Street, London W1J 8SB, United Kingdom; or Sanford C. Bernstein (Hong Kong)
Limited , Director of Compliance, Suites 3206-11, 32/F, One International Finance Centre, 1 Harbour View Street,
Central, Hong Kong, or Sanford C. Bernstein (business registration number 53193989L) , a unit of AllianceBernstein (Singapore) Ltd. which
is a licensed entity under the Securities and Futures Act and registered with Company Registration No. 199703364C, Director of
Compliance, 30 Cecil Street, #28-08 Prudential Tower, Singapore 049712.
OTHER DISCLOSURES
A price movement of a security which may be temporary will not necessarily trigger a recommendation change. Bernstein will advise as and
when coverage of securities commences and ceases. Bernstein has no policy or standard as to the frequency of any updates or changes to
its coverage policies. Although the definition and application of these methods are based on generally accepted industry practices and
models, please note that there is a range of reasonable variations within these models. The application of models typically depends on
forecasts of a range of economic variables, which may include, but not limited to, interest rates, exchange rates, earnings, cash flows and risk
factors that are subject to uncertainty and also may change over time. Any valuation is dependent upon the subjective opinion of the analysts
carrying out this valuation.
Bernstein produces a number of different types of research product including, among others, fundamental analysis and quantitative analysis.
Recommendations contained within one type of research product may differ from recommendations contained within other types of
research product, whether as a result of differing time horizons, methodologies or otherwise.
This document may not be passed on to any person in the United Kingdom (i) who is a retail client (ii) unless that person or entity qualifies as
an authorised person or exempt person within the meaning of section 19 of the UK Financial Services and Markets Act 2000 (the "Act"), or
qualifies as a person to whom the financial promotion restriction imposed by the Act does not apply by virtue of the Financial Services and
Markets Act 2000 (Financial Promotion) Order 2005, or is a person classified as an "professional client" for the purposes of the Conduct of
Business Rules of the Financial Conduct Authority.
This document may not be passed onto any person in Canada unless that person qualifies as "permitted client" as defined in Section 1.1 of NI
31-103.
To our readers in the United States: Sanford C. Bernstein & Co., LLC is distributing this publication in the United States and accepts
responsibility for its contents. Any U.S. person receiving this publication and wishing to effect securities transactions in any security
discussed herein should do so only through Sanford C. Bernstein & Co., LLC.
To our readers in the United Kingdom: This publication has been issued or approved for issue in the United Kingdom by Sanford C. Bernstein
Limited, authorised and regulated by the Financial Conduct Authority and located at 50 Berkeley Street, London W1J 8SB, +44 (0)20-7170-
5000.
To our readers in member states of the EEA: This publication is being distributed in the EEA by Sanford C. Bernstein Limited, which is
authorised and regulated in the United Kingdom by the Financial Conduct Authority and holds a passport under the Markets in Financial
Instruments Directive.
To our readers in Hong Kong: This publication is being distributed in Hong Kong by Sanford C. Bernstein (Hong Kong) Limited
, which is licensed and regulated by the Hong Kong Securities and Futures Commission (Central Entity No. AXC846). This publication is
solely for professional investors only, as defined in the Securities and Futures Ordinance (Cap. 571).
To our readers in Singapore: This publication is being distributed in Singapore by Sanford C. Bernstein, a unit of AllianceBernstein (Singapore)
Ltd., only to accredited investors or institutional investors, as defined in the Securities and Futures Act (Chapter 289). Recipients in
Singapore should contact AllianceBernstein (Singapore) Ltd. in respect of matters arising from, or in connection with, this publication.
AllianceBernstein (Singapore) Ltd. is a licensed entity under the Securities and Futures Act and registered with Company Registration No.
199703364C. It is regulated by the Monetary Authority of Singapore and located at One Raffles Quay, #27-11 South Tower, Singapore
048583, +65-62304600. The business name "Bernstein" is registered under business registration number 53193989L.
To our readers in the Peoples Republic of China: The securities referred to in this document are not being offered or sold and may not be
offered or sold, directly or indirectly, in the People's Republic of China (for such purposes, not including the Hong Kong and Macau Special
Administrative Regions or Taiwan), except as permitted by the securities laws of the Peoples Republic of China.
To our readers in Japan: This document is not delivered to you for marketing purposes, and any information provided herein should not be
construed as a recommendation, solicitation or offer to buy or sell any securities or related financial products.
For the institutional client readers in Japan who have been granted access to the Bernstein website by Daiwa Securities Group Inc. (Daiwa),
your access to this document should not be construed as meaning that Sanford C Bernstein is providing you with investment advice for any
purposes. Whilst Sanford C Bernstein has prepared this document, your relationship is, and will remain with, Daiwa, and Sanford C Bernstein
has neither any contractual relationship with you nor any obligations towards you.
To our readers in Australia: Sanford C. Bernstein & Co., LLC, Sanford C. Bernstein Limited and Sanford C. Bernstein (Hong Kong) Limited
are exempt from the requirement to hold an Australian financial services licence under the Corporations Act 2001 in respect
of the provision of the following financial services to wholesale clients:
To our readers in Canada: If this publication is pertaining to a Canadian domiciled company, it is being distributed in Canada by Sanford C.
Bernstein (Canada) Limited, which is licensed and regulated by the Investment Industry Regulatory Organization of Canada ("IIROC"). If the
publication is pertaining to a non-Canadian domiciled company, it is being distributed by Sanford C. Bernstein & Co., LLC, which is licensed
and regulated by both the U.S. Securities and Exchange Commission and the Financial Industry Regulatory Authority ("FINRA") into Canada
under the International Dealers Exemption. This publication may not be passed onto any person in Canada unless that person qualifies as a
"Permitted Client" as defined in Section 1.1 of NI 31-103.
Sanford C. Bernstein & Co., LLC., Sanford C. Bernstein Limited, Sanford C. Bernstein (Hong Kong) Limited , Sanford C.
Bernstein (Canada) Limited and AllianceBernstein (Singapore) Ltd. are regulated by, respectively, the Securities and Exchange Commission
under U.S. laws, by the Financial Conduct Authority under U.K. laws, by the Hong Kong Securities and Futures Commission under Hong Kong
laws, by the Investment Industry Regulatory Organization of Canada and by the Monetary Authority of Singapore under Singapore laws, all of
which differ from Australian laws.
One or more of the officers, directors, or employees of Sanford C. Bernstein & Co., LLC, Sanford C. Bernstein Limited, Sanford C. Bernstein
(Hong Kong) Limited , Sanford C. Bernstein (Canada) Limited, Sanford C. Bernstein (business registration number
53193989L), a unit of AllianceBernstein (Singapore) Ltd. which is a licensed entity under the Securities and Futures Act and registered with
Company Registration No. 199703364C, and/or their affiliates may at any time hold, increase or decrease positions in securities of any
company mentioned herein.
Bernstein or its affiliates may provide investment management or other services to the pension or profit sharing plans, or employees of any
company mentioned herein, and may give advice to others as to investments in such companies. These entities may effect transactions that
are similar to or different from those recommended herein.
Bernstein Research Publications are disseminated to our customers through posting on the firm's password protected website,
www.bernsteinresearch.com. Additionally, Bernstein Research Publications are available through email, postal mail and commercial
research portals. If you wish to alter your current distribution method, please contact your salesperson for details.
Bernstein and/or its affiliates do and seek to do business with companies covered in its research publications. As a result, investors should
be aware that Bernstein and/or its affiliates may have a conflict of interest that could affect the objectivity of this publication. Investors
should consider this publication as only a single factor in making their investment decisions.
This publication has been published and distributed in accordance with Bernstein's policy for management of conflicts of interest in
investment research, a copy of which is available from Sanford C. Bernstein & Co., LLC, Director of Compliance, 1345 Avenue of the
Americas, New York, N.Y. 10105, Sanford C. Bernstein Limited, Director of Compliance, 50 Berkeley Street, London W1J 8SB, United
Kingdom, or Sanford C. Bernstein (Hong Kong) Limited , Director of Compliance, Suites 3206-11, 32/F, One International
Finance Centre, 1 Harbour View Street, Central, Hong Kong, or Sanford C. Bernstein (business registration number 53193989L) , a unit of
AllianceBernstein (Singapore) Ltd. which is a licensed entity under the Securities and Futures Act and registered with Company Registration
No. 199703364C, Director of Compliance, 30 Cecil Street, #28-08 Prudential Tower, Singapore 049712. Additional disclosures and
information regarding Bernstein's business are available on our website www.bernsteinresearch.com.
CERTIFICATIONS
I/(we), Laura Nelson Carney, PhD, David Dai, CFA, Mark Li, Alberto Moel, ScD, Mark C. Newman, Michael W. Parker, Bhavtosh Vajpayee,
Senior Analyst(s)/Analyst(s), certify that all of the views expressed in this publication accurately reflect my/(our) personal views about any
and all of the subject securities or issuers and that no part of my/(our) compensation was, is, or will be, directly or indirectly, related to the
specific recommendations or views in this publication.
Approved By: NK
Copyright 2017, Sanford C. Bernstein & Co., LLC, Sanford C. Bernstein Limited, Sanford C. Bernstein (Hong Kong) Limited , and AllianceBernstein (Singapore) Ltd.,
subsidiaries of AllianceBernstein L.P. ~1345 Avenue of the Americas ~ NY, NY 10105 ~212/756-4400. All rights reserved.
This publication is not directed to, or intended for distribution to or use by, any person or entity who is a citizen or resident of, or located in any locality, state, country or other jurisdiction where such distribution, publication,
availability or use would be contrary to law or regulation or which would subject Bernstein or any of their subsidiaries or affiliates to any registration or licensing requirement within such jurisdiction. This publication is based upon
public sources we believe to be reliable, but no representation is made by us that the publication is accurate or complete. We do not undertake to advise you of any change in the reported information or in the opinions herein.
This publication was prepared and issued by Bernstein for distribution to eligible counterparties or professional clients. This publication is not an offer to buy or sell any security, and it does not constitute investment, legal or tax
advice. The investments referred to herein may not be suitable for you. Investors must make their own investment decisions in consultation with their professional advisors in light of their specific circumstances. The value of
investments may fluctuate, and investments that are denominated in foreign currencies may fluctuate in value as a result of exposure to exchange rate movements. Information about past performance of an investment is not
necessarily a guide to, indicator of, or assurance of, future performance.
BERNSTEIN GLOBAL SALES OFFICES
ZURICH
Talstrasse 83
CH-8001 Zurich
+41 (44) 227 7910
BERNSTEIN FUTURE OF TECHNOLOGY 2017: TESTING THE FENCES