5G: Hyperscale + AI + 5G = Sea Change
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Artur Pylak
FOR IMPORTANT DISCLOSURES 203.901.1633 /.1634
psagawa@ / firstname.lastname@example.org
May 29, 2017
5G: Hyperscale + AI + 5G = Sea Change
5G is the final key ingredient for the new TMT paradigm. Limited pre-standard deployments will begin next year, with a wide rollout after ratification in early 2020. The enhancements vs. 4G will be profound, with significant advantages for cloud-based AI. First, 5G will eliminate the capacity bottlenecks, spotty coverage, inconsistent speeds, nagging latency, and data plan limits that have constrained the ambitions and availability of current cloud-hosted AI applications. Second, 5G, with flexible service definitions, low costs and vigorous competition, will encourage a proliferation of access points (wearables, connected home devices, automobiles, etc.), seamless integration with WiFi and other networks, and an expansion of low priced unlimited data plans, allowing cloud-based AI much greater ubiquity, to the detriment of device-based alternatives. Finally, more connections and more usage means more data – IoT, B2B, transactions, etc. – to fuel the development of new AI-powered services. We believe that 5G will catalyze a shift to virtual assistants as a primary internet user interface, the rise of mass market augmented reality applications, and hastened adoption of autonomous transportation as a service, amongst other AI-supported use cases. The biggest beneficiaries will be the AI-powered cloud hosts (AMZN, MSFT, GOOGL and IBM) and consumer franchises (GOOGL, FB, and AMZN), with device-focused AI application strategies (AAPL) and internet companies without strong AI capabilities (PCLN, EXPE, SNAP, YELP, and others) will be disadvantaged.
- 5G is the final pillar for the next era of TMT. 5G deployments will begin in 2020, offering 10x improvements in wireless network speed, latency, availability, flexibility and cost, greatly improving user experience for mobile applications while enabling a wide range of new use cases, both mobile and fixed. We have written of the technologies, capabilities and use cases of 5G in detail (http://www.ssrllc.com/publication/a-5g-wireless-primer-the-final-ingredient-for-the-next-era-of-tmt/). We believe that 5G is the last ingredient to a generational transformation of information technology that began with emergence of low cost/high performance hyperscale data centers in the early ‘00’s and accelerated with the rise of deep learning AI over the past 5 years. 5G will give users uncompromised access to powerful AI platforms running in the cloud.
- 5G will enhance cloud AI performance. While cloud-hosted AI taps nearly unlimited AI-tuned processing and storage, 4G networks have been significant bottlenecks due to congestion, coverage holes, undependable speeds, and noticeable latency, damping user engagement and forcing app design compromises. 5G will mitigate these constraints, enabling developers to rely more completely on cloud-based AI solutions in both existing and emerging applications. We also expect 5G to spark even more vigorous carrier rivalry in the US, leaving data usage caps untenable and prices lower, reducing cost as an obstacle to broader AI platform engagement.
- 5G will extend the reach of cloud AI platforms. 5G’s ability to support low power, low speed (and low priced) connections will encourage new connected device types – e.g. IoT, wearables, etc. – many of which will be access points for cloud-based AI. Moreover, integration with WiFi and other unlicensed communication modalities will be inherent to 5G, further promoting a seamless AI platform experience across devices and venues (e.g. home, office, car, businesses, and the community) – a sizeable additional advantage for cloud-based AI solutions over applications that depend on the presence of a smartphone.
- 5G will generate new data to fuel cloud AI. The ubiquity and capacity of 5G will allow AI platforms to capture more data to drive inexorable improvement in existing applications and new data to enable fresh use cases. This data will come from sensors connected to IoT devices in homes, workplaces and communities, from microphones, cameras, touch screens, keyboards, pens, and other user input mechanisms, from autopilot systems in cars, and from field equipment deployed by industry.
- 5G plus AI will enable new paradigm applications. 5G performance, reach and data will greatly enhance the most anticipated AI use cases. We believe AI powered virtual assistants – taking input not just by spoken command, but from text, images, gestures, facial expressions, context and other sources – will become the primary way users interact with the cloud and their devices (http://www.ssrllc.com/publication/ai-assistants-the-next-user-interface-paradigm/). 5G brings ubiquity and eliminates lags that might hamper engagement. We expect fleets of autonomous cars to offer transportation as a service in many cities within 5-7 years (http://www.ssrllc.com/publication/autonomous-cars-self-driving-ambition/). 5G will enable vehicle-to-vehicle and vehicle-to-environment communications to enhance safety and efficiency. We believe augmented reality systems that interpolate digital images into a user’s perspective will require 5G’s bandwidth and latency to deliver appealing mass market applications (http://www.ssrllc.com/publication/a-5g-wireless-primer-the-final-ingredient-for-the-next-era-of-tmt/). Enterprises will use AI powered predictive analytics on mountains of new data made possible by 5G. We expect other applications, both consumer and enterprise, to emerge, catalyzed by the combination of AI and 5G.
- Cloud AI will beat local device AI. Today, AI applications running on a local device, like a smartphone or a car, have some clear advantages vs. cloud-hosted alternatives – e.g. responsiveness, availability, data plan usage, etc. These benefits will be largely mitigated by 5G, and cloud-based AI’s substantial advantages in processing power, cross-device reach and available data will tip the scales. In this scenario, devices (e.g. smartphones) may commoditize, while differentiation and user loyalty shift to cloud-based AI service platforms. Automobiles, where the premium for reliability is very high and where the heft and cost of the necessary processing and storage are less of a concern, are an obvious exception, but even there, cloud resources will buttress the mission critical systems on board.
- As usual, GOOGL, AMZN, MSFT and FB will benefit. Building powerful AI services will require data, talent and cloud infrastructure, factors that greatly favor the big consumer AI platforms – GOOGL, AMZN and FB chiefly, but also MSFT. In addition, these companies also have the scale to justify “mobile CDNs” that will push computing resources, including AI-tuned hardware, physically closer to 5G users to improve performance. Few other applications players will have the wherewithal to build their own AI-tuned infrastructure, a further boon to AMZN, MSFT, GOOGL, and IBM’s hyperscale cloud hosting operations. AAPL, whose vision remains very device centered, will suffer from its poorly developed cloud business as the smartphone market slowly commoditizes. We also believe 3rd party App companies (PCLN, EXPE, SNAP, YELP, etc.) that do not quickly adapt to a 5G-Cloud-AI future will face existential threats.
The Third Wave
The change has come in waves. First, hyperscale data centers changed the economics of computing and storage, removing performance and scale limitations in the process. This revolution spurred the rise of the Internet behemoths – e.g GOOGL, AMZN, and FB – and cued the rejuvenation of MSFT. Second, deep learning AI began to feed on the proliferation of data and the availability of cheap computing, and is changing the nature of software development and opening the door for remarkable new applications that had previously been unthinkable. The third wave will be 5G wireless (Exhibit 1).
Wireless innovation has been relentless and seemingly evolutionary – further enhancements to 4G will bring even faster speeds to users before the official 5G standard roll-out in 2020 – but we believe 5G is a bigger technical jump forward than many suppose. Connection speeds, latency, availability, network flexibility, and costs will improve by a magnitude or more, while vigorous carrier competition will push deployment schedules and pricing to the benefit of consumers. This will be a substantial boost to the performance and reach of cloud-based AI platforms, while providing a new flood of data to feed AI training.
Network performance and usage prices have been a substantial handicap for cloud-based AIs – connection is not always available when and where you want it, speeds may not be fast enough for graphics, lag is often unacceptable, and data plans have users rationing their usage. Device based processing and storage has been an attractive alternative for app developers, with AAPL as the biggest advocate. We expect 5G to greatly change this equation, spurring new consumer engagement with cloud-based AI applications.
5G will also allow cloud-based AI to reach users through new connections. AMZN’s successful Echo brought its Alexa AI virtual assistant into millions of homes, spurring GOOGL, and soon, AAPL and MSFT, to follow. These AI platform players are now racing to sign up partners who will put their AI assistants into home appliances, wearables, cars and other venues. 5G offers a low-cost connection option for these devices, along with provision for seamless integration with WiFi and other standards using unlicensed spectrum. It will also create enough volume around a single standard to drive chip costs dramatically lower – a needed stimulus for this new wave of devices. 5G will also generate mountains of new data to fuel the improvement of existing AI applications and to spur the creation of new ones.
So far, we see four categories of AI “killer apps” and all will benefit greatly from 5G. Virtual assistants will be more powerful, more responsive, and increasingly ubiquitous. Augmented reality will move from simple image and image processing features to the real-time integration of context appropriate digital content into our view of the natural world. Autonomous vehicles will add the ability to coordinate with their environment and with each other. Enterprise predictive analytics applications will have new end points and mountains of new data with which to work.
All of this is good for the companies at the top of cloud AI heap – GOOGL, MSFT, IBM, FB and AMZN. Their services will get much better, and their hosting platforms will attract ever more 3rd party customers looking for AI support. It is not good for the companies betting on device-based AI (AAPL) and on the app model in general (PCLN, SNAP, EXPE, YELP, etc.).
Exh 1: The Three Pillars of the New TMT Paradigm
In the Beginning, Google Created the Hyperscale Data Center …
Hyperscale -The internet was a big place, even in 1996, when Larry and Sergei began to consider the best way to build a tool to help users navigate the World Wide Web. Over the next several years, and the birth of a company, the two founders pressed forward on a radical new software architecture that eliminated the barriers to scale inherent in the client-server data centers and structured data bases that then dominated the IT world. Eventually, Google contributed the central ideas to the open source community, birthing a raft of new wave IT standards and fueling the growth of today’s internet behemoths (Exhibit 2). Along the way, Google also reinvented data center hardware, designing its own custom blade servers, racks, and facilities to ratchet up performance and drive out costs, ideas that have also since been mimicked by Amazon, Facebook, Microsoft, and a few others that dare to play in the big leagues of cloud-based computing (Exhibit 3). The best hyperscale operators have all in computing costs a magnitude lower than the enterprise average, with the ability to flexibly shift resources to accommodate workloads of almost limitless size (Exhibit 4). This has already proven a game changer with the rise of the internet giants, riding scale economies and network effects to dominance. According to Sandvine, Google, Facebook and Amazon are the source for more than 55% of North American consumer internet traffic via fixed lines and 49% from mobile networks – only accounting for the traffic from their hosted properties in the top ten (Exhibit 5-6). These percentages are up from 41% overall 5 years ago (Sandvine didn’t break out wireless traffic in 2011).
Exh 2: Timeline of Key Milestones in Hyperscale Computing, AI, and 5G
Exh 3: GOOGL’s technologies work in concert to power the world’s most powerful distributed computing platform
Exh 4: On-Premise versus Cloud Hosted Costs
Exh 5: Downstream Fixed Access Peak Web Traffic, 2013-2016
Exh 6: Downstream Mobile Peak Web Traffic, 2013-2016
AI – Google was also the key player in the next revolutionary technology. Having addressed the challenge of how to build an index of the entire internet on its network of hyperscale datacenters, Google turned to the other side of the equation – could it do a better job of interpreting the intent of its users? For decades, scientists had conceived of recursive, learning algorithms that could adjust their own code to return ever more appropriate answers – an AI technique called neural networking or machine learning. However, to accomplish anything complicated, these software techniques required enormous amounts of data and massive computing resources, two things that Google had in spades. Google began aggressively recruiting computer scientists with AI expertise. In 2011, Google launched the Google Brain initiative, led by Google Fellows Jeff Dean and Andrew Ng. The team pushed the state AI forward in areas like image recognition and language processing, and the Brain AI technology became the core of the next iteration of the core search platform. In 2013, Google hired Dr. Geoffrey Hinton of the University of Toronto, widely considered the most important contributor to the development of deep learning, and in 2014, it acquired Deepmind, a Cambridge England based think tank that, according to Yoshua Bengio (another of the founding fathers of AI and the brother of Google Fellow Samy Bengio, part of the Google Brain team), was led by a dozen of the top 50 AI scientists in the world (Exhibit 7).
Exh 7: AI Scientist and Citations by Company
Rivals followed Google into the breach. Microsoft ranks as the number two company in terms of its AI capability, followed by IBM, Facebook, and Amazon. The Chinese internet giants, whose AI scientists are much more difficult to track, are also formidable – particularly Baidu. Today, nearly every cloud service provided by these companies is infused with AI programming, in many cases already giving them extraordinary advantage over would-be competitors without the same machine learning firepower. With enough data and processing power, cloud-based AI has the potential to make almost any software better and to enable new application areas – e.g. virtual assistants, autonomous controls, augmented reality, predictive analytics, etc. – that were previously impossible.
5G is the Road That Leads to the Cloud AI Door
Existing wireless networks are a bottleneck for cloud-based AI applications. Speeds are not always sufficient for image and video data. Latency adds noticeable lag between action and response. Availability is inconsistent. Perhaps most importantly, data caps and high prices deter users from engaging as often or for as long as they might like. This has been enough to prompt developers to implement AI functions, such as voice recognition, image processing, or biometric identification, to run directly on mobile devices rather than in the cloud, taking expedience in exchange for performance (Exhibits 8-9).
Exh 8: Average Performance of US Mobile Networks, February 2017
Exh 9: Throughput Requirements by Application and Select Compression Technologies
5G will change the equation. The standard is on a path to ratification in early 2020, with progress evident in the many working groups within the 3GPP organization that will approve the specifications. The biggest contributors of intellectual property to the standard are the chip makers (mostly Qualcomm) and equipment manufacturers (Ericsson, Nokia, and Huawei), all of whom are hard at work designing pre-standard gear that will smoothly upgrade to standards compliance upon ratification (Exhibit 10). Major carriers are working with suppliers to qualify gear and set their plans for deployment – some in pre-standard mode beginning next year, but most in 2020 once the specs are solidly set. We see the advantages of 5G on five key vectors, speed, latency, availability, flexibility and cost (http://www.ssrllc.com/publication/a-5g-wireless-primer-the-final-ingredient-for-the-next-era-of-tmt/). Each removes a major obstacle to cloud-based AI engagement (Exhibit 11).
Exh 10: 5G Network Technology Patent Assignments
Exh 11: ITU 5G Criteria, Applications, and Progress
5G aims to enable up to 10Gbps connections with average throughput of at least 1Gps (Exhibit 12). That is roughly 20 times faster than today’s LTE networks, and a significant boon to image and video heavy applications, such as augmented reality. The latency is expected to fall to just 1ms, a 98% drop from current average LTE latency of about 50ms. This will reduce the apparent lag for all cloud-based applications – responses from virtual assistants will be snappier, augmented reality programs will be able to map to the real world in nearly real time, and self-driving vehicles will talk to each other to avoid accidents. Cell sites will have dramatically higher capacity and will even shift to meet surging traffic, greatly reducing congestion and its impact on the user experience. At the same time, advances in antenna technology and the emphasis on low cost small cells will allow carriers to more economically fill in empty spots in their coverage. Together, these improvements will give users more confidence that cloud-based applications will be available whenever and wherever needed. Finally, 5G will increase carrier flexibility to use new spectrum bands and to offer a more varied palette of services, while simultaneously driving both capex and opex per unit of network capacity substantially lower. We believe this will spur even more vigorous competition amongst the US carriers, perhaps even drawing a new participant or two to the market. T-Mobile’s provocations have already made unlimited data plans standard and pushed down wireless ARPUs, and we expect more of the same with 5G. Users will not be concerned that using cloud-based AI platforms will drive up their monthly bill.
Exh 12: 8 Key Capabilities of IMT-2020 (5G)
5G – Finally the Internet of Things Can Happen
For nearly a decade, the tech industry has been touting the “Internet of Things” under various terms and guises. Endless connected home demos at CES have shown talking appliances, thermostats, lights, car door openers and what not easily managed from a smartphone. Unfortunately infighting on just what wireless standard to use to coordinate the whole thing has been a major obstacle, yielding non-interoperability and high prices. 5G, designed to seamlessly integrate with WiFi and other common standards using unlicensed spectrum in the home and workplace, could be the ticket to broad industry agreement and scale component production. 5G is also designed to enable low power, low throughput (LPLT), low price services to coexist with typical mobile broadband. This will be an important element toward connecting innumerable machines outside the range of home or office WiFi. These devices may be conduits for cloud-based AI – already Amazon and Google have been tripping over themselves to sign on appliance makers, car brands, and others to support their AI assistants.
Exh 13: Wearable Shipment Forecast
5G’s flexibility and cost may also stimulate the rise of wearables – those watches, glasses, pieces of clothing, shoes, etc. designed with a reason to be connected (Exhibit 13). Today, these items presume a connection via a smartphone, but in the future, they may connect independently. Cloud based AI may prove a better way for users to interact with these items, and better marry them to a broad set of integrated services across many devices. This ubiquity – maintaining a common interface and experience as the user moves about the home, in the car, at work, and in the community – is a profound advantage for cloud-based solutions vs. device based ones. A smartphone doesn’t know what you do on your computer, your TV, your home speaker, or your office system, but a cloud-based platform can. For an AI assistant, that continuity means for more personalized services and answers.
5G – More Data, Better AI
More user engagement, and more devices in play means more data to fuel AI development. Existing applications will get better, and better, as the data provides more nuance to the AI’s understanding of its circumstances. New data will beget new use cases. Here we are again intrigued by the Internet of Things. Sensors in roads, in the municipal infrastructure, in vehicles, in industrial equipment, and many other places, all connected via LPLT and providing grist for development. The municipal and industrial use cases may take longer to evolve, but their potential to generate social and commercial value are extraordinary.
The Primary AI Use Cases in a 5G World
Thus far, four primary future use cases for AI have become apparent, and while others will undoubtedly emerge, these offer huge opportunities. All of them will be enhanced by 5G.
Virtual Assistants – Voice-based interfaces are the current rage, but we expect future AI assistants to meld a variety of inputs – images, sounds, text, gestures, eye movement, facial expressions, movement, context, and others, along with spoken words – to intuit the meaning of requests and anticipate user needs (Exhibit 14). We believe these assistants will begin to displace the app model as the primary way that users interface with the internet, with applications, and with their various devices. 5G will hasten this transition by increasing the richness of user interactions, by increasing the venues in which users may access their assistants and by offering new training data to improve their accuracy and the range of activities that they may control. (http://www.ssrllc.com/publication/ai-assistants-the-next-user-interface-paradigm/)
Exh 14: Voice Assistant Improvement Wishlist
Augmented Reality – We see apps like Snapchat and Instagram, interpolating digital images into user generated photographs and videos, as early stage instances of augmented reality. Now the digital content is extremely simple – rabbit ears, rainbows and the like – but future content will be far more complex, relating information specifically to the context of the user environment and juxtaposing it artfully into the perspective. Future augmented reality will also be real time, always on, and projected into the user’s field of view. It will fuel entertainment, information and work related use cases, and it will rely on high bandwidth 5G mobile networks with minimal latency and broad availability to work (Exhibit 15). We believe that the full promise of mass market augmented reality is likely a decade away (display hardware for real time, full field of view AR is far from ready), but we expect high value commercial applications to lead the way. (http://www.ssrllc.com/publication/vr-and-ar-the-reality-story/)
Exh 15: Current latency and router hops of major web services
Autonomous Machines – Self driving cars are making dramatic progress, and we expect fully autonomous robo-taxi services to begin testing in commercial markets before decade end (Exhibit 16). (http://www.ssrllc.com/publication/autonomous-cars-self-driving-ambition/) The initial driving systems will not depend on external communications – the risks of losing connection at a crucial moment are too high – but supplementary network-delivered capabilities will gain traction with 5G. Vehicle to vehicle communication would enhance the collision avoidance capabilities of the autonomous systems and enable “platooning”, when self -driving vehicles organize into closely spaced packs to increase efficiency. It would also allow for municipalities to communicate directly to the cars for traffic control, safety and law enforcement. There will also be other forms of autonomous machines – drones, delivery robots, maintenance robots, etc. – that will benefit from the high performance, low cost aspects of 5G.
Exh 16: Autonomous Car Release Target Dates
Predictive Analytics – 5G will enable data collection for many purposes – road conditions, remote metering, security cameras, system diagnostics, etc.. This data will be analyzed and used to improve the performance of commercial and government operations. These applications are more industry specific, and perhaps less glamorous than augmented reality glasses or self-driving cars, but the potential opportunity is huge (Exhibit 17). (http://www.ssrllc.com/publication/ai-sizing-the-market-big-bigger-and-biggest/)
Exh 17: Sizing the AI Opportunity – Three Tiers of AI Innovation
Cloud AI Will Trump Device AI
Apple is making some headlines with rumored plans to develop a custom GPU to handle deep-learning powered apps directly on its iOS devices, tying to its company commitment to protect user privacy by keeping as much data local to the device as possible. While this has merit in the current market environment with network contributed performance lag and spotty availability, we believe that the relative benefit for Apple customers will be short lived (Exhibit 18).
We expect the arrival of 5G over the next 5 years to tip the scales squarely toward cloud based solutions, removing performance barriers, broadening the reach and generating valuable data. 3rd party developers may be wary of focusing on device-based AI solutions that will not be portable to non-iOS devices or even available via other access methods in other venues. Coordinating between different iOS devices may even be proscribed. As better, more flexible, more available, AI applications appear via the cloud and begin to crowd out the traditional app interface model, the demand for ever more capable (and expensive) devices may peak and wane, beginning to commodify, as cheaper products with 5G access can offer the same leading edge cloud AI delivered performance as the most expensive models. Adding a more burly GPU, ASIC and/or PLD to a smartphone could add $40 or more to the bill of materials of a smartphone, while taking space and drawing power, all for functionality better delivered over the network.
Exh 18: Cloud-Based versus Device-Based AI Solutions
5G may also allow wearable devices – watches, glasses, etc. – to untether from smartphones, perhaps raising their attractiveness to the detriment of their former masters. We do not expect the same phenomenon for automobiles, where the contribution of onboard AI to cost, form factor and battery life are far less a concern, particularly in the context of self-driving systems.
Winners and Losers
We believe that a small number of companies have built the critical mass of scientific talent, proprietary data sets and hyperscale infrastructure sufficient to be broad leaders in the big categories of AI. Those companies are Google, Microsoft, IBM, Facebook, Amazon and Baidu (http://www.ssrllc.com/publication/37440/), and they have already begun to demonstrate dominance in the primary application areas – i.e. virtual assistants, autonomous controls, augmented reality and predictive analytics – while also providing AI hosting platforms for 3rd parties (Exhibit 18). 5G will solidify their already strong positions. Other players have enough AI capabilities to use the technology to successfully deliver focused solutions, typically tied to their core businesses. Companies like Salesforce, Adobe, Uber, Twitter, and a few others fall into this category.
While we are bearish on the prospects for ever more powerful smartphones, many chip makers will still benefit greatly from the 5G boost to cloud-based AI. We see Nvidia and Xilinx as exceptionally well positioned for the rapidly expanding hyperscale datacenter AI processor buildout and for automated vehicles, which must depend on reliable onboard computing for their AIs. Qualcomm, which will clearly benefit from modem and radio components for 5G capable devices, may also play a role in AI processing.
While Apple has begun to build a more impressive AI team through hiring and acquisitions, we fear that its steadfast focus on device-based rather than cloud-based will prove ill advised. We do not believe that consumers care as much about data privacy as does Apple management, and expect constrained functionality relative to cloud-delivered alternatives will be a competitive liability. We are also concerned for narrowly focused mobile app companies facing likely competition from the powerful cloud AI platforms. For example, Snap’s image processing forward messaging platform sits squarely in Facebook’s augmented reality ambitions, and we do not believe that Snap has the AI R&D firepower needed to stay ahead. Other companies with similar problems include Priceline, Expedia, Yelp, and others.
Exh 19: The SSR TMT AI Heatmap
©2017, SSR LLC, 225 High Ridge Road, Stamford, CT 06905. All rights reserved. The information contained in this report has been obtained from sources believed to be reliable, and its accuracy and completeness is not guaranteed. No representation or warranty, express or implied, is made as to the fairness, accuracy, completeness or correctness of the information and opinions contained herein. The views and other information provided are subject to change without notice. This report is issued without regard to the specific investment objectives, financial situation or particular needs of any specific recipient and is not construed as a solicitation or an offer to buy or sell any securities or related financial instruments. Past performance is not necessarily a guide to future results.