The Future of Semis: After Moore’s Law
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Tejas Raut Dessai
psagawa@ / firstname.lastname@example.org
October 8, 2020
The Future of Semis: After Moore’s Law
As tech continues to permeate into more of our lives and work, most forecasts project semiconductor sales to outpace economic growth. Of course, these forecasts are complicated by the global recession and the notorious industry cyclicality – we are more cautious about the expected upcycle than most YTD. Underlying the state of demand and the inherent cyclicality are several powerful trends that are shifting the global competitive leaderboard: 1. Moore’s Law is reaching its limits – increasing manufacturing scale advantages and weighing on integrated suppliers. 2. The Cloud will the biggest growth driver, not smartphones – low power ARM servers, AI accelerators (e.g. GPUs, FPGAs, and ASICs), high performance networking, and cheap memory/solid-state storage will benefit. 3. New markets (e.g. IoT, AR, autonomous vehicles) are emerging slowly – near term sales could disappoint but leadership established now could be valuable a few years down the line. 4. China is investing heavily to get in the game. In this context, we are seeing a competitive shift, with NVDA ascendant, QCOM and AVGO well positioned for post-recession, AMD with a hot hand, and INTC struggling. Integrated commodity chip makers (Samsung, Hynix, MU, etc.) and contract foundries (TMC, Samsung, Global Foundries, UMC, etc.) will compete on scale and efficiency as process advantages get harder to come by and Chinese fabs try to muscle in. Meanwhile, analog chip makers, without meaningful scale economies, will see relative business as usual.
- Global chip demand will outpace the economy across the cycle. Semiconductor industry revenues have grown at a 7.25% CAGR over the past 32 years, outpacing the global economy by 220 bp. While the growth has slowed as the industry has grown larger, we, and most observers, believe that there is considerable runway left for electronics to spread into other aspects and corners of the global economy, and thus, sustain better than market average growth across the cycle. Of course, that cyclicality, born of the massive capital investment needed to expand industry capacity, will continue, although concentration could dampen its impact.
- Moore’s Law is near its limits. Intel founder Gordon Moore posited that the number of transistors that could be etched on a chip would double every 18 months, a phenomenon behind the rise of electronics in so many aspects of life. However, the miniaturization that drove the law has hit unyielding scientific obstacles. With the viability of radical new approaches, such as quantum computing, many years away, competitive advantage will shift from process leadership toward efficient hardware and software design. This will also increase the advantages of scale in manufacturing, favoring the fabless/contract fab model over integrated players.
- The action is shifting to the cloud. Datacenter computing drove the first era for semiconductors, PCs the second, and smartphones since the 2007 introduction of the iPhone. The next era for chipmakers will be defined by cloud computing, with RISC processors (i.e. ARM and RISC-V), AI accelerators (i.e. GPUs, FPGAs, and ASICs), high performance networking (i.e. interfaces, switches, optronics, etc.), solid state storage, and memory the key product categories. Cloud computing will eat into demand for enterprise datacenters, smartphones (and other consumer devices), and even telecommunications, where 5G architectures will shift functionality from the cell site into the cloud. Automotive and industrial are now pressed by the global recession but should eventually contribute to growth.
- The “Next Big Things” are still years away. New applications, such as Internet of Things (IoT), Augmented Reality (AR), and autonomous driving systems are in the news, but volume markets for chips to enable them are not likely until the back half of the decade. Radical new technology – such as quantum computing, spintronics, holographic and DNA storage, etc. – is even further from commercial reality. Nonetheless, the investment to establish leadership is happening now and could generate significant long-term value.
- The Chinese wildcard. Chinese companies, supported by aggressive government policy and partly in response to the rising trade tensions with the west, are investing heavily in semiconductor manufacturing. To date, Chinese fabs have been generations behind the global leaders, a deficit that precludes making competitive memory chips, but the end of Moore’s Law will help them close the gap. China is also investing in advanced process equipment with the hope of ending reliance on US suppliers, but commercial parity is at best years away.
- Analog is what it is. Analog semiconductors are used to bridge digital systems with the physical world in applications like power supplies, signal processing, and sensors. Unlike digital semiconductors, analog designs are not easily adaptable to ever smaller geometries – allowing manufacturers to use older fabs – and are notoriously finicky in the transition from design to production. Analog design engineering talent is also in perpetual short supply. In this context, the analog market is less cyclical with generally higher, more stable pricing. As such, systems designers seek out digital alternatives, but in many cases, these are not now possible. As the automotive industry recovers from recession, as millimeter wave 5G deployments become more prevalent, and as IoT rises to become a volume market, analog chip suppliers will benefit, with growth likely somewhat above the broader industry.
- Industry leadership is shifting. Long-time king of the hill INTC would be vulnerable, even if it hadn’t fallen behind on its fab process upgrade timetable. NVDA, leading in cloud AI processors, fresh from acquiring hot datacenter networking player MLNX, and awaiting approval for its deal for RISC designer ARM, is in a powerful position. QCOM and AVGO each have heavy exposure to smartphones, but strong opportunities in the cloud and for some of the “next big things”. AMD is enjoying INTCs pain but needs for its investments in RISC and GPU to pay off. Integrated memory makers (Samsung, Hynix, MU, etc.) and contract fabs (TSM, Samsung, UMC, etc.) will see scale get even more important as process advantages become harder to come by. Analog will be what it is.
- How to play it. NVDA is a core investment for the next decade – even if the ARM deal is blocked. We would wait for QCOM and AVGO, expecting another shoe to drop on smartphone demand. In contrast, we expect AMD to make hay while the sun is shining – own it now but keep it on a short leash. XLNX will work once 5G deployment accelerates post pandemic. Memory makers have been choppy with investors betting on a cyclical upturn that has been thwarted a bit by the global pandemic. We’d wait a bit more. In contrast, contract fabs have run hard and we’d take profits for now. Analog semis have been dogged by their dependence on the troubled auto industry. We’d wait here as well.
All That and a Bag of Chips
Since 1988, the semiconductor market has averaged 7.25% growth, a full 220bp ahead of global GDP, across wrenching cyclicality that has driven wild swings in annual industry sales. 2020 was supposed to start a hotly anticipated upcycle after 4 years of declining demand for smartphones had left manufacturers with excess capacity. Unfortunately for investors, the pandemic stepped on the inflection point, giving the industry another down year and a lot of uncertainty as to when the economic conditions would fuel an inevitable recovery. We are inclined to be a bit more pessimistic than consensus on this timing.
Against this backdrop, the sector faces a paradigm shift driven by three main forces. First, Moore’s Law, which posits that the density of transistors implemented on a chip doubles every 18 months, has bumped into hard physical limits. Process improvements can still drive a bit more performance, but the pace has slowed and will continue to decelerate. This will reward efficient design over process leadership, increase the benefit of scale in manufacturing, and favor the fabless model over integration. Second, smartphones grew from 0 to 25% of global chip revenues over the past 14 years, but sales have plateaued. Going forward, we see hyperscale datacenters and cloud infrastructure as the primary driver of the market. RISC processors (i.e. ARM and RISC-V), AI accelerators (i.e. GPUs, FPGAs, and ASICs), networking components (i.e. interfaces, switches, optronics, etc.), solid state storage, and memory will be the hot products. 5G, slowed a bit by the pandemic, will kick in by 2022, and automotive/industrial will return to growth once the recession is behind us. We are not enthusiastic for a significant reacceleration of smartphone unit demand.
Third, there are a number of well hyped tech buzzwords, like the Internet of Things (IoT), Alternative Reality (AR), and self-driving cars, beckoning from across Gartner’s “Trough of Disillusionment”. Eventually, these will be important markets, but the opportunity for volume chip shipments are years away. Nonetheless, investment is already happening, and participation now is likely a prerequisite for leadership then. We also note that game changing technologies, like quantum computing, spintronics, DNA storage, holographic storage, and others, are even further out. Finally, China Inc. is investing heavily to get in the game, just as aggressive US trade policy is erecting barriers to keep it out. As the “Moore’s Law” race slows down, access to the latest fab tech may become less important but gaining competitive scale, usually a point in China’s favor but not in this case, will be a tough hurdle coming from behind.
Meanwhile, analog semis are a different animal. Design is paramount, new processes don’t help much, and scale is tough to achieve. Analog is far less cyclical, prices are more stable, and leaders tend to sustain better profits. Automotive and industrial applications are key end markets, so the recession is a damper, but we expect better than average growth from the segment on the other side.
To sum up, we like fabless chip designers with a focus on the cloud, which means NVDA and to a lesser extent, AMD. QCOM and AVGO are also likely long-term winners, albeit with a drag from their current reliance on smartphones. Like the analog players, XLNX will slowed by its exposure to automotive, but once 5G spending picks up, we think it’s a buy. As for contract fabs, TSM is the biggest and best but we’d wait to feel more confident the end of the recession is in sight. The same is true for memory makers Samsung, Hynix, and MU. Finally, INTC might be worth a trade, but the long-run picture is pretty ugly. We’d pass.
60 Years of Integrated Circuits
On September 27, 1960, Fairchild Semiconductor created the world’s first semiconductor integrated circuit (IC) and was promptly sued by Texas Instrument, which held patents for an earlier IC design that had not proved commercially viable. The patent war was eventually settled in a 1966 agreement with wide cross licensing between the companies. Thus, the birth of modern electronics and the beginning of decades of intellectual property litigation.
Over the subsequent six decades, semiconductor integrated circuits – now shortened to semiconductors or just semis – have enabled modern technology in waves of innovation. In the early days, the cutting edge was in mainframe computers and electronic communications, elevating IBM and AT&T to the head of the “Nifty Fifty”. The early ‘80’s brought us PCs and the first cell phones, made possible by advances in semiconductor technology that let these devices be small enough and cheap enough to be affordable. The adage was that “Intel gives, and Microsoft taketh away” meaning that advances in the power of chips was soaked up by increasingly complex software. This was the manifestation of “Moore’s Law” named after Intel founder Gordon Moore who noted in 1965 that the density of transistors that could be implemented on a semiconductor chip was doubling about every 18 months, and that the pace of improvement could be expected to continue for a long time.
By the new millennium, Moore’s law had progressed to the point that the computing power of a PC could be implemented in something much smaller and powered by a portable battery. This dramatically altered the
Exh 1: Global Semiconductor Revenues over the past 30 years, 1989 – 2019
Exh 2: Global Semiconductor Sales Leaders and Operating Structures, 2019
Exh 3: Integrated Manufacturing Semi Revenues over last 10 years, 2010 – 2019
Exh 4: Contract Manufacturing Semi Revenues over last 10 years, 2010 – 2019
semiconductor landscape. The ARM RISC processor architecture was the hot ticket. Mobile specialist, ARM licensee, and fabless pioneer Qualcomm rose into the top echelon of chip suppliers. Apple took its ARM-based core chip design in-house. RF specialists Qorvo, Skyworks, and, especially, Broadcom rode the wave. Contract fabs Taiwan Semiconductor Manufacturing and Samsung soaked up the smartphone volume to fuel a fierce process improvement race. By 2019, smartphones alone accounted for 25% of all semiconductor revenues, with PCs far back in the rearview mirror.
That brings us to now (Exhibit 1, 2, 3, 4). Contract fab giant Taiwan Semiconductor is the number one pureplay chip company by market cap. (Samsung is a touch bigger overall but its diversified businesses make it hard to slot vs. semiconductor specialists) The success of gaming, cryptocurrency, and artificial intelligence (AI) has driven a boom in graphics processors, fueling Nvidia’s rapid rise to second place on the silicon market cap list. Intel, the erstwhile long-time leader, has fallen to third as it struggles to keep up in the manufacturing process race while its x86 CPU architecture faces new challengers (Exhibit 5). Number four is Broadcom, propelled by CEO Hock Tan’s aggressive M&A strategy and its relative focus on wireless and datacenter markets. Wireless leader Qualcomm rounds out the top five, with a big recent run on investor speculation for a 5G driven smartphone upcycle in 2021. Red hot AMD is the best of the rest, pushing a $100B cap as it harvests market share with its 7nm x86 processors.
Exh 5: Highlight of Top Semiconductor companies with greater than $100B M Cap
What Comes Next?
The semiconductor industry faces four major forces for change. 1. The end of Moore’s Law; 2. The preeminence of the cloud; 3. Anticipation of “The Next Big Things”, and 4. The rise of China. These forces have already had significant effect on the industry, but their impact will only grow with time, specifically for the biggest revenue digital product categories. We note that change for the smaller but generally more profitable analog product categories will be less wrenching for the competitors that specialize in them.
The End of Moore’s Law
About a decade ago, scientists began to sound a warning that soon, the laws of physics would no longer allow engineers to reduce the width of the microscopic lines of metal needed to conduct electricity along the surface of an integrated circuit. As the lines got smaller, the potential for the signal to bleed from one line to
Exh 6: Global semiconductor industry faces four major forces of change
another would rise, eventually rendering the chips unusable. The industry found some ingenious methods to eke out further improvement – e.g. insulation laid between the lines, laying lines higher and lower to the surface for a 3D effect, and other techniques – but the pace of improvement has clearly slowed, and with the lines on the most advanced current process a mere 14 atoms wide, the end is clearly in sight (Exhibit 6, 7, 8).
This has major implications for the semiconductor market. First, without a steady pace of process improvement as the impetus for the price/performance trajectory of ICs, the onus will be on efficient hardware and software design to carry the day. Expect specialized chips to gain on general purpose solutions, for reduced instruction set (RISC) processors (e.g. ARM or RISC-V) to gain on complex instruction set (CISC) processors (e.g. x86), and for scale efficiencies to become the most important competitive factor for chip manufacturing. This also reduces the process expertise barrier against would be entrants to the IC fabrication market, of note given the obvious interest of the Chinese (Exhibit 9).
Exh 7: No. of Transistors on an Integrated Chip for leading Processors, 1971 – 2018
Second, the integrated chip making model, whereby competitors design their own solutions and manufacture them in their own fabrication facilities, will lose traction against fabless rivals, who will be free to choose the cheapest contract fab partner. This sort of dynamic is already playing out for Intel, whose repeated delays in bringing its latest process improvements to production have boosted bitter rival AMD, which moved to the fabless model with the 2009 spinoff of GlobalFoundries. The last stronghold of the
Exh 8: Number of transistors which fit into a microprocessor – in Log Scale
Exh 9: Snapshot of Moore’s Law and its limitations
integrated model is memory, where design is a commodity and chip density is paramount to price/performance. We expect scale advantages to grow and cyclicality to ease as a result.
The Preeminence of the Cloud
For more than a decade, semiconductor market growth has been driven by the global adoption of smartphones. In the first 10 years after the 2007 introduction of the iPhone, unit sales in the category went from essentially zero to more than 1.5B, accounting for nearly 30% of all 2016 semiconductor industry revenues. Unfortunately, the law of big numbers reared its head and with an installed base of roughly 4 billion users, smartphone demand hit a plateau, with global volumes stagnant over the last 4 years. Meanwhile, data processing – a grab bag category that includes chips for PCs, private enterprise datacenters, and cloud-based hyperscale datacenters – has returned to growth after years of decline, largely on massive capex by the major cloud operators. We believe that this will be the biggest source of growth for the semiconductor industry over the next decade.
Capital spending on cloud datacenters jumped 46% in 2018, as the major operators (Amazon, Microsoft, and Google lead in commercial hosting, but platforms for massive consumer franchises, such as Facebook, are big spenders too) stepped up to address demand. The big 2018 splurge yielded a 1H2019 hangover as the spending dipped, but contrary to expectations, the spree started up again in the back half of last year and accelerated with the pandemic (Exhibit 10, 11, 12).
While some analysts continue to pound the table on a big 5G driven smartphone super-cycle for 2021, we fear that global recession and scarce 5G service availability will dampen enthusiasm for $1,000+ flagship devices (Smartphones: They Picked the Wrong Year to Call a Supercycle). Rather, we see continued cloud infrastructure spending – supporting work-from-home,
Exh 10: Global Hyperscale Cloud Capex Spending and Growth – last 5 years
Exh 11: Snapshot of Global Smartphone Unit Shipments – Growth has Plateaued
Exh 12: Snapshot of Key Consumer Technology Adoption in the US
streaming media, online gaming, e-commerce, and other internet-based activities that have been boosted by the pandemic – as the primary driver for the semiconductor market. Indeed, as 5G finally rolls out, we believe that it will enhance demand for those cloud applications as it should dramatically reduce the lag experienced when interacting with them (Exhibit 13).
A cloud-driven IC market will be quite different from the smartphone driven market of the past (Exhibit 14, 15, 16). (Hyperscale Semiconductor: Processor Diversity Coming to the Cloud) Traditional x86 CISC CPUs will see pressure in the datacenter from low power RISC CPUs based on ARM or the open source RISC-V architectures. This will be particularly important for content delivery networks that push application performance via hyperlocal mini datacenters. As AI applications become an ever-greater of the mix, we see specialized processors – AI tuned GPUs, flexible FPGAs, and use case specific ASICs – capturing disproportionate share of the growth. This will also drive demand for memory and solid-state storage, with the hyperscale operators soaking up DRAM and CDNs leading the transition to SSD. High speed networking – interfaces, switches, optronics, etc. – will also be a growth category in a cloud driven market.
In contrast, we believe the performance of the cloud will lessen the impetus to upgrade older smartphones, as more innovation is served from the network rather than implemented on the device. Watches and other peripheral devices will shore up demand a bit, but we expect chips for this end market to cycle around a rough stasis. Automotive, another consumer driven category, is down fairly hard in the pandemic but is bound to recover as the post crisis market emphasizes electric vehicles and ADAS systems, both of which materially raise the semiconductor content per vehicle. The industrial market has also been hit by the COVID-19 recession but should also be a net contributor to semiconductor market growth in recovery.
Exh 13: Modern Computing Stack is Evolving to Support Data-hungry Applications
Exh 14: Centralized hyperscale processing is displacing legacy architecture
Exh 15: Highlight of Emerging Standards in the Cloud-Driven Computing Stack
Exh 16: Brief summary of RISC-V Architecture and Notable Implementations
Anticipating the “Next Big Things”
Futurists are always jumping the gun. For example, the “Internet of Things” has been a buzzword hype-fest for years as pundits described a world of devices in every possible venue – homes, factories, streets, stores, schools, and so on – all coordinated via low power, low cost wireless connections, collecting data, acting on user commands, fulfilling routine tasks, blah, blah, blah. The smart home and smart factory have been just around the corner for a long time, ever since someone put a modem in a soda machine in Finland so that you could pay for a Coke via text message.
There are signs that some of this may be finally getting some traction – we’ve written about the “Smart Home” (Hold the Metaverse: The Cloud Assistant Era is Coming) – but from a chip supplier perspective, we believe that it will be a few years before the volumes of low priced sensors, microcontrollers, and wireless networking chips for IoT are a meaningful piece of overall IC demand (Exhibit 17, 18). The same is true for augmented reality (AR), a “next big thing” touted by both Apple and Facebook, but not yet close to the magic combination of compelling technology, mass market pricing, and killer applications to make it more than a modest niche. We’ve written of this too (The Long Road to AR Glasses), and note that the optical hardware needed to interpolate high resolution images into the context of a user’s field of view remain far too bulky, power hungry, and expensive to allow a mass market product – even IF the social reaction to a “smart glasses” product is less negative than we fear.
While we remain bullish on the prospect of on demand robo-taxis deployed to specific, geo-fenced markets as an alternative to ride hailing services, the reality is that the volumes of sensors, AI-focused processors, memory, and storage for the 10’s of thousands of self-driving vehicles needed to launch the service in a
Exh 17: Summary of Catalysts Driving Innovation in Home Networking
Exh 18: Global Smart Home Device Installed Base Forecast by Type, 2019 – 2023E
dozen or more markets will not move the needle for chipmakers looking for multi-million unit annual opportunities. Eventually, we will get there, but probably not in this decade (Exhibit 19, 20, 21).
Still, investment in these and other future use cases continues apace, and while the commercial payoff is far away, the potential is huge. We believe that leadership established in the near-term will be likely to pay off with substantial growth in the long-term. Qualcomm, Nvidia, Apple, and Alphabet have made substantial commitments to these opportunities, and could have future advantage in components, or in Apple’s case, integrated devices, that address them.
The future is even further away for a range of radical alternatives to existing transistor-based paradigms. Traditional semiconductors rely on the binary nature of electrical charge as positive or negative, while quantum computing exploits a distribution of possible states in subatomic particles and the interconnectedness of particles that have been “entangled” to perform multiple operations simultaneously. This technology has huge potential to improve processing performance for complicated analytics and for modeling real world uncertainty with obvious relevance for AI, cryptography, and other valuable applications. Arguably, IBM and Alphabet are at the forefront of the research on quantum computing, but the road ahead is awfully long, and a commercial quantum product is unlikely in the next ten years.
Spintronics, a portmanteau of spin and electronics, proposes using the directional “spin” of electrons to add a further dimension to the ordinary positive/negative binary that underlies traditional integrated circuits. The idea is believed to offer particular advantage to memory, storage, and semiconductor laser products, but
Exh 19: Average Semiconductor Component Revenue Per Vehicle, 2005 – 2025E
Exh 20: Share of ADAS Semiconductor by Category in Passenger Cars, 2025E
Exh 21: Industry Forecast for ADAS specific Semiconductor Sales, 2017 – 2025E
Exh 22: Industry Estimates for Growth in IOT End-Nodes, 2020 – 2027E
like quantum computing, it is many years away from commercial reality. The same is true for a number of proposed alternative models for data storage. For example, scientists posit that DNA molecules have the
potential to store more than a million times more information per gram than traditional magnetic storage media. While it seems unlikely that DNA read/write times will ever be competitive with solid-state storage, the sheer capacity potential could revolutionize archival applications.
In these potentially paradigm shifting technologies, it is too early to call out winners, although IBM, Alphabet, and Microsoft stand out for their aggressive investment in them. Should one of these, another existing enterprise making similar investment, or even a future startup gain advantage, the pay off would be enormous.
The Rise of China
China is currently dependent on the developed world for most of its chip supplies and all of its semiconductor fabrication equipment, leaving it vulnerable to the trade war tactics currently being played against it (Exhibit 23, 24). Obviously, addressing this dependence is a priority for Chinese government industrial policy. Between national and regional investment funds, more than the equivalent of $150B has been set aside for investment in establishing a fully independent chip industry in the country.
Still, relevance in semiconductors is easier said than done. China lacks the manufacturing process expertise needed to compete with the likes of Samsung and Taiwan Semiconductor. Indigenous leader SMIC has
Exh 23: Snapshot of Global Semiconductor Consumption Share, 2003 – 2019
Exh 24: Global Semiconductor Production Share by Company HQ, 2019
Exh 25: Chinese Semiconductor Market is Dominated by US Based Firms, 2018
struggled to implement a 14nm manufacturing process, while its Asian rivals are already moving past 7nm to 5nm. Samsung and TSM work closely with US and EU equipment suppliers, like Applied Materials, Lam Research and ASML, to develop and implement their new processes, an option that is not available to Chinese companies in the midst of a trade war. Efforts to develop homegrown IC manufacturing technology, like Extreme Ultraviolet lithography (EUV), start from far behind with substantial security controls by governments wary of Chinese attempts to import the machines or the scientists that designed them. Without access to advanced technology and with relatively scant domestic expertise, Chinese firms are many years behind world-class standard. Even if the erosion of Moore’s Law gives cover for China to play catch up on technology, scale is the other significant problem. IF they could make world-class memory and processors, it would be exceedingly difficult reach competitive costs without comparable volumes. This will be a long row to hoe (Exhibit 25, 26).
It is possible that Chinese policy, set in 5-year plans, could shift emphasis toward chip opportunities that would not be so affected by its inability to compete at the highest level of manufacturing. In recent years, Chinese companies have filed for hundreds of patents on chips designed for AI acceleration, dovetailing with an established base of AI software expertise. While we see western AI markets dominated by Nvidia, due to the competitive moat provided by the strength of its CUDA programming environment amongst software engineers, and by Alphabet’s Tensor Processing Unit ASIC – the company has, by far, the biggest and most decorated AI development team in the industry – AI development is still in the early innings. China still starts from behind against excellent competition, but their aggressive investment remains a significant wildcard in handicapping future leadership.
Exh 26: Summary of likely focus areas for China’s semiconductor policy near-term
Good Old Analog
Analog semiconductors are different. Most of the real world is not defined by ones and zeros, but rather by continuous distributions of values. Analog components use the variable nature of electric voltage to represent and process these nonbinary signals, whereby digital components must convert natural signals into flows of ones and zeros that then approximate any analog aspects of data. There are many applications – such as sensors, power management, or the transmission of radio waves – that are inherently analog in ways that render digital approximations inadequate. This is the realm of analog ICs.
Analog designs are finnicky, susceptible to interference and tricky to lay out. These problems become far worse at small geometries and designs are difficult to port from one manufacturing process to another, so most analog components are fabricated on older, well-established equipment in order to get consistency. As such, Moore’s Law is not really a consideration for this market. It also follows that analog circuit design is entirely different from digital, and thus, requires entirely different engineers with entirely different skills to design in analog. The process is often described as more of an art than a science, as textbook layouts can fail in practice and debugging a design can be counterintuitive. This rewards experience in a field where experience is seemingly always in short supply and establishes a significant entry barrier to would-be rivals without the proper base of expertise. As such, analog chip companies tend to manage higher margins than their digital counterparts and industry capacity and pricing is less cyclical (Exhibit 27).
Analog parts are an important piece of a number of long-term growth markets, such as IoT, electric vehicles, self-driving systems, 5G, optical systems for AR, etc., and we expect the analog market to have somewhat higher growth than the overall industry. However, as we noted earlier, much of the demand for these applications is well in the future while current major end markets for analog chips, such as automotive and industrial applications, have been hit by the pandemic-induced recession. In this context, we would wait.
Winners and Losers
After nearly 40 years at the top of the industry, Intel no longer rules the roost. Taiwan Semiconductor Manufacturing Company is now the number one semiconductor player by market cap (N.B. Samsung
Exh 27: Historical Global Analog Semiconductor Sales, 2011 – 2019
Exh 28: NVDA Revenue and Growth Rates over last 5 years, 2015 – 2020
carries a slightly higher valuation but semiconductors are only about a third of company sales) We believe TSM will sustain its leadership in manufacturing going forward, but are concerned that expectations for the rest of this year and next could be aggressive given the global recession. Shares have run up nearly 80% from March lows and we suggest caution.
We are more bullish on one of TSM’s major customers, Nvidia. The company’s core GPU franchise is well positioned for the pandemic, with video gaming demand and AI acceleration for cloud datacenters seeing strong demand and well positioned to sustain momentum in recovery as well. The recent acquisition of high-performance networking chip designer Mellanox diversifies Nvidia’s products for datacenters, with meaningful cross selling and cost synergies in the combination. The recently announced deal to acquire ARM from Softbank may face significant government opposition from China, but should it gain approval, we see major upside for investors. While license fees from smartphones may be ARM’s primary franchise, we believe Nvidia could use its expertise to jumpstart efforts on RISC CPUs for the cloud datacenter market, and conceivably, to increase the adoption of the CUDA programming rubric via licensing of core GPU tech to 3rd parties. The pandemic boosted Nvidia’s topline growth to more than 40% YoY but consensus projects that to drop to less than 20% in CY2021. We believe analysts are too pessimistic for cloud datacenter and gaming spending and see meaningful upside to these numbers (Exhibit 28, 29, 30).
Broadcom and Qualcomm are numbers 4 and 5 on the semiconductor market cap list and both carry expectations for a 2021 5G smartphone boom. Consensus topline growth projections for Qualcomm are particularly aggressive given the global recession and trade restrictions on shipments to China. We note
Exh 29: ARM and MLNX give NVDA lots of room for further margin expansions
Exh 30: Consensus estimates for NVDA leave sufficient room for upside surprises
reports that Chinese OEMs had been stockpiling chips in the event of a prolonged standoff, which suggest that recent quarters may have seen artificially high sales that will make for difficult compares in 2021. While we see Qualcomm as a long-term winner, with opportunity to expand its smartphone leadership into the datacenter and automotive sectors, we see more risk than reward in the near-term.
Broadcom also carries exposure to the smartphone market but is balanced by its strong position in high performance networking components used by hyperscale datacenter operators. Moreover, the investment narrative is more about margin expansion than it is topline growth. In this context the risk reward for 2021 looks a bit more positive than with Qualcomm, even though we have doubts that CEO Tan Hock can find enough M&A targets to keep his roll-up strategy going for much longer.
AMD has been the hottest stock in the sector, up almost 80% YTD, its 2007 move to fabless operations validated by Intel’s public struggles to keep up with TSM and Samsung. Consensus is modeling a hard pullback from the 40% growth experienced in the most recent quarter, leaving ample room, we think, for the company to deliver upside surprise. Longer term, AMDs reliance on CISC x86 architecture CPUs will be tested by the expected arrival of RISC alternatives for hyperscale platforms, but for now, we are comfortable with the stock. The same is true for FPGA leader Xilinx. The stock took a hit with the ban on sales to Huawei, softness in 5G deployments, and exposure to the automotive sector, but strength in datacenter for AI acceleration is offsetting those issues.
As for the rest, we fear consensus may be still jumping the gun on the cyclical turn for integrated memory makers – it would be hard to deliver upside if global smartphone volumes are disappointing, even if sales to the datacenter market surprise. As such we’d be cautious on MU and Hynix and wait to get greater
Exh 31: Key Financial and Valuation Metrics for Top Global Semiconductor Players
confidence in an economic recovery. Similarly, we see long term upside for analog chipmakers like AMD, NXP, ADI, TXN, ON, MRVL, and others, but are concerned that investors are a bit ahead of things with 2021 expectations (Exhibit 31, 32).
Exh 32: Summary of Winners for Next Phase of Semiconductors