AMD: The Challenger

Like most investors in Asia, I’ve been following Taiwan Semiconductor Manufacturing Company (“TSMC”) for a long time. TSMC was the world’s first independent foundry, offering semiconductor designers a way to manufacture their chips without incurring the heavy cost of the capital investment required. Instead, by aggregating the capital-intensive part of the value chain and ploughing its earnings back into R&D and more advanced processes, TSMC has accelerated the development of the ecosystem for everyone.

While COVID made 2020 a year none of us will ever forget, it should also be remembered as the year Intel announced it would be severely delaying its 7nm node, effectively ceding the leading edge to TSMC - semiconductor jargon for the ability to manufacture at scale the densest and most power-efficient transistors. Intel’s misstep was the result of decisions taken years earlier, specifically to forego the design and manufacture of less-power-intensive CPUs for the massive Mobile market, and then to reject the use of Extreme Ultra-violet (“EUV”) lithography for its advanced processes. Its loss represented a tectonic shift in the nexus of global technology: the most advanced hardware and software are now dependent on chips made by TSMC in Taiwan.

I summarise Longriver’s investment strategy as “Global Value Investing, from Asia”. It draws on my deep experience investing in Asia, and how observing the interaction between Asian and international companies as suppliers, customers, competitors and imitators has often led me to discover superior businesses in other parts of the world. For all its world-class consumer-facing apps, digital services and electronic devices, Asia is entirely dependent on the Western tech stack, particularly semiconductors and semiconductor capital equipment. But unlike my peers with a mandate to invest only in Asia, I now get to invest in other parts of the tech stack, too, including Advanced Micro Devices (“AMD”), a client of TSMC, which directly benefits from the latter’s process lead.

The World’s Best Business Model:

Before I introduce AMD, it’s important to first introduce Moore’s Law, which named after the eponymous engineer, posits that the number of transistors on an integrated circuit would double every two years. This increases the amount of computational power for a given cost or, seen the other way, lowers the cost for a given amount of computational power. This is because semiconductors require less voltage and become more power efficient as they fit more, smaller transistors into a given area.

Moore’s Law is not a law of physics; it is simply an exponential relationship based on empirical observation from manufacturing experience. But it is perhaps the best and most important business model of the last century. As shown below in this diagram from Dutch semiconductor capital equipment maker ASML, Moore’s Law has enabled a virtuous cycle: as integrated circuits (“chips”, “semiconductors”) become better and cheaper, they enable the design and development of new, superior and more affordable digital devices and services; superior performance spurs consumers to upgrade, while lower prices broaden the market; some of the profits from these new devices is re-invested into the research and development of better and cheaper integrated circuits; and the process repeats.

This virtuous cycle has driven the development and ubiquity of our digital age. It is also why the semiconductor industry, while cyclical in the short term (due to high fixed costs, long lead times to expand supply and inventory cycles), can be counted on to enjoy secular growth in the long term.

ASML follows up with this visualisation of the global technology stack, showing the industry’s fifty largest firms and their place in the ecosystem, from Semi Equipment to Semi Manufacturers to Semi Design to Hardware to Software & Services. The size of each logo represents its Earnings Before Interest & Tax (“EBIT”) as a proportion of the overall USD 688 billion ecosystem. You will surely recognise most of these companies; the key is to understand how interdependent they are. The hardware and software which enables our digital world would not be possible without the highly advanced semiconductors which undergird them; likewise, it would be impossible to fund the development of those highly advanced semiconductors if the software and application layers weren’t so profitable too.

OVERVIEW:

This brings us back to AMD, a fabless semiconductor company which is Intel’s primary competitor in the market for Central Processing Units (“CPUs”), general-purpose processors which act as a computer’s brain. AMD also boasts a portfolio of accelerated computing processors, including Graphics Processing Units (“GPUs”), Field-Programmable Gate Arrays (“FPGAs”) and Data Processing Units (“DPUs”). The table below shows the composition of AMD’s revenues in 1Q23, and hints at the inventory cycle which followed the surge in demand COVID created for PCs and laptops to work from home. Note that the figures for the Embedded segment are flattered by the acquisition of Xilinx, which closed at the end of 1Q22 (more on this below).

THE SECOND SOURCE:

AMD was founded in 1969 when eight engineers left Fairchild Semiconductor to start their own venture, following Gordon Moore and Robert Noyce, who had left Fairchild a year earlier to found Intel. The eight engineers recruited Jerry Sanders, an ambitious engineer and salesman who had recently been fired from Fairchild, and he joined them on condition of becoming President. Sanders was gritty and resourceful, and steered AMD until his retirement in 2004.

AMD did not always share in the ecosystem’s profitability. In fact, it captured very little value at all for most of its first five decades because Intel’s dominance left a scant surplus for its competitors. From 1991 (the earliest data I can access) to 2013 (the year before AMD’s current management took office), AMD’s average return on equity was negative 55 per cent, it earned negative free cash flow of USD 1.7 billion, and its share count increased by more than four times.

Like Intel, AMD developed as an “integrated device manufacturer” (“IDM”), which both designed and fabricated chips. However, AMD never boasted Intel’s leading technology or, later, its brand power. Instead, its break came in 1975 when it became the ‘second source’ for Intel’s 8085 processor: to diversify its supply chain, IBM obliged Intel to licence its IP to AMD and to give AMD some share of its production. In 1982, Intel went further by licensing its x86 architecture to AMD to produce a series of clones under the AMD brand. x86 would become the dominant architecture for CPUs across the market, as it still is today.

As a second source, however, AMD’s economics were terrible. It was the marginal producer, and its orders were always first to be cut in a downcycle. Nor did it have the scale to match Intel’s investment in leading-edge fabs. To the detriment of the company’s balance sheet, Sanders infamously refused to outsource manufacturing, saying, “Real men have fabs”. AMD’s relationship with Intel soured just as it looked like AMD was gaining traction and scale. Intel launched a ruinous lawsuit in 1987 to terminate AMD’s x86 licence and even went as far as to offer OEMs and retailers rebates to not use AMD chips.

There were ups and downs over the decades, but by the early 2010s, AMD was failing. In 2015, its market cap hit a low of just over USD1bn, down more than ninety-five per cent from its pre-GFC high.

LISA SU’S TURNAROUND

AMD’s turnaround is, therefore, nothing short of a corporate miracle. Under the leadership of CEO Dr. Lisa Su (蘇姿丰博士), who took office in 2014, AMD has turned its CPUs from ‘cheap but inferior’ to ‘cheaper and often superior’. This is especially so when measured on a Total Cost of Ownership (“TCO”) basis, which incorporates ongoing costs like reliability, support and power consumption. For example, AMD estimates that over four years, its 3rd Gen. EPYC Server CPU can reduce the total cost of ownership by up to 35% compared to an equivalent Intel Xeon Server CPU. For AMD’s data centre customers who operate at hyperscale, this can translate into huge savings over time.

Since 2017 when AMD launched the first generation of its Zen architecture PC & Laptop CPUs (named “Ryzen”) and Server CPUs (named “EPYC”), it has rapidly taken market share from Intel’s quasi-monopoly. There are many ways to slice the data, but I find these two charts illustrative. The first shows AMD’s share of the overall x86 CPU market, which basically covers all CPUs except those used for mobile phones. It underscores the momentous gains AMD has made, especially in Servers.

The second chart shows share specifically within the Server market. It highlights AMD’s dramatic gains in 2022, a period when its supply was constrained by TSMC reaching the limits of its capacity. It’s worth noting that Intel’s pricing umbrella has allowed AMD to substantially raise its ASPs while still offering value in absolute terms and per unit of compute. The chart also shows the small but meaningful rise of ARM and RISC-V architectures which prioritise power savings and flexibility, respectively.

AMD’s success is thanks to several big technical bets Su and her team made early in their tenure, pushing on all four levers to improve performance: manufacturing process, design-technology co-optimisation, design innovation and advanced packaging.

A focus on high-value and high-margin segments like gaming consoles, data centres and high-performance computing. This was smart because those markets are also the most demanding and care most about improvements in performance and the total cost of ownership. Clients in these segments are also sophisticated enough to be open to and able to rewrite software tuned to Intel’s CPUs to get the best performance out of AMD’s CPUs.
A ‘clean sheet’ design refresh for the micro-architecture of their new Zen family of CPUs, emphasising performance, power efficiency, memory and security. Intel, in contrast, is now on the fourteenth generation of its Pentium processors and is long overdue for a refresh.
A ‘chiplet’ approach which uses multiple smaller dies (chiplets) instead of a single monolithic die. The inherent latency issues in this approach were overcome with Infinity Fabric, AMD’s proprietary high-speed interconnect technology, which allows the chiplets to work in unison as if they are one. Chiplets improve manufacturing yields (because their smaller surface area is less prone to defects); permit manufacturing flexibility (because not all chiplets in a package have to be made on the same node); accelerate time-to-market (because chiplets are less complex to design); and allow for modularity at scale (because different combinations of common chiplets can easily tailor processors to different customers’ diverse needs; at last year’s Investor Day, AMD’s Chief Technology Officer Mark Papermaster said the company has more than forty different chiplet designs in production).
Shifting manufacturing from GlobalFoundries to TSMC. AMD spun out its manufacturing arm in 2009 (four years after Sanders retired) to form GlobalFoundries, which was intended to compete with TSMC and others as an independent foundry with AMD as its anchor customer. However, when GlobalFoundries abandoned its pursuit of leading edge nodes in 2018, the door was opened for AMD to switch to TSMC. This gave AMD process parity with Intel and, in 2020, put it ahead when TSMC took the leading edge. This is critical to understanding AMD CPU’s advantage in power efficiency: as mentioned earlier, semiconductors require less voltage and become more power efficient as they fit more, smaller transistors into a given area.

The combination of these successful bets has let AMD bring new products to market at an impressive cadence and with a reliability that underscores Intel’s recent poor execution. By improving performance and reducing TCO, AMD has been able to raise its prices. Going fabless has transformed its economics: TSMC aggregates the capital-intensive fabrication while AMD earns high margins and high returns on capital for its strengths in design. Thanks to this, AMD’s balance sheet looks radically different today than it did seven years ago.

At an event this month, AMD launched its fourth generation EPYC server CPUs, code-named Genoa, Genoa-X, Bergamo and Siena, which confidently embody its strengths and ability to modularise its design to suit different applications. The chips will be made on TSMC’s 5-nanometre node and boast 96 cores (more cores let the CPU work on more calculations simultaneously), with Bergamo offering 128. They also introduce next-generation DDR5 Memory and CXL inter-connect, which will be particularly helpful for AI applications which require substantial memory and high-speed bandwidth.

If you prefer a plain English summary, then I like this one from trade journal The Next Platform: “In terms of design wins, AMD landed a few good punches with Rome [its first generation], and definitely had Intel on the ropes with Milan [its third generation]. With the “Genoa” Epyc 9004s launched today – and the follow-on derivative “Genoa-X” and “Bergamo” and “Siena” server CPUs coming next year – AMD is wielding the metal chair.”

ADDING Xilinx and Pensando:

To lean further into High-Performance Computing, Data Centres and Enterprise, Su acquired Xilinx in 2020 for USD49bn (paid for with new shares equal to roughly one-third of AMD’s then shares outstanding) and Pensando in 2022 for USD2bn (in cash).

Xilinx was founded in 1984 and specialises in FPGAs, integrated circuits which can be reprogrammed after manufacturing. FPGAs can be used for prototyping or as an alternative to ASICs by users who do not have enough scale to justify the cost of custom silicon. Since FPGAs can be tailored to the specific logic of an application, they offer lower latency compared to GPUs and CPUs. This makes them particularly useful in situations where real-time data processing is required, which is why Xilinx has been particularly successful in markets such as communications, automotive and industrial.

Victor Peng, former CEO of Xilinx and now President of AMD, described the deal this way to a German magazine: “The two companies complement each other perfectly. Xilinx is very diversified, the company's success is based on solid businesses with steady and long-term growth. AMD, on the other hand, has grown enormously fast in recent years. When the two companies are combined, we’ll enjoy the benefits of a diverse customer and market base, and also experience enormous growth rates. This is often not the case, either companies grow very fast but are typically not very diversified, or they are very diversified but don't grow as fast in return. We will have both. AMD will move us forward in the data centre space and we can introduce AMD technologies in our core markets, such as automotive, communications and industrial. There is a huge TAM opportunity open to the company as a whole.”

Pensando was founded in 2017 and specialises in Data Processing Units (“DPUs”), processors designed to offload and accelerate networking, storage, and security tasks that a server's CPU would typically handle. DPUs are programmable and can help move computed data across a network, whether to the CPU, GPUs, or storage.

By combining CPUs, GPUs, FPGAs and DPUs under one roof and leveraging the modular design of its chiplets, AMD can now offer an even broader portfolio of processors optimised for different workloads - particularly to accelerate data centre computing.

Critically, Xilinx and Pensando bring a substantial bench of software engineering talent to AMD. As programmable circuits, both FPGAs and DPUs needed this talent to deliver the best results to clients. Su unified AMD’s AI and software efforts under Peng’s leadership, recognising his experience at Xilinx using the software layer to drive extra performance at the hardware layer. His big job is to get AMD’s software library, ROCm (Radeon Optimised Computing), to become a viable competitor to Intel and Nvidia.

The Empire Strikes Back?

An honest assessment would concede that AMD’s turnaround has been as much about Intel’s failures as its own successes. AMD is a direct beneficiary of Intel losing the leading edge to TSMC because it gives AMD’s CPUs a power efficiency advantage, right out of the gate. And by sticking to monolithic designs, Intel’s chips have been wracked by delays. Intel’s vertically integrated model has high fixed costs and high operating leverage. So, while it generated enormous profits during good times, it is now experiencing the reverse: the impact of lost market share is creating financial pressure at just the time it is facing the maximum technical pressure. How will it bail out a leaky ship with a leaky bucket?

But are these pressures abating? Under CEO Patrick Gelsinger’s leadership, Intel has made a series of hard decisions and slaughtered a number of sacred cows:

It will attempt to retake the leading edge by progressing an unprecedented five nodes in four years, and has finally accepted the need to embrace EUV to do so. This will culminate in the 18A node which Intel has guided to be manufacturing ready by 2H24.
In the meantime, it will outsource the production of high-end CPUs to TSMC, the first time it has ever relied on a third-party manufacturer to do so.
It established Intel Foundry Services (“IFS”) to make its manufacturing arm accountable to third-party clients.
It is attempting to acquire Israeli company Tower Semiconductor to bring much-needed customer service skills to IFS.
It will use a chiplet design for Meteor Lake, its fourteenth-generation PC CPUs.
It is experimenting with a technology it calls PowerVia to achieve a performance and cost advantage over TSMC.
In 1Q23, it dropped prices to stabilise its market share against AMD.
And to pay for the massive investment in process leadership, it cut headcount, cut its dividend and was the loudest voice lobbying for the USD 52 billion USA CHIPS & Science Act.

Intel’s success is far from guaranteed, but we should at least acknowledge the possibility and its potential impact on AMD. Intel estimated that the total addressable market (“TAM”) for the leading edge would grow at a 12% CAGR from 2021 to 2030, while the TAM for mature nodes would grow at just a 2% CAGR. These are just guesses but give a directional idea of the magnitude of TSMC’s current advantage - and the advantage TSMC conveys to its clients. If Intel succeeds on time with its 18A node, it will have beaten TSMC. The latest we’ve heard is that Intel confirmed in 2022 that it will use ASML’s High-NA lithography tools and that it completed chip tape-outs in March of this year. Will clients have enough confidence in Intel’s roadmap to design their systems around 18A chips? I’m not convinced. But it bears close watching.

Even if Intel fails, it continues to enjoy an enormous amount of inertia. Despite a decade of missteps, its CPU market share remains stubbornly high. It is firmly entrenched with the PC and Laptop OEMs, with whom it has dealt for decades. And while hyperscalers and the most demanding users are prepared to re-write code to optimise for AMD CPUs, many Enterprise customers appear unwilling to do the same. Investor Gavin Baker calls this Intel’s “exorbitant privilege”. Look closely at the slide below, and you’ll see that Intel - ostensibly a vertically integrated hardware design and manufacturing company - boasts 19,000 software engineers and claims that “software is one of our competitive advantages”. (In comparison, AMD has ~5,000 software engineers, many of whom came with Xilinx and Pensando).

In many cases, software written for one x86 CPU will run perfectly fine on another x86 CPU without any modifications—though it may not be fully optimised for the second CPU's microarchitecture. However, if the software uses specific features or technologies unique to Intel's CPUs - like certain instruction set extensions or other proprietary technologies - then re-writing the software to run optimally on AMD's CPUs could involve more work. In some cases, it may be necessary to write separate versions of certain parts of the software for each CPU microarchitecture.

Writing software to take advantage of specific CPU microarchitectures - like Intel and AMD's x86 CPUs - involves understanding the specific features and capabilities and then optimising software to use those features best. This involves lower-level programming and understanding complex factors such as CPU instruction sets, cache utilisation, multithreading, etc. These are tasks usually performed by experienced systems programmers and compilers, and they require knowledge that is quite specialised. This is the power of Intel’s inertia.

The Next WAVE, THE NEXT Challenge

Jensen Huang, co-Founder and CEO of Nvidia, calls the surge of interest in ChatGPT this year, “AI’s iPhone moment”. His point is that, like the iPhone, the technology undergirding ChatGPT (large language models) has been in development for many years but only just found the product to tie it together and make it a household name. This has sparked an arms race of sorts as Big Tech companies in America (and, to a lesser extent, China) rush to integrate AI into their products - and build out the data centre infrastructure necessary to support it. For this reason, AMD CEO Lisa Su estimated that AMD’s recent Data Centre & AI Technology event that the AI market will grow from USD 30 billion today to USD 150 billion by 2027.

In fact, the balance of Data Centre budgets has been shifting towards GPUs - and away from CPUs - for years. Before ChatGPT, machine learning was already vital for recommendation engines across Search, social networks, eCommerce and more. GPUs are perfect for the matrix algebra on which machine learning and AI are built: while CPUs can quickly perform complex calculations in series, GPUs are designed with thousands of cores to quickly perform relatively simple calculations in parallel (see the Appendix, below, for a visualisation).

The challenge for AMD, however, is that it is a very distant second to Nvidia in this market. In addition to superior hardware, Nvidia boasts a deep software library known as CUDA (Compute Unified Device Architecture), written specifically to allow developers to use NVIDIA's graphics processing units (GPUs) for general-purpose processing, a technique known as GPGPU (General-Purpose computing on Graphics Processing Units). CUDA has several advantages:

First-Mover Advantage: NVIDIA was one of the early pioneers of GPGPU computing, and CUDA was one of the first widely-adopted platforms that allowed developers to easily harness the power of GPUs for non-graphics tasks. This early lead gave CUDA a significant head-start in terms of adoption and development.
Wide Adoption and Support: Because of its early lead, CUDA is supported by many software applications, particularly in fields like machine learning, scientific computing, and video processing. This wide base of existing software makes NVIDIA's GPUs attractive to many users and developers.
Comprehensive Ecosystem: CUDA is more than just an API. It's a complete ecosystem with a compiler, libraries, development tools, and support for programming languages like C, C++, and Fortran. This comprehensive ecosystem makes it easier for developers to build, debug, and optimize their CUDA applications.
Performance: NVIDIA's GPUs are known for their high performance in parallel computing tasks, and CUDA is specifically designed to take advantage of this. The CUDA platform allows developers to closely manage and optimize the GPU's resources, leading to highly efficient and performant applications.

CUDA code can only run on NVIDIA GPUs. This is the core of Nvidia’s business model: the software allows it to sell its hardware at extremely high margins. AMD and Intel also have their own parallel computing platforms: AMD has ROCm (mentioned above) and Intel has oneAPI. But neither has been as widely adopted as CUDA, partly due to NVIDIA's early lead in the GPGPU computing space and its current incumbency. CUDA has set the standard, is taught in universities and occupies a huge part of developer mindshare.

In addition to launching CPUs designed explicitly for AI, AMD is trying to compete with Nvidia in the GPU market along several vectors:

Beef up its software capability: as described above, this was almost certainly a key reason for acquiring Xilinx and Pensando.
Go open-source: AMD is taking an orthogonal approach to Nvidia and embracing open-source communities and independents like Hugging Face and PyTorch. However, ROCm’s aim to be compatible with a range of GPUs and CPUs adds complexity. It can make it more challenging to achieve the same level of performance and efficiency as a platform designed for a specific hardware architecture.
A unified software stack: At the 2022 Investor Day, Victor Peng pointed out that AMD is in a pole position to write software that works well across its whole suite of processors, not just GPUs.
Heterogeneous Compute: AMD’s Instinct line of accelerators combines its GPUs and CPUs in one package for superior integration and performance. At its Data Centre & AI Technology event earlier this month, AMD showcased how the Instinct MI300, the latest in this line, could run a large language model on a single system. In the future, AMD’s management has made it clear they will use their chiplet designs to integrate FPGAs and DPUs for further acceleration.
Cross-selling to existing clients: CEO Lisa Su mentioned this on AMD’s 1Q23 earnings call, though the pool of potential buyers is so limited it’s hard to see how this will be a useful tactic unless AMD sacrifices its CPU margins to promote a bundle.
Focus on inference: Machine learning and AI can be broken into two categories, training and inference. The former uses hundreds of billions of parameters to form a model, which requires vast computational power. Inference applies the model to specific inputs to produce a local result. AMD may have a better chance of focusing on inference, where workloads will be less demanding and better suited to CPUs and FPGAs. Nvidia’s high prices leave a lot of room for customers to consider other options.
Cheaper prices.

Ironically, AMD’s best hope may be to become a second source, again: Nvidia’s high margins and limited production capacity are a bottleneck on its clients’ growth and could even compromise their strategic position (e.g. Amazon’s AWS). Their motivation must therefore be very high to find an alternative supply. There was a rumour in May that Microsoft is working with AMD to do just this. While quickly denied, Microsoft’s wording left room for speculation.

So, once again, AMD will be competing with a powerful and dynamic incumbent (which is also a client of TSMC). I suspect AMD overpaid for Xilinx, driven by fear of being snookered by Nvidia’s attempted acquisition of ARM, which would have given Nvidia a massive advantage in the development of power-efficient CPUs to pair with its GPUs. At least AMD overpaid with its own overvalued equity. It remains to be seen how well it can now deliver on its strategy and integrated portfolio of CPUs, GPUs, FPGAs and DPUs - especially in the face of Nvidia’s own efforts in heterogeneous computing, like the Grace Hopper superchip which directly competes against AMD’s Instinct MI300.

Valuation:

AMD traded for a large part of the last decade on an infinite PE because it had no earnings. To keen observers, its technical progress and sales momentum were apparent well before they manifested in the income statement. The stock began to de-rate in 2022 as concerns grew over a post-COVID inventory cycle. It has re-rated since, due to excitement about AI and reductions in the sell-side’s forecasts.

Sidebar: one thing I have come to realise is that the financial guidance from AMD’s Investor Days is, for all intents and purposes, purely aspirational.

Wrapping It Up:

My investment process emphasises selecting companies with a proven track record of profitable growth. AMD would not have met my criteria when Lisa Su became CEO in 2013, and I would have missed the 100-bagger performance she delivered. I’m OK with that because turnarounds are not what I’m trying to identify and deliver.

Fast forward to today, and AMD has indeed turned around: it has carved a large share of the PC and Server CPU market away from Intel, wresting a monopoly into a duopoly. It has delivered handsome returns on capital for the last seven years and repaired its balance sheet. These results were, of course, flattered by the huge demand boost seen during COVID, which could not have been reliably forecast. But looking forward, there are structural reasons why AMD is likely to continue to take share in the PC and Server CPU market: TSMC’s process leadership and AMD’s own innovations in design-technology co-optimisation, design innovation and advanced packaging.

This pattern of a challenger with a superior offering taking market share from a troubled incumbent matches many of my other investments, including HDFC Bank, Techtronic Industries and Adyen. Intel CEO Pat Gelsinger’s recent willingness to make hard decisions has given me a reason for pause, but I do not think his odds of success are high.

Semiconductors are superior cyclicals, with windows to buy and sell them. Nine months ago, when I first invested in AMD for clients, the CPU market was entering a downcycle as the supply chain digested excess inventory it had accumulated during the COVID boom in PC and Laptop sales. While it would hurt earnings, it was clear that this cycle would pass, eventually. And the odds seemed good that AMD’s superior line of Data Centre servers would continue taking share from Intel. I ascribed zero value to AMD’s GPU business and the words “artificial intelligence” never crossed my lips. I will therefore be the first to admit I got lucky for the wrong reasons. It’s funny, though, how often that happens when buying with a margin of safety.

The final conclusion to draw from this analysis is the centrality of TSMC to the semiconductor ecosystem. I have described three companies - AMD, Intel and Nvidia - competing to deliver the most advanced CPUs and GPUs to power machine learning and artificial intelligence. And while it may be hard to predict which of those three will prevail, all three are clients of TSMC and depend on it for their competitive advantage.

Disclosure: Longriver Investment Partners Limited manages the Longriver Partners Fund and separately managed accounts. These portfolios may hold or trade securities discussed in this article, and their holdings may change without notice.