What this covers
Jay uses AMD's history to explain how the semiconductor industry is built. AMD settled on CPUs in the 1980s but kept missing roadmap and ship dates. Customers moved to Intel, which then had more profit to reinvest. The 2006 ATI purchase added GPUs but left AMD second to Nvidia as well. The 2008 fab spin-off into GlobalFoundries is offered as a textbook case for specialization. Around 2015 to 2016 Intel stalled at roughly 10 nanometer, three to four years behind TSMC, and AMD, building on TSMC, captured the full gains of Moore's Law.
The middle of the conversation is industry economics. Fabless design is capital-light, with a team of about 20 able to produce a chip. A leading-edge fab costs about $7 billion, and manufacturing costs rise nearly as fast as performance. Only TSMC and Samsung can make the most advanced chips, and TSMC's plants are all in Taiwan, which Jay treats as a geopolitical choke point. He sets the $52 billion CHIPS Act against TSMC's roughly $44 billion of capex in 2022 alone. He also contrasts AMD's 2021 margins with Intel's, which are weighed down by fab depreciation and utilization reportedly in the fifties.
The forward-looking half is more contested. Jay argues that x86 software optimization protects incumbents, and that moving to ARM means costly re-optimization. Nvidia's CUDA locks in AI developers. The largest internet companies, which consume about 70 percent of server CPUs, are now designing their own chips. He cites Google's TPU cutting its data center builds by half, but warns that a custom chip built only to save money is at best break-even. He predicts specialized ASICs will take over AI work and that the industry will swing back toward vertical integration, with manufacturing staying specialized.
Jay argues that AMD's history shows semiconductor advantage comes from execution, software ecosystem lock-in, and asset-light design, and that the industry is shifting from a few general-purpose CPU giants toward custom silicon designed by hyperscalers and specialized fabless firms.
- Intel and AMD rivalry was decided by product-cycle execution, and AMD's missed roadmaps cost it share in a self-reinforcing loop
- x86 software optimization creates switching friction that protects incumbents, while ARM's migration cost is the last 20% of manual work
- Hyperscalers designing their own chips, with Broadcom and AMD acting as ASIC partners, is a pendulum swing back toward vertical integration without manufacturing
Specialized accelerators and custom silicon are replacing general-purpose CPUs as Moore's Law slows.
- Data centers are moving from roughly 100 percent CPU toward something like 50 to 60 percent CPU with 40 to 50 percent other accelerators, a model called heterogeneous compute that mixes CPUs with custom chips.
“we're starting to see these data centers go from a hundred percent CPU to 50%, 60% CPU and 40% everything else.”
- Data center servers have been almost entirely Intel x86 chips for about 30 to 40 years, and the data center was essentially a giant room full of CPUs before AI workloads started shifting it.
“If you think of servers in the data center, it's an Intel x86 chip. It's like a hundred percent.”
- What is called AI is in practice linear algebra, which is more efficiently done on GPUs than CPUs, so the software behind products marketed as AI typically runs on Nvidia GPUs in data centers.
“Really what it is linear algebra. And for a variety of reasons, you don't want to use a CPU to do linear algebra.”
- The largest internet companies, the Super 7 including Google, Facebook, Amazon, Microsoft, Baidu, Alibaba and Tencent, are designing their own chips, and along with a few others they consume about 70 percent of server CPU output, which makes customers into competitors.
“The seven biggest data center companies, the Super 7: Google, Facebook, Amazon, Microsoft, Baidu, Alibaba, Tencent. Those companies are all designing their own chips.”
- The semiconductor industry faces a pivotal transition in which general-purpose computing's 40-year dominance is giving way to specialization because Moore's Law gains are slowing.
“Let me start off by saying general compute has been the pattern for the last 40 years. We're now at a transition point.”
- Industrial and automotive companies such as John Deere are becoming chip designers, with John Deere designing its own chips for autonomous tractor driving, which Jay expects to spread to automotive, industrial and aerospace firms.
“They're actually designing their own chips now to do autonomous tractor driving essentially.”
- Entering the CPU market remains hard because it requires a sizable team and deep software support, though many Chinese companies are now building CPUs and GPUs, and ARM is gaining as its power and cost advantages kick in.
“In China, especially we're starting to see a lot of companies come up with GPUs and CPUs”
- Intel's founders understood in the 1960s that chips could be made cheap enough to be treated almost as disposables for consumer electronics, a view most of the industry lacked at the time.
“understood that you could make chips cheap enough to power all kinds of devices to the point where you could almost think of it as disposables”
- ARM-based CPUs are reaching an inflection point because companies see better performance and lower bills from their lower power consumption, which makes ARM more attractive despite the porting friction.
“We're getting now to an inflection point where ARM is starting to look a lot more attractive”
- A chip is a collection of tiny electronic circuits integrated together, and these integrated circuits can perform math very quickly and accurately, making chips the unit of computation.
“you can make all of those circuits together that are integrated together, integrated circuits, can do math”
Custom chips are rational and strategically advantageous for big hyperscalers, not merely cost savers.
- At Google scale, where every search runs through multiple AI iterations for billions of users, a special-purpose ASIC doing AI math is more efficient than a GPU, so data centers will increasingly use custom chips for AI.
“It's even better at doing AI math than the GPU.”
- Moore's Law, the doubling of transistor density roughly every 18 months, has slowed so that the cadence is now closer to three to four years, even though it drove decades of compounding improvement in productivity.
“I said 18 months a few minutes ago, now it's more like three years, four years.”
- Custom chips used to be impractical because a company designing its own chip would find a new CPU generation with roughly double the performance out before the chip shipped, but when CPU gains slow to four or five years, building a custom chip becomes rational.
“But if it's not taking four or five years to see those performance gains on CPUs, suddenly making my own chip makes a lot more sense.”
- A company designing its own chip only to save a few dollars on chip cost is at best break-even, because the team, backend support, working capital and foundry payments are expensive, so the justification must come from strategic advantage elsewhere, such as Google's TPU letting it cut the number of data centers it must build by 50 percent.
“if you're just going to do it to save a few dollars on the cost of that chip, it doesn't make sense.”
- The closest analogy to hyperscalers designing their own chips is grocery store private label, though the chip case is far more complex and the decision is driven by strategic advantage rather than cost savings alone.
“I think there are decent analogies, private label groceries.”
- Google designed a video coding unit, the VCU, for YouTube's compression and decompression work, using TSMC to manufacture it, and it likely saves hundreds of millions of dollars a year in operating and capital expenditure without competitor chips being published against it.
“It's got to save them hundreds of millions of dollars a year in OpEx and CapEx.”
- Because a few large buyers account for most high-margin data center chip demand, the merchant chip makers' customers are becoming their competitors, a dynamic Jay argues is unusual for an industry this complex.
“And now you're starting to see your customer has become your competitor.”
- Google said its TPU AI chip allowed it to cut the number of data centers it had to build by 50%, when it was building one roughly every nine months at about $1 billion each.
“Having this chip allows us to reduce the number of data centers we have to build by 50%.”
- Google's YouTube video chip and AI chips show that hyperscalers can design special-purpose silicon for their own workloads without needing merchant suppliers, changing the industry's structure from a few general-purpose vendors.
“there was no competitor. Usually, Google releases their chip and there's some competitor who's going to publish a white paper”
Fabless design is capital-light and specialization-driven, earning better margins than fab-owning rivals like Intel.
- Fabless chip design is a fantastic business because it is capital light: a team of about 20 people can design powerful chips in a year or two and sell them at good prices, without owning equipment or heavy capex.
“It's capital light. You don't have to build machinery.”
- Fabless companies like AMD have returned much of their margin improvement to shareholders through dividends and buybacks, in contrast to Intel which raises money from every source including governments to fund its capex, and AMD has become more acquisitive, buying two companies in 2021.
“AMD like a lot of the fabless companies have been seeing great margin improvement over the last decade and have been returning a lot of that cash to shareholders through dividends and share buybacks.”
- Fabless companies like AMD have better margins than Intel because Intel carries multi-billion-dollar fab depreciation, and with fab utilization reportedly in the fifties rather than the roughly 80 percent needed for profitability, the fixed-cost burden is hurting Intel badly.
“To really be profitable, you have to run a fab at close to 80% capacity utilization.”
- In 2008 AMD decided to spin off its fabs into what became GlobalFoundries because its internal fab dependence caused under-investment, late processes and late products, and Jay calls the split the textbook case for why specialization and focus create shareholder value.
“if you were to write a business school case study on why specialization and focus matter for shareholder value, AMD is like the textbook example”
- Since the 1990s the semiconductor industry has moved toward fewer companies at the leading edge, making the foundry bottleneck a national security concern.
“fewer and fewer number of companies have been able to produce at what we call the leading edge”
- Chips are a good business, but the answer depends on whether the company is a fabless designer, which is capital-light and attractive, or a manufacturer, which is a treadmill of constant reinvestment.
“Yes, chips are a good business. I think we have to sort of break it up into those two parts of it.”
Intel and AMD outcomes turned on execution, management and market position rather than money.
- The GPU market is split roughly between Nvidia at about 70 percent and AMD at about 30 percent, a steady-state position that prevailed through the 2000s and 2010s but is now changing.
“Nvidia is 70-ish percent of the market, depending on which category you're looking at, AMD is the other 30%”
- Intel's failure to stay at the leading edge was not primarily a money problem, because it had large cash flows; it stemmed from bad management and capital allocation decisions that were organizational and structural, though enough money, time and motivation could theoretically fix it.
“If it were a money problem, we wouldn't necessarily be in this place right now. Intel had plenty of money”
- AMD bought ATI in 2006 to diversify beyond Intel into GPUs, but this placed it in a second-place position against Nvidia in GPUs, just as it was already second to Intel in CPUs.
“They acquired a company called ATI, which was making GPUs.”
- In the first decades of Intel and AMD, the decisive variable was the ability to execute on the right product at the right time, and Intel was somewhat better at execution from early on and then much better.
“I think that ability to execute on time is what really made the difference.”
- AMD settled on CPUs as its core product in the 1980s, but repeatedly missed roadmap and ship dates, which led customers to switch to Intel, which then had more profits to reinvest, creating a self-reinforcing loss spiral.
“So it was a self-reinforcing cycle for a long time.”
- The Intel-AMD rivalry has been central to the story or myth of Silicon Valley for a long time.
“For a long time, that rivalry between AMD and Intel was very central to the story or the myth of Silicon Valley.”
Advanced chip manufacturing is concentrated in few firms and enormously costly to replicate.
- TSMC earns record margins and cash flow from a near-monopoly at the leading edge, but it must spend tens of billions of dollars every year on R&D and capex just to keep moving, and it had years where capex exceeded revenue.
“they're going to have to spend tens of billions of dollars every year to sort keep that moving.”
- Only two companies, TSMC in Taiwan and Samsung in Korea, can produce chips at the most advanced seven-nanometer process nodes, creating a strategic and geopolitical choke point because TSMC's plants are all in Taiwan.
“there are really only two companies that can produce chips at the most advanced process nodes, seven nanometer, that's TSMC in Taiwan and Samsung in Korea.”
- The fabless chip design market is roughly $400 to $500 billion, and including foundries, equipment, software, test and packaging the total global semiconductor market is roughly $800 to $900 billion.
“the market for fabless semis today is about 400, $500 billion”
- The cost of building an advanced semiconductor fab is about $7 billion, and manufacturing costs rise at nearly the same rate as performance gains, which is a corollary to Moore's Law that is rarely discussed.
“the cost of manufacturing those chips increases at almost the same rate as the performance gains”
- The US CHIPS Act provides $52 billion over five years to subsidize domestic semiconductors, which is roughly comparable to TSMC's roughly $44 billion in 2022 capital expenditure alone, so catching up is theoretically possible but requires very large sums.
“the chips act, which just passed in the US to subsidize the US semiconductor industry is $52 billion to be allocated over five years”
Intel's manufacturing stall lets AMD, working with TSMC, capture Moore's Law gains.
- The most important lesson from AMD for investors is to understand a company's ability to execute, which depends on organizational structure and internal decision-making that are hard for outsiders to judge, and AMD's reputation for poor execution persisted for decades despite a decade of roadmap delivery.
“I think it's around understanding your ability to execute.”
- Around 2015 to 2016 Intel fell off the Moore's Law curve and is stuck at roughly 10 nanometer, about two process nodes or three to four years behind TSMC, which let AMD working with TSMC capture the full benefit of Moore's Law.
“Intel started falling off the Moore's law curve, and their performance improvements started slowing down”
- Moore and Grove left Fairchild in 1968 to found Intel, and a team led by Jerry Sanders left a year later to found AMD; the two started as colleagues and became intense rivals.
“A year later, another team, led by a guy named Jerry Sanders, left to go form Advanced Micro Devices, AMD.”
- Because Intel stalled, AMD, working with TSMC, now captures the full performance gains of Moore's Law, and it has been taking share from Intel in both PCs and especially data centers.
“Intel started falling off the Moore's law curve, and their performance improvements started slowing down”
Merchant chip makers should back custom ASIC projects to funnel their GPU and CPU sales.
- Merchant chip companies like AMD and Broadcom should support customers' custom ASIC projects, accepting lower but decent contribution margins, because this creates a sales funnel that leads to selling additional chips such as GPUs or CPUs for the same board or system, and resisting the trend entirely would be disastrous.
“I think a big mistake for them would be to resist it entirely and dig their feet in and say, "No, no, no."”
- Custom chips by big customers will erode the total addressable market for merchant chip companies and hurt some more than others, but will not be existential because hyperscalers only design a few strategic chips and still need many others.
“It will definitely mean an erosion of TAM, there will probably be a couple companies that do get hurt worse than others”
- Companies like AMD, Broadcom and Marvell provide ASIC design and back-end support to internet customers at decent but non-recurring margins, which they accept because it funnels into sales of GPUs, CPUs and other board components.
“It's not great business. It's decent margin because I already have the team working”
Chip industry structure repeatedly splits and spins out through history in a pendulum-like pattern.
- Starting in the 1990s the industry split into companies that design chips and companies that manufacture them, and most chip designers are now fabless, handing designs to third-party foundries, which together with Moore's Law form the two pillars of the industry.
“Most chip companies today are fabless and they work with foundries.”
- The commercial US semiconductor industry traces back to William Shockley, whose technical vision was strong but whose management was poor, and the engineers who left his team went on to found the chip companies that followed.
“we could trace all of this back to one guy named Shockley, who had great technical vision of how this stuff was supposed to go”
- Bell Labs, part of AT&T, invented the integrated circuit in the 1950s because its telecom switching network needed dense circuits, and the whole chain from service to chip manufacturing was vertically integrated under one monopoly.
“Bell Labs at that point was part of AT&T, the ultimate monopoly, completely vertically integrated.”
- Fairchild Semiconductor set up a separate chip division inside a camera-maker, and visionaries who recognized chips' potential left to found Intel in 1968 and AMD a year later, with AMD's founding team led by Jerry Sanders.
“A year later, another team, led by a guy named Jerry Sanders, left to go form Advanced Micro Devices, AMD.”
- The semiconductor industry moves like a pendulum: it went from fully vertically integrated in the 1950s to fully abstracted by the 2000s, and is now swinging back toward companies designing their own chips, though manufacturing will remain specialized.
“I think the best way to view that is to think of this industry as moving with a pendulum.”
Capital should prioritize diversification into industrial, automotive and networking markets.
- AMD bought Pensando for about $2 billion and Xilinx for about $49 billion in 2021 to add special-purpose and FPGA capabilities, preparing for heterogeneous compute, though Jay notes the Xilinx price was high for a company with about $800 million in revenue.
“They also paid 49 billion dollars for it, which is a lot of money for 800 million revenue company.”
- A capital allocator at AMD or Intel should prioritize diversification into industrial, automotive, telecom and networking, and, with a broader remit, invest heavily in software capabilities, because Nvidia's CUDA software layer is why it dominates AI GPUs.
“If I were given a slightly broader remit, I would be looking long and hard at software and how to improve my software barriers.”
- If Jay were allocating capital at AMD, he would prioritize diversification into industrial and automotive systems, plus telecom and networking, because these markets will matter in the future.
“I think that's the easy thing to do. There aren't a lot of targets left to buy”
- Nvidia's CUDA software layer gave programmers low-level access to GPUs, and when developers used it for AI in the 2000s, it became the default AI GPU platform, a software advantage AMD has struggled to match.
“Nvidia years ago created a little software layer called Kuda”
x86 software ecosystem lock-in makes switching to ARM costly despite ARM's advantages.
- Software optimized for years on x86 instructions creates a barrier to entry: switching between AMD and Intel x86 chips is easy for software companies, but moving to a different architecture like ARM is costly because software must be re-optimized.
“It's hard though, for anyone else to get into that market because all of this is run on x86 instruction sets.”
- Chips do not operate in a vacuum, so the massive software ecosystem optimized for x86 over a generation is the main reason it has been hard to displace, and chip companies that build CPUs must staff software engineers as well as hardware designers, a hard transition.
“Chips do not operate in a vacuum. They are vehicles for executing software”
- Porting software from x86 to ARM is about 80 percent automated with cross-compilation tools, but the last 20 percent is labor-intensive manual optimization that nobody wants to do, and this friction keeps people from switching even as ARM's power and cost advantages grow.
“You can press a button and cross compile your software from x86 to ARM, but then that's 80% of the work.”
Apple's vertical integration of silicon and software pays for its chip team.
- Apple designs its own silicon for Mac and iPhone, spending roughly $1 billion a year in operating expense on its chip teams, and because Apple owns the operating system it can tightly integrate hardware and software for performance and battery advantages that more than pay for the team through revenue and customer lock-in.
“In both cases, Apple owns the operating system that runs on those chips. So they're able to link the two tightly together”
- Apple's in-house chip division was staffed by people poached from Intel and AMD, and its design team resembles the design teams of merchant chip companies, handing finished designs to a back-end operations team for foundry production.
“They're all people who were poached from Intel and AMD.”
- Apple's M-series laptops have a clear power advantage with much longer battery life and often no fans, and the A-series phone chips give crisper, smoother graphics than Android devices, benefits consumers can perceive.
“The M1 Apple laptops have a pretty clear power advantage. The battery life is much longer, the fan doesn't go on.”
Consolidation to fewer chip companies has lifted industry gross margins.
- In 2021 AMD earned $16 billion in revenue with about 48 percent gross margin and 20 percent operating margin, which Jay calls decent for a fabless company, and the industry consolidated from about 2,000 chip companies in 2000 to about 200 today, driving margin improvement.
“In 2021 AMD had revenue of 16 billion dollars and they had gross margins of 48% on that”
- Over the last 20 years semiconductor gross margins rose from 20-30 percent to around 40-plus percent as the industry consolidated from roughly 2,000 chip companies to about 200.
“probably 2000 chip companies in 2000, there are probably 200 today. And so we've seen this immense improvement in gross margins across the industry”