What this covers

Host Patrick talks with Andrew Homan, a managing partner at Maverick Silicon, and Chris Miller, author of Chip War, about how the chip industry is structured and where AI is pushing it. Miller supplies much of the history and geopolitics, and Homan the investor's view. The conversation traces the industry from fully integrated firms to the fabless design plus foundry split, and now to cloud and AI companies pulling chip design back in-house. Along the way, Intel's fall is read through the innovator's dilemma, and TSMC's lead is traced to its foundry-only focus and to Taiwan's dense supplier cluster.

The middle of the episode gets technical. Homan compares today's AI buildout with the telecom bubble, citing fiber capex at 200 to 500 percent of operating cash flow and only about 2 percent of fiber lit in 1999. Both guests then turn to physical bottlenecks: the Memory Wall, GPUs outrunning networking, CoWoS packaging constraints, and data center chips that never prioritized power. Also covered are the extreme precision of EUV lithography, Japanese chemical suppliers working at 99.9999% purity, and hyperscalers' custom chips aimed at specific workloads. Homan's claim that young semiconductor firms can run 35 to 40 percent operating margins is flagged as contested.

The last stretch is geopolitical. Miller argues no country can make cutting-edge chips alone, which makes China's self-sufficiency push harder than anything attempted before, a view marked as contested. He describes Huawei and SMIC stuck with second-tier tools because they cannot buy ASML's best machines, and says the roughly $40 billion CHIPS Act shifts manufacturing only at the margin. He also sets out three US goals: reduce Taiwan concentration, stay ahead, and deny China.

Sharpest takeaway

Homan and Miller argue that AI's economic rent will concentrate in the silicon layer and the ecosystems around it, that durable chip moats come from silicon-software integration and manufacturing ecosystems rather than raw scale, and that US-China politics, not antitrust substance, now shapes semiconductor deal outcomes.

  • Intel's decline reflects the innovator's dilemma: success in PCs and data centers made it risk-averse about new platforms like mobile and AI
  • Moats in PC and mobile came from silicon plus software, and the manufacturing moat in Taiwan rests on a dense supplier ecosystem that is hard to replicate
  • Chinese antitrust reviews have been used as political leverage, so deal approval depends on US-China relations more than deal substance

The claims · ranked79 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.86

Manufacturing moved overseas not primarily for labor cost, since fabs are highly automated, but because of ecosystem dynamics that are hard to reverse, and this makes companies prefer investing in design over manufacturing.

causalhigh valueestablishednovelty 3/4durability 4/4· Chris Miller

“It's cheaper because of the ecosystem dynamics, and those are very hard to reverse.”

0.86

Describing PC and mobile moats as silicon alone misstates their source: the moat was silicon plus software and the interface between them, which is why Intel chips are in half the world's PCs and software is often a large portion of the moat.

causalhigh valueestablishednovelty 3/4durability 4/4· Chris Miller

“it was silicon plus software and the interface between the two. Why are there Intel chips in half the world's PCs? It's because of the software that's built on top of Intel's architecture.”

0.86

In the telecom bubble, capex by companies like Global Crossing, Level 3, and Qwest averaged about 200 percent of operating cash flow and sometimes 500 percent, and only about 2 percent of installed fiber was lit up as late as 1999 and early 2000, making that buildout unsustainable.

factualhigh valueestablishednovelty 3/4durability 4/4· Andrew Homan

“was on average about 200% of their operating cash flow. ... only 2% of it was actually lit up and being used.”

0.81

TSMC's comparative advantage rests on Taiwan's dense manufacturing ecosystem of chemical suppliers, materials producers, and tool repair experts within about an hour and a half of each other on high-speed rail, which lets fabs fix and prototype quickly in ways Arizona or Japan cannot match.

causalhigh valueestablishednovelty 3/4durability 3/4· Chris Miller

“they're all right there in a very small country within an hour and a half of each other on the high-speed rail”

0.81

Huawei designs China's leading AI chips and works with SMIC to make them, but SMIC cannot buy the most advanced ASML lithography tools, so it must produce near-cutting-edge GPUs with second-tier machinery and hits scale limits.

factualhigh valueestablishednovelty 3/4durability 3/4· Chris Miller

“SMIC isn't allowed to buy the most advanced tools, including the newer lithography systems from ASML.”

0.81

Data center chips have not focused on power consumption over the past decade, so power has not been brought down, whereas mobile chips improved compute per watt because battery life constrained design, and Nvidia now faces customer demands to take power seriously.

causalhigh valueestablishednovelty 3/4durability 3/4· Chris Miller

“in data center chips, power has not really been a key focus the past decade.”

0.81

A couple of Japanese chemical firms hold unique knowledge to produce certain chemicals at 99.9999% precision, making chemicals one of the most sensitive links in the supply chain alongside tools and design IP.

factualhigh valueestablishednovelty 3/4durability 3/4· Chris Miller

“the chemicals are made in many cases by a couple of Japanese firms which have the unique knowledge to produce chemical X with 99.9999% precision.”

0.81

TSMC overtook Intel as the leading-edge manufacturer around the 10-nanometer node, roughly seven or eight years ago, because it focuses solely on the foundry model and learns from the most sophisticated design customers.

causalhigh valueestablishednovelty 3/4durability 3/4· Andrew Homan

“they're just maniacally focused on just one business model, right? All they are doing is the foundry model”

0.81

As Moore's Law benefits get harder to capture on the front end, value is shifting to back-end advanced packaging, which historically was ignored, and CoWoS packaging was a bottleneck limiting Nvidia GPU output.

causalhigh valueestablishednovelty 3/4durability 3/4· Andrew Homan

“that's why you're seeing so much interesting innovation on the back-end side, so the advanced packaging side of the equation”

0.81

There is a flywheel in which better chips produce more compute, which is used first to design and manufacture better chips, since laying out chips with 10 billion transistors requires computing and lithography shape optimization now involves extraordinary computation.

causalhigh valueestablishednovelty 3/4durability 3/4· Chris Miller

“There's a sort of flywheel of we make better chips, we produce more compute, and we use it first and foremost to make better chips and produce more compute.”

0.80

Embracing the innovator's dilemma is painful because it requires sacrificing the products people built careers on and promoting people from new business lines that threaten the old ones, so the profitable status quo is preferred until it is too late.

causalhigh valueestablishednovelty 2/4durability 4/4· Chris Miller

“It's much easier to keep the status quo, especially if the status quo looks like it's profitable. And it was profitable for a long time. And by the time you realize it's no longer profitable, it's often too late.”

0.80

When the iPhone launched with AT&T as launch partner, AT&T had a $250 billion market cap versus Apple's $70 billion, but fourteen years later Apple is at $3.5 trillion and AT&T at $150 billion, showing that Apple extracted the hardware and App Store economics.

factualhigh valueestablishednovelty 2/4durability 4/4· Andrew Homan

“Apple just extracted all the economics out of that transition, capturing obviously the hardware with the iPhone, also the software with the App Store.”

0.79

A young semiconductor company doing a couple hundred million dollars of sales can run 35 to 40 percent operating margins, far better than software companies burning cash, because software is capital intensive later in life while semis are capital intensive early.

factualhigh valuecontestednovelty 4/4durability 3/4· Andrew Homan

“capital intensity hits you later; in semis, it hits you early.”

0.78

A key limit on AI performance is the memory wall: data must move between processor and memory at speeds limited by electrons, so architectures that bring memory closer, or use optical interconnects, could lower the cost of compute and reshape who leads in chip production.

causalhigh valuecontestednovelty 3/4durability 4/4· Chris Miller

“One of the key limiting factors in performance improvements is the need to move data back and forth from processor to memory.”

0.78

Intel's early role in the PC revolution offers a template for being early in a growth cycle, as it rode the wave for two and a half to three decades despite DRAM-style cycles, because more PC buyers dampened visible cyclicality.

causalhigh valueestablishednovelty 3/4durability 3/4· Chris Miller

“because more people were buying PCs, you didn't really notice those cycles anywhere near the way you notice the DRAM cycle”

0.76

EUV lithography involves hitting tin with a laser heated to about 100 times the temperature of the Sun, then bouncing the light off the smoothest mirrors in the world, a process as precise as hitting a golf ball from Earth into a hole-in-one on the Moon, which explains why chips do not come off the line perfect.

factualhigh valueestablishednovelty 2/4durability 4/4· Andrew Homan

“we're going to use this laser to heat this tin to 100 times the temperature of the Sun.”

0.75

Hyperscalers are building custom chips purpose-built for specific workloads such as search or social media feeds because this can significantly lower their cost to serve, as their NVIDIA GPU spend grows toward ever-larger totals.

causalhigh valueestablishednovelty 2/4durability 3/4· Andrew Homan

“custom-build a chip that is very purpose-built for one specific workload, whether that be search, for example, or some sort of social media app that will allow them to potentially lower their cost to serve those customers significantly.”

0.75

The chip industry began with fully integrated firms that designed and produced their own chips, shifted to fabless design plus TSMC foundry, and now large AI and cloud companies are reverting toward designing their own chips to control everything from software to silicon.

factualhigh valueestablishednovelty 2/4durability 3/4· Chris Miller

“they're also reverting to the prior practice of designing their own chips and getting really deep into the weeds of the technicals of how their chips are produced”

0.75

Chip production concentration in Taiwan is effectively unavoidable because economies of scale mean few locations can reach the required size, perhaps Korea too, and data center builds for foundation models are so expensive that only oil-rich or very large economies can justify them.

causalhigh valueestablishednovelty 2/4durability 3/4· Chris Miller

“you're just not going to have multiple locations at the size of Taiwan where chips are produced.”

0.73

No country can produce cutting-edge chips alone, and China's attempt at full self-sufficiency is harder than anything tried before because the last 30 years favored specialization, with Taiwan in manufacturing, Japan in chemicals, and the US in design.

causalhigh valuecontestednovelty 3/4durability 3/4· Chris Miller

“No one can do it on their own right now. The U.S. can't produce cutting-edge chips on its own. Japan can't. Taiwan certainly can't.”

0.72

The CHIPS Act's roughly $40 billion is a lot of money but only moves manufacturing location at the margin, since TSMC spent almost that much on capex in a single recent year, so reshifting manufacturing will take a long time.

factualhigh valueestablishednovelty 3/4durability 2/4· Chris Miller

“TSMC spent almost that much in CapEx last year. And so even really substantial government money is enough to move the needle, but it moves the needle on the margin”

0.71

Chip yields are low because manufacturing is the most complex process humans have attempted, billions of transistors must nearly all work, and tolerances are nanometers rather than millimeters, while TSMC rolls out a new, tighter process every year, forcing yield learning to restart each cycle.

causalhigh valueestablishednovelty 2/4durability 3/4· Chris Miller

“cars can have fault tolerance measured in millimeters. For certain parts in chips, you're talking about nanometers, so a thousand times more precise.”

0.71

The US government has three goals: mitigate chip-making concentration around Taiwan, keep US firms technologically ahead in every possible segment, and prevent adversaries, China above all, from accessing cutting-edge AI chips.

normativehigh valueestablishednovelty 2/4durability 3/4· Chris Miller

“One is to mitigate chip-making concentration around Taiwan. Number two is to keep U.S. firms out in front technologically in every segment that's possible. And then three is to prevent adversaries”

0.71

GPU performance has been rising much faster than networking and memory performance, so GPUs cannot be fully utilized, and if networking and memory cannot keep up, bottlenecks may slow the ability to train ever-larger models.

causalhigh valueestablishednovelty 2/4durability 3/4· Andrew Homan

“GPU performance has been increasing at a much faster rate than the networking and memory performance.”

0.70

The US wants most of the world's GPUs deployed in the United States and is wary of letting too many go to other countries, due to competition with China and a belief that training most foundation models domestically brings real benefit.

normativehigh valueestablishednovelty 2/4durability 2/4· Chris Miller

“the U.S. wants most of the world's GPUs to be used in the United States and is wary of letting too many get deployed in other countries”

0.70

Intel's effort to buy Tower Semiconductor was waved through by Western regulators but China refused to review it, illustrating how China's approval can block a small foundry deal.

factualhigh valueestablishednovelty 2/4durability 2/4· Chris Miller

“Western regulators all waved it through, and then China refused to review it.”

0.70

Hyperscalers such as Meta, Microsoft, and Amazon are moving downward into chips by buying AI accelerator startups, integrating them, and hiring hundreds or thousands of employees with expertise they never needed before, while chipmakers like NVIDIA move upward.

factualhigh valueestablishednovelty 2/4durability 2/4· Chris Miller

“They've all bought AI accelerator startups, they've all tried to integrate them, they're all trying to understand how to bring the networking together.”

0.70

Inference, once thought to be compute-light, is becoming compute-intensive because OpenAI's Strawberry model requires much more compute and a much longer inference period to improve reasoning.

factualhigh valuecontestednovelty 3/4durability 3/4· Andrew Homan

“an area of AI that was thought to be compute light—that being inference—is now compute-intensive as well.”

0.69

In early 2023 a frontier model company's cost base was about 90 percent compute, and the cost to train a competitive model rose from about $10 million in 2022 to $100 million in 2023, $1 billion in 2024, $10 billion in 2025, and a figure Larry Ellison cited of $100 billion in 2026.

factualhigh valuecontestednovelty 3/4durability 2/4· Andrew Homan

“literally 90% of it was compute, right? ... to train a competitive model in 2022 was $10 million. In 2023, it's $100 million. In 2024, it's a billion dollars; in 2025, it's $10 billion.”

0.68

Technology has undergone an architectural change roughly every decade, moving from mainframes in the 1980s to PCs, then to mobile, with cloud arriving simultaneously, and AI is now in the very early innings of a paradigm shift with broader implications than any prior one.

factualhigh valuecontestednovelty 2/4durability 3/4· Andrew Homan

“every decade there seems to be what I would describe as an architectural change.”

0.68

Intel's problems are a function of its success: it was too successful in PC chips and data centers for too long, which made it risk-averse and hesitant to invest in new products, so it missed mobile and is now struggling with AI.

causalhigh valuecontestednovelty 2/4durability 3/4· Chris Miller

“Because Intel's problems, in my view, are actually a function of their success. They were too successful in producing chips for PCs, too successful in data centers for a long time, and it made them risk-averse and hesitant to invest in new products.”

0.65

Today's hyperscaler capex is about 50 percent of their operating cash flow, in line with the last five years, so they could double capex and still self-fund, and GPU utilization in their clouds is nearly 100 percent, unlike the idle telecom fiber.

factualhigh valuecontestednovelty 3/4durability 2/4· Andrew Homan

“their CapEx is only about 50% of that, which is dead in line with where it's been over the last five years.”

0.62

Just as fears of a polysilicon shortage for solar in 2006 and 2007 led to capacity arriving and then oversupply, the fear of a power shortage for data centers will likely be solved by market response.

forecasthigh valuecontestednovelty 3/4durability 2/4· Andrew Homan

“capitalism usually finds a way to work. And I remember if you go back to solar, for example, in the 2006, 2007 timeframe, there was a fear that we were going to run out of polysilicon”

0.61

The more advanced chips a system needs, the more lower-tech chips it requires around them, because a new car with self-driving needs many sensors and data-transfer chips, which benefits analog and sensor makers like Analog Devices.

causalhigh valuespeaker onlynovelty 4/4durability 3/4· Chris Miller

“The more advanced chips we need, the more less advanced chips we need to be in the systems around them.”

0.61

Even with more competition and falling prices, prior transitions in PC, server, and mobile still created very large silicon-layer businesses, because the total addressable market grew so much that great companies were still built.

causalhigh valuecontestednovelty 2/4durability 3/4· Andrew Homan

“Despite there being more competition, prices did come down. But the TAM was so large that still great companies were built and created over those time periods.”

0.58

Data centers may ask utilities for more power than they will use because of a double-ordering dynamic, where they request their highest possible nameplate capacity to secure access, so actual usage could be lower than forecasts.

causalhigh valuespeaker onlynovelty 4/4durability 2/4· Andrew Homan

“will the real usage actually be less than what they're stating their nameplate capacity will be? Just because there's almost like a double ordering dynamic?”

0.57

The key focus should be finding ways to minimize the money needed to build AI compute, because admiring the scale of spending is not enough and cost is likely to be one of AI's key gating factors.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Chris Miller

“the key focus right now should be finding ways to minimize the amount of money that needs to be spent in order to build out the computing we need for AI.”

0.57

There are two kinds of antitrust in chips: real review from Western regulators and political review from China, and many recent deals were sunk not because of real concern but as part of the political game with the US.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Chris Miller

“There's actually two types of antitrust that matter in the chip industry. One is real antitrust from Western regulators, and two is fake antitrust review from Chinese regulators.”

0.57

Edge AI is in very early innings, with China and the edge still at roughly the first inning compared with data centers in innings three or four, and products like glasses, cars, and industrial equipment could define it.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Chris Miller

“Unlike data center, where we're inning three or four right now, for data center we're still inning one on the edge.”

0.57

There is a real dearth of capital for semiconductor companies at Series B and beyond, as tape-out, IP, and EDA and emulation costs, which can total $50 million, ramp, and few players are willing to commit that level of capital, unlike software.

factualhigh valuespeaker onlynovelty 3/4durability 3/4· Andrew Homan

“But in that middle section—let's call it Series B and beyond—where the capital needs really start to pick up, I think there's a real dearth of capital in that market.”

0.57

Analog-based companies that do not widen their aperture to add intelligence to historically dumb chips will be disrupted, just as Nvidia disrupted the data center incumbents, and the edge wave is coming for them next.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Andrew Homan

“if you don't embrace change here and widen your aperture to be able to put intelligence into your historically kind of dumb chips, you're going to be on the wrong side of change”

0.57

NVIDIA's most plausible vulnerability is that its flexible GPU is less efficient for a specific use case; if transformer architecture becomes the standard, startups can isolate one use case and do it much cheaper, though NVIDIA could respond with custom chips but has chosen not to.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Andrew

“that will potentially be that opportunity for companies, for startups that are able to just isolate that one very specific use case and do it much cheaper.”

0.57

The AI ecosystem has three layers: chips and cloud at the bottom (NVIDIA, AMD, TSMC), foundation models in the middle (OpenAI, Anthropic, Gemini), and applications on top (ChatGPT, Office, Copilot, Tesla Full Self-Driving); the bottom layer is the most likely to capture significant economic rent in almost any scenario, while the top two layers are opaque.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Andrew Homan

“to me, almost in any scenario, I see that being the area that's going to capture a significant amount of economic rent”

0.53

Foundation model companies are the modern analogue to telecom overbuilders because they spend far more than their operating cash flow, which limits the set of firms able to fund $100 billion models to a few incumbents.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Patrick

“So if I were to take your analysis and apply it to today, the place that looks like telco is the foundation model companies.”

0.53

When people talk about a bubble it usually means there is not one, since people called this a bubble twelve months ago and it has not proven so, but semis is a growth cyclical that will have digestion periods that can be bought.

normativehigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“I think when people talk about bubble, it means you probably aren't in a bubble.”

0.53

Data center chips are around inning three or four of their innovation cycle while edge chips are only at inning one, so the edge is where new companies can emerge, because NVIDIA has concentrated resources on the data center and perhaps decreased focus on edge products.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Chris Miller

“Unlike data center, where we're inning three or four right now, for data center we're still inning one on the edge.”

0.53

NVIDIA's DNA is not oriented toward low-power chips, which are critical for battery-operated edge devices, and its data center revenue growth is so large that its auto business has been flat over eight quarters.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“part of that is just it's not in their DNA as much to be offering low-power chips”

0.53

Startup GPU clouds like CoreWeave can move faster than Amazon or Microsoft, as CoreWeave had H100s out in March 2023 while Azure and AWS took roughly six months longer to offer them, and aligning with a chip company gives the startup an advantage.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“they had their H100s out in March of 2023, and you couldn't spin up an instance of H100 on Azure or AWS until probably almost like six months later.”

0.53

Incentivizing consumers to set air conditioning warmer during summer peak load, as in Atlanta in August, could lower peak demand and reveal more excess capacity than investors assume, because the broader grid's real utilization is poorly understood.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“can you somehow incentivize folks to turn their air conditioning to a warmer temperature and bring that peak level down?”

0.53

Huawei started in telecoms and networking equipment in Shenzhen, grew by selling first in China and then internationally, learned from foreign customers, and became closer to the Chinese state over the last 10 to 15 years, shifting into a near-monopolistic supplier of AI chips for a constrained Chinese market.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Chris Miller

“now it's the Chinese AI monopoly player.”

0.52

The main risk is the time horizon for returns, because hyperscalers can spend regardless of return for the next couple of years but then need real business models, and if scaling laws hold, capex scales with model size.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Chris Miller

“the next couple of years we have a runway in which the hyperscalers will spend regardless of return. And then at some point, we're going to need to see real business models emerging.”

0.52

Of ten semiconductor startups, one or two will likely IPO and most will be acquired by large incumbents, which collectively generate about $200 billion of free cash flow a year and can pay one to three billion dollars per deal.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Andrew Homan

“if there's 10 companies, my guess is one or two will end up going public”

0.52

Durable defensibility for chip startups will come from intellectual property and the unique value of solving a pain point, because scale is supplied by third-party foundries that startups outsource to.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Andrew Homan

“the scale piece comes from the foundry players. These startups are designing the chips for the most part”

0.52

Semiconductor startups face three pain points: access to design tools from Synopsys or Cadence, access to IP building blocks, and the foundry tape-out and manufacturing process, and Maverick aims to ease these by aligning incumbents in each bucket.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Andrew Homan

“there's three big pain points for a semiconductor startup.”

0.52

Semiconductor investing became a backwater because many semi-focused funds of 20 to 30 years ago shifted their early-stage expertise to software, which captured imagination, and that expertise later retired, though capital should flow back into hardware as its value creation grows.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Andrew Homan

“a lot of that expertise just retired and left the industry. And semis are challenging.”

0.52

Going directly after NVIDIA's GPU is a challenging task, so the better opportunities lie in the surrounding bottlenecks of interconnect, memory, power, and packaging, because the data center model of five years ago is breaking.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Andrew Homan

“I think it is challenging to go after the predator that is at the very top of the food chain right now—Nvidia being that player—and there's all these interesting areas around that GPU, whether it's interconnect or memory or power-related”

0.52

Because large buyers fear becoming dumb pipes, they want a diverse, healthy silicon supplier base; they pursue custom chips, established incumbents such as AMD, and startups, creating a three-bucket strategy to avoid dependence on NVIDIA.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Andrew

“So it's going to be a combination of custom chips that these hyperscalers are developing themselves.”

0.52

Eventually a merger approved everywhere except China will push a company to defy the veto, but because China is large and most firms need its market, companies with Chinese customers or assets must in practice accept its no.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Chris Miller

“unless you're in a niche where you don't care about Chinese customers and you don't have any assets in China—you've got to listen when they say no.”

0.52

The vast majority of semiconductor startup founders are much more experienced than software founders, often having run divisions at large public semis companies, so the average founder age gap can be around 20 years, making experience-based management teams the norm.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Andrew Homan

“the vast majority tend to be much more experienced. So these are individuals that have run divisions at the large public semiconductor companies”

0.52

Government works at government speed while the chip industry works at Moore's Law speed, so the long-run priority is for US chip firms to sustain exponential growth, since the whole economy is levered to Moore's Law and political issues will sort themselves out if they do.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Chris Miller

“government works at government speed. The chip industry works at Moore's Law speed.”

0.51

Since ChatGPT's release, NVIDIA's fundamentals have tracked its stock price closely, so its earnings multiple is flat to slightly down, yet many active managers still do not own it because they fear it is too expensive or already discovered.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“the fundamentals have tracked dead in line with the stock price. And so if anything, I think the earnings multiple is probably flat to slightly down”

0.51

Valuations for private semiconductor companies have been palatable relative to software, and the semiconductor index has returned roughly 25 percent CAGR over the last decade, so an illiquidity premium should make private semis outperform public ones.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“this is an area where, at least to date, you've seen valuations actually be quite palatable, much different than what you see in software”

0.50

Semiconductors remain a cyclical business with periodic digestion periods, but it is a growth cyclical and an attractive end market, so downturns are good times to deploy capital from a position of strength even though the path up is not a straight line.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Andrew

“I've been doing semis for long enough to know that this is still a cyclical business, despite what people will say.”

0.49

Unlike telecom fiber, GPUs in cloud deployments run at nearly 100% utilization, so the AI buildout is financed sustainably and used, though whether LLM value will justify the spend is still debatable.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Andrew

“it is nearly 100%. There's nothing sitting around idle”

0.48

The UAE is the surprising upside player in AI, having invested heavily and produced sophisticated Falcon models, while China is the surprising downside player because its AI capital expenditure is not matched by US big tech levels, except possibly ByteDance.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Chris Miller

“the surprise player right now is the UAE. The Emirates have put a ton of money into AI.”

0.48

The semiconductor industry is roughly $700 billion today versus about $800 billion for software, and Andrew expects semis to exceed software by the end of the decade as human effort shifts into compute.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Andrew Homan

“I think the semiconductor market will exceed the software market by the end of the decade”

0.47

The Middle East holds one of the very large pools of capital that could fund frontier AI spend, and the U.S. is comfortable with Gulf investment in U.S.-based data centers, such as the Microsoft and Mubadala vehicle, but is wary of frontier data centers located in the Gulf.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Chris Miller

“the U.S. is a lot more worried about investment in the Middle East because then the U.S. has less leverage, the Middle East has more.”

0.44

Andy Grove, an early Intel leader, wrote Only the Paranoid Survive about exactly this kind of paradigm-shift risk, yet Intel's management still missed the transitions to mobile and now AI.

factualestablishednovelty 1/4durability 4/4· Andrew Homan

“Interestingly, Andy Grove, one of the initial leaders at Intel, wrote a book called 'Only the Paranoid Survive,' exactly about this.”

0.34

The semiconductor supply chain is far more interwoven and longer than software, with many siloed players, so no single person can hold a view of the whole picture, and even Chris finds that understanding the full chain is almost impossible.

factualspeaker onlynovelty 2/4durability 3/4· Chris Miller

“there are so many steps in the supply chain that actually it's really hard, almost impossible, to have a view of the whole picture.”

0.31

Semiconductor innovation relies on a wider talent pool of thousands to hundreds of thousands of technical engineers rather than a handful of 100x engineers, though figures like Jim Keller are exceptional.

factualspeaker onlynovelty 2/4durability 2/4· Andrew Homan

“I think it is actually probably a wider talent pool than maybe what you're seeing with some of these AI kind of frontier model places”

0.31

The list of companies able to sustain frontier model spending of around $100 billion is very short, and mainly includes companies with large operating cash flows or large Middle Eastern capital pools, which may make the technology an incumbents' dream.

forecastspeaker onlynovelty 2/4durability 2/4· Andrew

“that list is pretty short.”

0.31

The risk is that a single company, Nvidia, captures all the economic value of this transition, so a vibrant startup community and diverse silicon supplier base are needed to avoid over-concentration.

normativespeaker onlynovelty 2/4durability 2/4· Andrew Homan

“what happens if there's one company that's just capturing all the economic value of a transition like this?”

0.31

John Gaddis gave Chris an intellectual blank check by telling him to study whatever he wanted for his PhD, which built the confidence that with the right tools one can enter any field and produce something of value.

normativespeaker onlynovelty 1/4durability 4/4· Chris Miller

“John said 'Study whatever you want to study' and gave me sort of an intellectual blank check”

0.28

The data center that existed five years ago, with an Intel processor at the center plus simple networking and storage, was wiped out by an AI tsunami, and NVIDIA's Jensen Huang is riding it, so everything in the data center must be rebuilt because GPU power needs are off the charts.

causalspeaker onlynovelty 2/4durability 2/4· Andrew Homan

“you had a tsunami, an AI tsunami, that basically wiped that data center out, and Jensen's riding his surfboard on top of that tsunami.”

0.28

Improved semiconductor technology and biotech/genetics are likely to intersect in the next decade rather than this one, and PhD research at leading universities suggests this will lead to company formation.

forecastspeaker onlynovelty 2/4durability 2/4· Chris Miller

“not this decade, but next decade there'll be a very interesting intersection of improved semiconductor technologies and biotech.”

0.26

The kindest thing anyone has done for Andrew was his high school and college coaches' and teachers' mentorship, which he tries to replicate at Maverick, and a rowing coach's unprompted check-in years later showed that someone still cared long after he left.

normativespeaker onlynovelty 0/4durability 4/4· Andrew Homan

“the mentorship that I got from my teachers and my coaches over that period of time had such a huge impact on who I am today.”

0.26

Software businesses that used to be people-intensive, with large sales forces and R&D teams, are shifting toward cost structures where compute and chip spending dominate, as shown in hyperscaler capex numbers.

factualspeaker onlynovelty 2/4durability 2/4· Andrew Homan

“is that now changing where just the cost of business is being much more compute and chip driven versus people driven?”

0.26

The semiconductor list of top companies is much shorter than AI's, and half of the top AI companies were founded in the last decade, while semis has fewer new company formations because large incumbents dominate.

factualspeaker onlynovelty 2/4durability 2/4· Chris

“Half of them were founded in the last decade. Do the same in semiconductors, and it's a very different ratio.”

0.21

Stable reading advice: the Grove book should be mandatory reading for the whole industry, not just Intel management, given the repeated pattern of missed transitions.

normativespeaker onlynovelty 1/4durability 2/4· Andrew

“Why is this not mandatory reading for the entire industry”