Can AI Keep Growing Exponentially? Let’s Ask SemiAnalysis…

PodcastBig Technology PodcastDylan PatelOct 7

The discussion examines the scale and limits of AI infrastructure investment, from a projected $2 trillion in U.S. capital expenditures to constraints on chips, power, financing, and execution. It also weighs AI labs' customer retention, inference economics, and competitive pressure against risks such as cloud vulnerabilities and uncertain returns on infrastructure. Dylan Patel sees a possible path for Anthropic to reach a $10 trillion valuation, contingent on roughly $1 trillion in revenue and broader market multiples.

AI Infrastructure Semiconductors Technology Stocks Market Valuations Venture Capital Risk Management

Preview · 19 more points in the app

Read the Full Summary

Key Takeaways

The guest argues that countries see data centers and GPUs as strategic infrastructure because access to compute may shape whether they import or export AI tokens. The guest also says U.S. infrastructure supports users abroad, including in Europe, by hosting training and serving inference.

Dylan Patel estimates U.S. AI infrastructure and supply-chain CapEx will approach $2 trillion next year, while hyperscaler quarterly CapEx growth is expected to ease from 116% year over year in Q3 2026 to below 100% in Q4 and the 70% range the following Q1. The guest expects investment to remain massive but says resource, financing, and organizational constraints—and growth deceleration from a larger base—will eventually slow spending, making revenue and profit growth need to outpace it.

The speaker estimates that a $1 trillion AI infrastructure investment would require $250 billion in annual revenue over six years to repay on a straight-line basis, while also saying AI-lab inference revenue has grown tenfold year over year. Anthropic is said to cover combined training and inference compute costs, and its roughly 75% inference gross margins challenge the view that AI inference is inherently lower-margin than software, though the speaker does not establish that those margins will persist.

The guest says GPU demand exceeds supply, but self-build data-center campuses can be delayed or left as empty shells before major chip purchases because construction is a smaller share of project costs than GPUs, HBM, and optics. That flexibility contrasts with the five- or six-year neocloud contracts the guest describes, which commit customers to agreed capacity and pricing after providers secure land, power, and sometimes chips.

GEO-POLITICS

Sovereign Compute — The guest argues that countries view data centers and GPUs as strategic infrastructure they want to control. Access to that compute may determine whether a country becomes a net importer or exporter of AI tokens.

“And people, even when they're raising money, they use this word a lot because they view compute as infrastructure. And I think that's exactly right…”—

TECHNOLOGY

Chip Design Workforce — The speaker says the number of U.S. chip-design R&D engineers has remained roughly flat for 20 years, even as the sector's economic value has expanded sharply. NVIDIA, Broadcom, and other companies illustrate how value has grown without a comparable increase in designer headcount.

“Like chip design is a great example. Right? Chips…”—

Compute Buildout — The guest says GPU demand currently exceeds supply, while chip production, data-center construction, and deployment constrain how quickly capacity can grow. Self-build campuses can remain empty shells or be delayed before major chip purchases because construction is a small share of project costs compared with HBM, optics, and GPUs. The guest says it will be easier to determine when projects need to slow if demand for knowledge work or intelligence reaches limits.

“What Dylan's referring to with the potential to slow down or pause or cancel projects is more with respect to the self build sites, which is to say like, Meta…”—

AI Drug Discovery — Jordan Nanos argues that a very large Anthropic could create value by entering businesses such as cancer drug development, rather than relying only on per-token software usage. He points to Eli Lilly’s returns from GLP-1 research as an example of the potential payoff from successful drug development.

“I think that they get into other areas of business…”—

ECONOMY

AI Capital Spending — Dylan Patel estimates U.S. capital expenditures across AI infrastructure and its broader supply chain will approach $2 trillion next year. Hyperscaler quarterly CapEx growth was cited at 116% year over year in Q3 2026, with consensus expecting it to fall below 100% in Q4 and into the 70% range in the following Q1. The guest expects investment to remain massive but says triple-digit growth cannot continue indefinitely as expansion and resource constraints bind.

“Yeah. I mean, the first thing I... When that chart was going around, I was like, well, People's numbers are too low. It's actually much higher than that already. Really?…”—

Global AI Infrastructure — The guest argues that U.S. AI infrastructure spending supports users abroad because the country hosts training and serves inference for other countries, including Europe. In the speaker's view, this makes the economic impact of U.S. AI infrastructure larger than that of earlier physical-infrastructure investment cycles.

“But for AI infrastructure, your CapEx in America is actually a lot of it's being used to serve outside of America uses. Right?…”—

Expansion Constraints — The guest says it is increasingly difficult for hyperscalers and neoclouds to marshal the resources needed to add another gigawatt of capacity or raise capital for projects. These organizational and financing constraints may limit further infrastructure expansion, even though current investments are producing strong returns.

“One is it becomes really hard for the biggest neo clouds and hyperscalers to just organize and marshal the resources required to build another gigawatt on top of what they're…”—

MARKETS

NVIDIA Backstops — The speaker says NVIDIA is supporting $588 billion in off-balance-sheet backstops, an amount compared with 2.5% of the U.S. M2 money supply. NVIDIA can act as an investment-grade off-taker to reassure creditors financing GPU deployments, then direct capacity to a higher-paying customer rather than use the GPUs for its own research. The speaker says this support also helps NVIDIA maintain a diverse customer base instead of relying only on hyperscalers.

“And the number that we have is that currently we've got tracking of $588,000,000,000 of off balance sheet backstops that NVIDIA is supporting right now, which is one of yet…”—

Supply-Chain Financing — The speaker says NVIDIA's financing support extends beyond GPUs to construction, memory suppliers, fabs, and industrial equipment. This points to potential financing bottlenecks across the broader supply chain, not just the cost of buying chips.

“And then maybe the last thing to say about this is that it's not just the chips. Like, we track a whole wide range of things that they're backstopping…”—

FORECASTS

Anthropic's Valuation — Dylan Patel says Anthropic could become the world's first $10 trillion company. He estimates that reaching that valuation might require revenue of roughly $1 trillion, potentially in 2028 or 2029; the outcome depends partly on revenue multiples and broader stock-market valuations.

“Anthropic could even be the first company to be a $10,000,000,000,000 valuation company in the world…”—

AI Infrastructure Scale

Patel argues U.S. AI-related capital spending is approaching $2 trillion, or roughly 5–6% of GDP, and exceeds the scale suggested by historical infrastructure comparisons.

CapEx Growth Limits

The guests expect continued infrastructure expansion but say growth will eventually decelerate as capital, construction capacity, and supply-chain resources become harder to marshal.

AI Investment Payback

Patel says revenue trails infrastructure spending, but current AI demand and the potential growth of lab revenue could justify investment; he also rejects claims that Meta faces imminent solvency trouble.

AI Revenue Potential

The guests argue that AI returns may come from industries such as drug discovery, robotics, and chip design, while Patel says automation could expand output even if labor’s share of the economy falls.

Knowledge Work Disruption

Patel expects high-agency workers to remain influential in the near term, while warning that knowledge work could be transformed if AI becomes better than humans across most tasks.

Neo Cloud Contract Risk

Nanos distinguishes long-term neo-cloud capacity contracts from hyperscalers’ more flexible self-build plans and says current GPU demand still outstrips the supply that providers can bring online.

Anthropic Valuation Stakes

Patel sees a possible path for Anthropic to become a $10 trillion company, but says such valuations would require enormous revenue and could intensify concerns about concentrated wealth and social disruption.

Anthropic Inference Margins

Patel says Anthropic’s inference margins compare favorably with leading software companies and argues that sticky customers and low acquisition costs could make AI labs unusually attractive businesses.

NVIDIA’s Neo Cloud Backstops

Nanos says NVIDIA’s financing guarantees support a diverse GPU-cloud market and help providers fund chips, construction, and other supply-chain needs amid strong customer demand.

Neo Cloud Security Gaps

Patel describes serious security weaknesses at some neo clouds, including customer data exposure, while Nanos explains why sovereign AI efforts may put government workloads on shared infrastructure.

Oracle Cloud Execution

Patel separates Oracle’s strong cluster operations from its financing and construction risks, including fixed-margin contracts, debt-market concerns, and delays at a New Mexico site.

Google’s Compute Strategy

The guests say Google is directing a declining share of its expanding compute to DeepMind, while its TPU deals and talent losses could leave it further behind OpenAI and Anthropic.

Frontier Models In-House

Patel says leading labs have kept some advanced models internal, while Nanos notes that building high-value end-user products could offer stronger incentives than selling model access alone.