SemiAnalysis' Dylan Patel: Why the $2 Trillion AI Buildout Can Pay Off

SemiAnalysis CEO Dylan Patel says US AI capex is heading for about $2 trillion a year, bigger than any infrastructure boom in history, and that revenue can catch up. In this episode of the Big Technology Podcast, Alex Kantrowitz talks with Patel and Jordan Nanos, the SemiAnalysis researcher behind the ClusterMax 3.0 ranking of GPU clouds, about where the money to pay for the buildout comes from, why Anthropic may be a better business than software, Nvidia’s $588 billion in financing backstops, how badly most neoclouds handle security, Oracle’s awkward position, Google DeepMind’s shrinking share of compute, and why the frontier labs are keeping their best models in house.

TLDW

Patel thinks the popular charts understate the buildout: counting the whole supply chain, US AI capex is closer to 5 or 6% of GDP than 3%, partly because American data centers serve Europe and Asia too. Both guests expect spending growth to slow simply because the base is so large, while lab revenue (Anthropic grew roughly 10x this year) pays back the capex faster than skeptics expect. The second half moves to the plumbing: Nvidia backstopping neocloud debt, a cloud where SemiAnalysis accidentally saw a foreign intelligence agency’s workloads, what ClusterMax does and does not measure, and how contract length decided who profited from rising GPU prices. It ends on Google’s compute allocation and the labs’ growing habit of holding back finished models.

Thoughts

The sharpest point in the conversation is about contract length, and it explains more of the GPU cloud market than any ranking does. Oracle and CoreWeave rank near the top of ClusterMax for running clusters well, but they locked most of their capacity into long contracts, and Oracle signed its OpenAI deals at what Patel describes as essentially a fixed margin. When GPU rental prices rose, they could not capture it. Nebius, with the shortest average contract length among the top clouds, could reprice and “really jack up pricing.” So operational quality and financial outcome came apart: the best operators did not get the best returns, the most flexible ones did. Anyone reading a neocloud’s backlog as pure strength should ask the follow-up question of what price that backlog was locked in at, and whether the market has moved since.

Patel’s closing claim deserves more attention than it will get: the gap between public models and the real frontier is now a choice the labs make. He says Anthropic had Mythos ready in February, that its safety cards mention two more models since, and that the public got a reduced version (Fable) months later, while OpenAI sat on Astra and has shown math results from another model it has not released. His explanation is commercial as much as safety-driven: “the moment you make it external people can distill it.” Nanos frames the fork plainly. If per-token API margins are the prize, labs release. If the end applications (a drug, a product, an internal research loop) are worth more, the capitalist move is to pace the frontier for yourself. His bet is “minimum viable models,” released just far enough ahead of Chinese open source to keep the lead. If that is right, benchmarks of public models measure what the labs chose to sell, not what they can do.

The case that an AI lab is a better business than software is stronger than the usual “low-margin compute reseller” story allows. Patel compares Anthropic to Salesforce: both at roughly 75% gross margins, except Anthropic got there far faster, has almost no customer acquisition cost where SaaS can spend up to 30% of revenue on it, and has R&D growing slower than revenue. He also argues the revenue is sticky, since a better Chinese model rarely shifts existing customers’ spend. The weak point is the one he names himself: all of this assumes the lab keeps a frontier model worth paying for, which is why the race dynamics in the last chapter matter so much to the valuation.

The security story should worry buyers more than investors. SemiAnalysis says it accidentally saw other tenants’ storage and Slurm jobs on one cloud, including a national security agency from a country in the top 10 by GDP. Nanos explains why it keeps happening: in a seller’s market, neoclouds are rewarded for raising money and racking GPUs, not for service quality, and the failures in security and reliability are “dead simple” to fix. When there is exactly one facility in a country with the newest Nvidia racks, even a government will share it. Of 323 providers SemiAnalysis tracks, few clear its bar, and that will not change until supply catches up and customers can walk away.

Patel’s answer to “where does the money come from” is the most honest part of his optimism. He does not say AI creates jobs; he says it accelerates the decades-long shift of income from labor to capital, because tokens are capital doing labor. His hope, which he calls possibly irrational, is that the economy grows so fast that labor’s dollars stay flat or grow slowly while its share falls. His chip design example supports it: the number of US chip designers has been flat for 20 years while the industry exploded. But the same logic powers his darker scenario, where a $20 trillion Anthropic concentrates gains so heavily that people “overthrow governments.” The bull case and the bear case are the same mechanism; distribution is the variable.

Key Takeaways

Patel puts next year’s US AI capex, including the wider supply chain, at roughly $2 trillion, or 5 to 6% of a GDP of about $30 trillion, well above the 3.6% figure in a widely shared chart.
Much of America’s AI capex serves other countries: most of Anthropic’s and OpenAI’s European inference runs in US data centers, unlike railroads or fiber, which served the country that built them.
Hyperscaler capex grew 116% year over year in Q3 2026. Nanos expects real limits late next year from organizing another gigawatt of construction and raising capital in the trillions.
Revenue trails capex by design: $1 trillion of GPUs depreciated over six years needs about $250 billion a year in revenue on a straight-line basis, while the data center shells last around 15 years.
Patel expects OpenAI and Anthropic together to reach $600 to $700 billion by December next year and possibly $1.5 trillion by the end of 2027, and says only $3 trillion of AI revenue by 2031 would be a real problem.
He calls claims that Meta’s spending is “corporate suicide” silly: Meta can stop buying new capacity at any time and its cash flow covers what it has signed.
Neocloud contracts with labs and hyperscalers are typically five or six years for full capacity with no monthly cancellation; the slowdown risk sits in self-built campuses that can sit as empty shells.
Anthropic’s inference runs at roughly 75% gross margins, similar to Salesforce, with near-zero customer acquisition cost, which Patel argues makes an AI lab a better business than SaaS.
SemiAnalysis tracks $588 billion of off-balance-sheet backstops from Nvidia, covering not just GPUs but construction, memory, fabs and industrial equipment.
Nanos says Nvidia’s backstops are pragmatic: it wants many customers rather than dependence on three or four hyperscalers, and the compute it uses for its own research is a small share of revenue.
ClusterMax rates how good a cluster is to rent (reliability, security, storage, networking, Slurm and Kubernetes), not whether a stock is a buy. Oracle ranks high on service while its New Mexico site is delayed by a pipeline problem.
Google DeepMind’s share of Google’s compute has been falling, and Patel says it will have the least compute of the three frontier labs by year end as Google sells TPUs to Anthropic.
Patel says the labs now keep months of progress internal, releasing new models mainly when competition or events like an IPO force it.

Chapters

00:51 The AI Buildout Is Bigger Than the Charts Show

Kantrowitz opens with a chart comparing AI spending to the railroads as a share of GDP. Patel says the numbers are too low: near $2 trillion of US capex against roughly $30 trillion of GDP is 5 to 6%. Nanos agrees growth must slow from triple digits, as the supply chain for equipment, construction and capital strains, and Patel adds that 10x growth like Anthropic’s this year simply cannot repeat off a $100 billion base.

08:01 How Much Revenue Pays for It

Kantrowitz reads out bearish estimates, including Bain’s figure that hyperscalers need $6 trillion a year in revenue by 2031. Patel thinks everyone’s revenue requirements are too low because AI revenue will be much higher, walks through six-year depreciation math, and says Anthropic is already profitable on revenue minus compute, with OpenAI a few quarters behind. Nanos argues people lack the imagination to see revenue beyond coding subscriptions, in areas like drug discovery and self-driving.

14:19 Where the Money Comes From

Patel says companies will shift R&D and cost of goods sold toward AI, and that AI will speed up the long decline in labor’s share of GDP. He hopes the economy grows fast enough that labor income holds steady in dollars. Asked about his past line that knowledge work is doomed, he separates the next few years, when high-agency people do the most with AI, from a later world where AI beats humans at nearly everything.

19:46 Can Big Customers Walk Away From Neoclouds?

Nanos clarifies that neocloud contracts are not cancellable month to month; they run five or six years for full capacity. The flexibility belongs to self-built campuses, where construction is a small share of total cost and a company can leave a shell empty. A downturn would need demand to slow, and every model SemiAnalysis runs still shows demand outstripping supply.

24:17 The Anthropic IPO and the Economics of an AI Lab

Patel explains his post that Anthropic at $2 trillion could go to zero or $20 trillion. He is bullish, but a $20 trillion Anthropic implies trillions in revenue and capabilities so strong that gains could concentrate and destabilize society. He then compares Anthropic’s 75% inference margins and near-zero acquisition cost to Salesforce, and says pacing the frontier is limited by the race between labs and China.

34:05 Nvidia’s Backstop Universe

Nanos describes Nvidia acting as an investment-grade offtaker so neoclouds can raise debt, then passing capacity on when someone pays more. SemiAnalysis tracks $588 billion of these backstops. He rejects the idea that it props up a bubble, pointing to Nvidia’s real research use of GPUs and its interest in having many customers rather than a few hyperscalers.

38:54 Most Neoclouds Are Bad at Security

Ilya Sutskever’s post warning that rogue agents could take over neoclouds followed a SemiAnalysis article on poor neocloud security. Patel says open source models can hack dozens of neoclouds, and that SemiAnalysis once saw other customers’ data on a cluster, including a foreign government agency. Nanos ties it to sovereign AI: when only one facility has the newest hardware, everyone shares it.

43:10 ClusterMax, Oracle and Why Contract Length Matters

Patel stresses ClusterMax is not a stock pick. Oracle rates highly as a cloud while facing a fixed-margin OpenAI deal, debt worries and a New Mexico site delayed by pipelines. Nebius, with short contracts, raised prices as GPU rents rose. Nanos says the many unrecommended providers reflect a seller’s market that rewards building fast over service quality.

51:24 Google, TPUs and DeepMind’s Compute Gap

Patel says the share of Google’s compute going to DeepMind keeps falling as Google sells TPUs to Anthropic, summed up as Thomas Kurian winning and Demis Hassabis losing. Nanos says Gemini needs a niche because nobody wants the third or fourth best coding model, and Patel warns the gap widens if OpenAI and Anthropic each add 5 to 10 gigawatts next year.

54:52 Are the Labs Keeping Their Best Models at Home?

Patel says yes, citing Anthropic’s Mythos and OpenAI’s delayed releases, and expects both to ship new models soon under competitive and IPO pressure. Nanos weighs per-token API margins against the bigger prize of end applications and bets on minimum viable releases that keep the US labs just ahead of Chinese open source.

Notable Quotes

“If AI revenue is only $3 trillion in 2031, then yeah, we’ll have big problems.”
Dylan Patel, on what would actually break the buildout

“People lack the creativity to understand how returns can be viewed beyond the Claude Code subscription and per token pricing regime of today.”
Jordan Nanos, on where future AI revenue comes from

“It’s arguably a better business than software is, is an AI lab.”
Dylan Patel, comparing Anthropic’s margins to Salesforce

“You would like them to have an air gap data center, but if there’s exactly one facility in the entire country with GB300 NVL72, I mean, it’s a sellers market.”
Jordan Nanos, on governments sharing neocloud clusters

“Nobody wants the third or fourth or fifth best coding model.”
Jordan Nanos, on why Gemini needs a niche

“They’ve kept, you know, months of progress internal only, because the moment you make it external people can distill it.”
Dylan Patel, on the gap between public and internal models

Watch the full conversation with Dylan Patel and Jordan Nanos on the Big Technology Podcast here.

SemiAnalysis’ Dylan Patel: Why the $2 Trillion AI Buildout Can Pay Off