Dylan Patel: $11T of AI capex through 2029, $5T of it borrowed

SemiAnalysis founder Dylan Patel returns to the Dwarkesh Podcast with one number that anchors the rest of the conversation: his firm models about $11 trillion of AI capex between 2024 and 2029, roughly $6 trillion funded from cash flow and more than $5 trillion raised as credit. Much of the episode is an argument about what borrowing that $5 trillion does to everyone who is not building a data center.

#The margin that lets two labs outbid everyone

The mechanism is unglamorous arbitrage. Base compute costs roughly $10–15 million per megawatt, and at that price almost anyone can turn a profit: rent a GB300 rack, download open Kimi weights, serve them with vLLM or SGLang, list the endpoint on OpenRouter. The frontier labs work at a different multiple — Anthropic's revenue per megawatt has reached $50 million, and Patel thinks the model layer is heading for $100 million. Anthropic turned a profit in Q2; he believes OpenAI may have in Q3.

That gap is why compute keeps concentrating. Both labs went from roughly 2 gigawatts at the start of the year to more than 5 by its end, taking about 30% of all compute added in 2026. Signed contracts put them at 40–50% next year, and Patel expects 70–80% of incremental compute by 2028 — on newer silicon (GB300s, TPUv7, Trainium3) worth 3–5x the performance per watt of two-year-old parts, so their share of usable FLOPs runs ahead of their share of watts.

Why doesn't the supply chain expand into that gap? By Dwarkesh Patel's arithmetic, $6 billion of fab capex buys a gigawatt of annual production capacity, and a gigawatt generates roughly $100 billion of end revenue a year — a hundredfold spread. Dylan Patel's answer is the bullwhip: the signal takes years to reach Carl Zeiss and ASML, the industry still plans for about 100 EUV tools a year by 2030, and the world remains capital-constrained.

#The non-consensus call: less inference, not more

Patel's sharpest claim is that the labs will spend a shrinking share of compute on inference. He pegs today's split at roughly 60% training and 40% inference — and within training, 50% research and only 10% development. Anthropic's Mythos pre-training run, he says, was under 200 megawatts for about two months, with RL smaller still at any single site; the rest of a multi-gigawatt fleet goes to research. If a megawatt sold as tokens returns $100 million but the same megawatt pointed at research buys a better model, boards choose the model. His evidence that it is already under way: Anthropic keeps adding compute every month while its revenue additions have plateaued.

#What $5 trillion of credit does to everyone else

Extraordinary data-center returns pull capital away from everything else and push rates up. Meta has been raising at 5–6%; Patel sees no reason it would not pay 8%, and a 250 bps move propagates to every other borrower — banks, telecoms, consumer packaged goods, sovereigns. Dwarkesh Patel notes that corporate income tax is under 10% of US federal revenue while payroll and income taxes are over 80%, and that debt service already eats about 20% of federal tax revenue; a five-point rate rise, with $2 trillion of annual borrowing on top, would push that above 60%. Citing economist Basil Halperin, he expects a second Volcker shock — the 1980s episode in which roughly 40 countries, mostly Latin American, defaulted — and names Pakistan and Nigeria as candidates. A higher discount rate would also crater bond-proxy equities: why pay today's price for Johnson & Johnson or a railway, Dylan Patel asks, if your discount rate is 8–10%?

#Why it matters

The centralization argument does not depend on recursive self-improvement. Dwarkesh Patel's version: frontier FLOPs grow 4–5x a year while the compute needed for a given capability falls about 3x a year, so the effective "AI population" inside a lab grows roughly 10x annually — from perhaps 10 million AI laborers this year to 100 million, then a billion. By the end of the decade, he argues, a single lab plausibly commands more labor-equivalent than there are people on Earth. Neither finds a force that reverses it. "Every force is screeching towards centralization," Dylan Patel says. "And that's scary as hell."

The one counterweight they identify is regulation, and they say it already bites: the best model in the world, Patel claims, was trained in February, OpenAI has not released Astra and paused training for two weeks, and Anthropic is sitting on its next Mythos. That slows the US labs more than it slows open Chinese models — and by capping revenue per megawatt it also caps the labs' ability to outbid everyone for compute.