SemiAnalysis founder Dylan Patel said on August 25 that OpenAI and Anthropic absorbed roughly 30% of all AI compute added to the world this year, and that 40-50% of next year's new compute is already committed to the two of them. Speaking on the Dwarkesh Podcast, he put both labs at about 2 gigawatts at the start of 2026 and above 5 gigawatts by year-end. The interview's pull quote is his summary of the trend: "Every force is screeching towards centralization."
Key facts
- OpenAI started 2026 at 2 gigawatts, Anthropic at less than 2; both are above 5 by year-end, a 3-4x increase.
- Those increases account for about 30% of all compute added worldwide this year, with 40-50% of next year's already "signed and penned and inked."
- Patel dates the interview's furthest forecast to end-2028: north of 50 gigawatts per lab, 100 gigawatts combined, and 70-80% of incremental compute.
- Primary source: Dwarkesh Patel's interview with Dylan Patel, published August 25, 2026.
The gigawatt numbers are the headline, but they are not the mechanism. The number that actually drives the argument is a piece of inside baseball about revenue density.
For most of the cloud era, a megawatt of datacenter capacity generated somewhere in the range of $10-15 million a year. Patel says frontier model serving has broken past that: "In the case of Anthropic, the revenue has gone as high as $50 million per megawatt." He describes what that unlocks in plain terms: "if I spend 10 bucks on inference capacity, I actually generate 50 bucks of revenue, and then I can turn around and incrementally spend all of that profit on training."
That is a flywheel, and it explains centralization better than any story about who has the best relationship with Nvidia. Whoever converts a watt into the most revenue can pay the most for the next watt. Everyone else -- enterprises, universities, smaller labs, cloud customers renting capacity for non-AI work -- is bidding against a buyer whose willingness to pay is set by a much higher return. Think of it as two bidders at an auction where one of them earns five times as much from every item won. The auction does not need to be rigged for the outcome to look inevitable.
Anthropic's own disclosures support the revenue half of that story. In its Google and Broadcom compute announcement, the company says run-rate revenue "surpassed $30 billion -- up from approximately $9 billion at the end of 2025," and that multiple gigawatts of next-generation TPU capacity come online starting in 2027. A separate post confirms access to up to one million TPUs. What Anthropic does not publish is any gigawatt figure for its current footprint, so the 2-to-5 trajectory rests on Patel alone.
The strongest published counter-argument comes from Epoch AI. In Frontier labs don't use most AI compute (yet), researcher Josh You estimates that the compute OpenAI used for research, training and inference at the end of 2025 was "around 10% to 15% of the world's operational AI compute supply," and that adding Anthropic, xAI, and the labs inside Google and Meta still leaves the group "probably still under half the world total." Global AI computing power, he writes, has grown to roughly the equivalent of 20 million Nvidia H100s.
You's deeper point is the one that should temper the forecast. If the top labs do capture most of global compute, their growth stops being a function of how much money they can raise and becomes a function of how fast the world can manufacture chips. With AI capital expenditure "already approaching $1 trillion per year," he argues, accelerating production beyond that "would require dramatic economic changes." The centralization curve contains its own brake.
There is also a gap between what the interview is titled and what its guest actually says. The headline reads "Anthropic & OpenAI will have most of the world's compute by 2028." In the transcript Patel claims 70-80% of incremental compute, and when asked to translate 100 gigawatts into a share of total world compute, he backs off: "I think that may be a little difficult, given that by 2028 they've taken 70-80% of incremental compute. And I'm not sure what happens to markets then." He raises his own accounting caveat too -- when Amazon serves Anthropic models through Bedrock, "that counts as Anthropic compute in our worldview."
The interview's second half runs further out. Dwarkesh Patel's own summary frames it as whether ">$10T of total AI capex we'll see by the end of the decade will cause a sovereign debt crisis, where hyperscaler debt raises interest rates, drives non-AI exposed countries into bankruptcy, and crashes non-AI equities." Dylan Patel's supporting argument runs through the American tax base: corporate income is under 10% of federal revenues while payroll and income taxes make up more than 80% and would shrink under automation, at a time when roughly 20% of tax revenue already goes to servicing debt.
The honest caveat is that essentially all of this is one analyst's model, stated conversationally. The gigawatt figures, the per-megawatt revenue, and the capex projections are SemiAnalysis estimates, not audited disclosures, and the two labs involved confirm neither. What is checkable is the direction: Anthropic's revenue really did more than triple in eight months, and the TPU contracts really are multi-gigawatt. The argument is that those two facts compound. Whether they compound all the way to 70% of the world's new chips is a forecast, and Epoch AI has published a serious reason to doubt it.
Originally published on Ground Truth, where every claim is checked against the primary source.
