Anthropic disclosed that its run-rate revenue "surpassed $30 billion -- up from approximately $9 billion at the end of 2025," in the announcement of a multi-gigawatt compute partnership with Google and Broadcom. That is more than a tripling in roughly eight months, and it is the clearest public evidence for a claim that has been driving a lot of industry forecasting: that frontier model serving now generates far more revenue per unit of datacenter capacity than anything the cloud industry was built around.
Key facts
- Run-rate revenue surpassed $30 billion, up from approximately $9 billion at the end of 2025.
- The Google and Broadcom partnership brings multiple gigawatts of next-generation TPU capacity online starting in 2027.
- A separate announcement confirms Anthropic will have access to up to one million TPUs, with substantially increased capacity during 2026.
- Primary sources: Anthropic's Google and Broadcom compute announcement and Expanding our use of Google Cloud TPUs and services.
Two things are worth separating here, because they get merged in coverage.
The first is what the number is. Run-rate revenue annualizes a recent period rather than reporting what was actually collected over the past year. For a company growing this fast the distinction is large: $30 billion run-rate means the most recent measured period, extrapolated, would produce $30 billion over twelve months. It is a legitimate and commonly used figure, and it is not the same as $30 billion booked. Anyone comparing it to a public company's trailing revenue is comparing different quantities.
The second is what it implies about compute economics, which is the part with consequences.
For most of the cloud era, a megawatt of datacenter capacity was worth somewhere in the low tens of millions of dollars a year in revenue. Serving frontier language models appears to have broken that ceiling. In an August 25 interview, SemiAnalysis founder Dylan Patel put the old baseline at "$10-15 million per megawatt" and said that for Anthropic specifically, "the revenue has gone as high as $50 million per megawatt."
If something like that ratio holds, it changes who can afford the next chip. Patel's framing: "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." A buyer earning five times as much per unit of capacity as everyone else in the auction does not need a supply agreement to win it - it can simply pay more. That is the mechanism behind essentially every current forecast of compute concentrating at a handful of labs.
Anthropic's own disclosures corroborate the revenue half of that story and are silent on the rest. Neither announcement states any gigawatt figure for Anthropic's current footprint. So the widely repeated trajectory of "under 2 gigawatts at the start of the year to above 5 by year-end" is an outside estimate, not a company disclosure, and should be attributed that way.
What Anthropic does state is the direction of its buildout. The Google and Broadcom partnership brings multiple gigawatts of next-generation TPU capacity from 2027, and a separate post confirms up to one million TPUs. Choosing Google's TPUs at that scale is itself a notable strategic fact: it is the largest public commitment by a frontier lab to an accelerator that is not an Nvidia GPU, and it gives Anthropic a supply path that does not compete directly with every other AI company for the same parts. Anthropic has also raised heavily to fund it, as detailed in its Series H announcement.
The honest caveats are worth stating plainly, because this is the category of number most likely to be repeated carelessly.
Run-rate figures are self-reported, unaudited, and chosen by the company for the moment they are published. Growth from $9 billion to $30 billion in eight months is extraordinary, and extraordinary growth rates are also the ones most sensitive to which month you annualize. Anthropic does not break out how much of that revenue comes through partners rather than directly - a distinction Ground Truth has covered before, in Amazon booking $53 billion on Anthropic that is not revenue. Revenue is also not profit; none of these disclosures address the cost of serving.
There is also a serious argument that the per-megawatt economics driving all of this cannot extend indefinitely. Epoch AI researcher Josh You argues in Frontier labs don't use most AI compute (yet) that the top labs combined were still probably under half of world AI compute at the end of 2025, and that with AI capital expenditure "already approaching $1 trillion per year," continued concentration would eventually require an acceleration in global chip production that "would require dramatic economic changes." Revenue density can rise faster than supply for a while. It cannot do so forever.
Originally published on Ground Truth, where every claim is checked against the primary source.
