Anthropic's 518 billion dollar compute bill is mostly locked in. OpenAI's revenue run rate nears 70 billion dollars

Two numbers arrived on 29 September that belong side by side. One is how much of the AI build-out's bill is already signed. The other is how fast the revenue meant to pay for it is growing.

Start with the bill. On Monday, Reuters reported that Anthropic's draft IPO prospectus lists 518 billion dollars of future cloud, computing and infrastructure obligations. On Tuesday it reported the breakdown. According to Reuters, which has seen the confidential filing, Anthropic expects to spend at least 518 billion dollars over a decade with six partners, and about 80 percent of that sum is non-cancelable or must be paid regardless of usage. The prospectus has not been made public, and Anthropic did not immediately respond to a request for comment.

The partners are the companies that supply most of the world's AI computing. Google: at least 111.1 billion dollars between April 2026 and July 2033, and, in the filing's words as quoted by Reuters, if Anthropic's actual spend falls short, it must pay Google the difference. Amazon: 110 billion dollars between May 2026 and April 2036, on similar terms. Microsoft: 31.4 billion dollars from November 2026 to May 2033, non-cancelable except if Microsoft commits an uncured material breach. Broadcom: about 161.2 billion dollars of equipment lease obligations, largely non-cancelable, the single largest line. And xAI, now part of SpaceX: agreements that could reach 84.5 billion dollars for Nvidia-based capacity through 2029, but largely cancelable with 90 days' notice. Reuters also reports that AMD has committed to purchase up to 5 billion dollars of Anthropic stock and to supply computing capacity expected to exceed 20 billion dollars.

Why the split matters more than the total. A headline commitment can be soft: options, capacity reservations, deals that lapse if demand does not show up. This schedule says most of it is not soft. Roughly 414 billion dollars, on our arithmetic from Reuters' 80 percent, is owed whatever happens to usage. The one large flexible block is xAI's. For Google, Amazon, Microsoft and Broadcom, that is contracted demand stretching into the next decade. For Anthropic, it is a fixed cost set against revenue that, per Reuters' earlier story, was about 4.6 billion dollars in 2025, with nearly a quarter from two customers and many large clients not on long-term contracts.

Now the revenue. Also on Tuesday, Axios reported that OpenAI's annualized revenue run rate has grown more than 70 percent since the start of the third quarter, to almost 70 billion dollars. Business-to-business revenue more than doubled over the same period, and, according to a source who spoke to Reuters, OpenAI made more consumer revenue in the quarter than in all of 2025. Reuters confirmed the run-rate figure through its own source. OpenAI has not commented. In August Bloomberg reported the run rate had passed 40 billion dollars, so the reported figure is up by roughly three-quarters in about seven weeks.

One of the nine conditions on this map is lab revenue: the top three labs lifting combined annualized revenue from about 100 billion dollars in July to at least 180 billion by the end of 2026. Seventy billion a year is about 5.8 billion a month. That is above the 4 billion a month we use as a low case for a leading lab and below the 8 billion that would signal a takeoff. It is the first run-rate figure we have been able to attach to that condition, and it points the right way.

Put the two stories together and you get the shape of the build-out in late 2026. Revenue at the leading labs is growing faster than almost anyone projected a year ago. The commitments made to meet that demand are larger still, and they are mostly fixed. Neither fact cancels the other. The question our infrastructure-economics condition asks, who pays for the build-out, now has a sharper answer: the labs have signed for it, the suppliers have booked it, and it works only if revenue keeps compounding for years.

Now the limits. Both stories rest on documents and people the companies have not put on the record. The prospectus is a confidential draft; a public S-1 could differ. The OpenAI figure comes from anonymous sources and is a run rate, not recognized revenue, and Axios says it could not learn the company's expenses. Obligations over a decade are not bills due this year. And a spending plan is evidence of demand for suppliers, not profit for the company making it.

On the map, Anthropic stays amber: meaningful, mixed, nothing red. OpenAI stays amber, reflecting both this revenue figure and its earlier decision to rule out a 2026 IPO. The lab-revenue condition moves from not enough evidence to supports. Infrastructure economics stays amber, with another data point against.

What would change the picture: Anthropic's S-1 appearing on the SEC's website with audited figures and the full schedule; either lab disclosing its own run rate or monthly revenue; a partner confirming or resizing its commitment; the xAI capacity being cancelled or cut.

The short version: the revenue is growing fast. The bill is already signed.