The numbers were always large in private. On Sunday they became public. The Information reported that Anthropic, the artificial-intelligence company preparing to sell shares to investors, has signed agreements for at least 14.8 gigawatts of computing capacity since October, with a combined contract value that could reach $517 billion. It is the fullest accounting yet of what it costs to promise the world a frontier model, and it reframes everything investors thought they knew about Anthropic’s balance sheet.
The scale becomes clearer in context. Late last year, Anthropic presented investors with a plan to spend roughly $180 billion leasing servers through 2029. Nine months later, the contracts it has actually signed are worth as much as three times that figure, and they stretch far beyond the planning horizon of the earlier budget. The gap between the guidance and the commitments is now the central number in the company’s pre-IPO story.
Most of the capacity comes from the two cloud giants that also happen to be Anthropic’s earliest corporate backers. Agreements with Amazon and Google cover about 11 of the 14.8 gigawatts, over terms of around a decade and at a contract value estimated above $300 billion. The deals complete a transformation in which Anthropic’s investors became its principal suppliers of infrastructure, their cloud businesses standing to earn back much of what their equity stakes might pay out.
The remaining capacity is spread across a cast of AI clouds, data-center developers and one rocket company. Anthropic leased a gigawatt of servers from Microsoft’s Azure in November; it signed a contract worth up to $45 billion with SpaceX in May; last month it agreed separate deals with the cloud startups Lambda and Nscale, worth $35 billion and $45 billion respectively; and it added smaller agreements with the Norwegian data-center firm Volta and with AMD, alongside a joint $50 billion program with Fluidstack to build data centers in Texas and New York. A deal worth roughly $9.1 billion with Riot, the bitcoin miner that has been converting its power assets to AI hosting, fills out the roster.
The pattern behind the deals is a sudden change in demand. Anthropic has said that products like Claude Code, its coding agent, grew explosively this year, an abrupt shock that forced its procurement team to rewrite plans on short cycles. The company needed guaranteed capacity for training and for serving agents, and it needed it faster than its cloud partners could deliver on ordinary terms, so it signed contracts that trade flexibility for volume.
Those contracts carry a particular financial shape. Some are take-or-pay arrangements, obliging Anthropic to pay whether it uses the capacity or not, which converts what look like options into rigid liabilities. The company’s annualized revenue has reached roughly $65 billion, a figure most software companies never approach, but it still runs far behind the commitments now accumulating on its books. The arithmetic of the AI business has always been that way: pay for the future first, collect revenue later.
The disclosure arrives as Anthropic prepares to go public, at valuations that investors and analysts have floated as high as $2 trillion. The compute contracts serve two purposes in that story. They are evidence that Anthropic has secured the capacity to keep racing OpenAI, which has said it plans to secure 30 gigawatts by the end of the decade. And they are obligations that the proceeds of the offering will be expected to help fund, which means the prospectus will have to explain how a company with large revenue and larger commitments intends to close the gap.
The same report carried a second piece of news about how Anthropic plans to handle its employees after a listing. The company is considering requiring rank-and-file workers who want to sell shares to do so on preset schedules, rather than selling at moments of their choosing. The approach, still under discussion, would spread insider selling across time and remove the optics of employees cashing out at peaks, a concern for any company whose staff hold large blocks of paper wealth.
The employee-selling idea reflects a specific anxiety. When a company this closely watched goes public, every insider transaction is parsed for signals, and a wave of early selling by employees who have waited years for liquidity can dent the stock and the story. Preset plans trade away the employees’ freedom to time their sales for a measure of predictability, the same logic that public-company executives already follow with trading plans of their own.
The two disclosures belong together. The compute contracts bind Anthropic to its suppliers for a decade and to a payment schedule that assumes the AI boom continues. The employee-selling plan binds its workforce to a timetable designed to keep the stock story clean. Both are forms of commitment made before the market can vote, and both will be scrutinized by the same investors Anthropic is about to court.
Whether the strategy holds will depend on questions the company has not yet answered: how much of the $517 billion is firm obligation and how much is capacity Anthropic can shed if demand falters, and whether revenue growth can catch the commitments that now run years ahead of it. Anthropic has placed the largest bet in the industry on the durability of AI demand, and it has now shown the world the size of the wager. The prospectus will show how it intends to pay.
