Meta projected spending up to $10bn a year on Anthropic’s AI models

Meta has become one of Anthropic’s largest customers, at the same time as Mark Zuckerberg has been attacking the company in public. At one point this year Meta internally projected that it could spend as much as $10bn a year on Anthropic’s models, The New York Times reported, citing two people who declined to be identified discussing private information.

Anthropic estimated in July that its yearly revenue would pass $65bn. A $10bn line from a single customer would have been a significant share of that. Five people with knowledge of the two companies told the paper that Meta had become a heavy user of Anthropic’s products.

Eli Tan and Kalley Huang reported the figures. Meta and Anthropic both declined to comment.

The two companies compete directly. What has not been reported before is how far their finances are entangled while they do it. Meta already pays Microsoft hundreds of millions a year to rent AI models, which TNW reported this month. Google and Amazon have committed $73bn to Anthropic while building rival models of their own.

What Zuckerberg said in public

Zuckerberg did not name Anthropic or its chief executive Dario Amodei. He wrote that leading AI labs were trying to consolidate power while painting a future “filled with doom”. The line came in a 6,500-word essay published this month.

“If those labs lead, then the balance of power will favor larger institutions over individuals,” he wrote. The essay is part of a wider public campaign. Meta has been running advertisements telling the public it is betting on people, and Zuckerberg has cast unnamed competitors as the ones automating work away.

None of that is visible in the spending. Meta has been buying the rival lab’s models at scale for most of the year.

Meta’s engineers went first

The spending surged at the start of the year. Meta engineers began using Claude Code heavily, Anthropic’s coding tool. By April, employees were competing on internal leaderboards to see who could burn the most Anthropic tokens, a practice the industry calls tokenmaxxing.

Meta executives discussed the $10bn projection around that time. Then token costs climbed and the company removed the leaderboards. In June, Meta told employees it was on track to spend billions on AI use this year, and that it would build a better system to manage that spending. Microsoft did much the same thing in August, setting division-level budgets.

Meta has cut back on some Anthropic spending this summer. It is still spending hundreds of millions of dollars a month with the company, two people told the paper.

The agent was built on Claude

Meta used Anthropic’s models to power and test the consumer agent it calls Hatch internally, four people said. Zuckerberg has described it as Meta’s next breakthrough. It is meant to work “24/7 on your behalf to help achieve your goals and improve your life, your health, your relationships, your finances,” he said.

That will change at launch. The version that debuts publicly will run on Meta’s own latest model, one person said. TNW reported last week that Meta has considered charging up to $199.99 a month for the agent.

Meta is also building a model to match Anthropic’s strongest, known internally as Watermelon. In July the company paused a stage of development called pretraining, then resumed it, four people said. That delayed the release to at least October. Meta has not said when either Hatch or Watermelon will arrive.

The IPO in the background

Meta’s leaders know what their spending does to a company heading for a listing. Anthropic is preparing an initial public offering that could value it at $2tn, and TNW reported this month that it may seek to raise as much as $100bn.

Nat Friedman, Meta’s head of AI product, has told some employees what follows from that. If Meta uses only its own coding tools or ones from OpenAI, it could lower Anthropic’s revenue before the offering, two people said.

Meta has released an update to Muse Code, the coding tool it launched in August. Its employees are increasingly using that instead of Anthropic’s tools. Microsoft’s AI chief said in June that his company wants to eliminate what it pays Anthropic, so the tactic is not Meta’s invention.

“It’s like watching dueling generals compete on the battlefield. Everybody’s gunning to try and destabilize that I.P.O,” said Tomasz Tunguz, an investor at Theory Ventures.

The money runs both ways

In June, Anthropic approached Meta with an offer of its own. It proposed buying up to $10bn in computing power from Meta’s data centres over two years, a plan TNW covered when the talks first surfaced. Neither company has announced a deal.

The two $10bn figures point in opposite directions. One is what Meta modelled paying Anthropic for models. The other is what Anthropic offered to pay Meta for compute. They are separate proposals from the same month, and neither has turned into confirmed spending.

Meta could not hire them

Meta spent much of last summer trying to poach AI researchers from rivals with nine-figure compensation packages. It hired some high-profile names from OpenAI and Google. It largely failed to recruit from Anthropic, three people with knowledge of the companies said.

Microsoft was an early OpenAI investor, and the two now stress their independence from each other. Nvidia runs multibillion-dollar partnerships with Meta and Google while both design their own chips. Buying from a competitor is the normal condition of this industry, not the exception.

The Times noted its own position in the story. It has sued OpenAI and Microsoft over copyright infringement of news content, and both companies deny the claims.

Two dates make the rest of this testable. Watermelon is due no earlier than October, and Anthropic’s offering is the deadline Friedman was pointing at. Whether Meta’s monthly payments keep falling before then is the number to watch.