Meta Projected $10B Annual Spend on Anthropic Despite Public Rivalry

Meta has emerged as one of Anthropic's largest customers, with internal projections earlier this year suggesting the company could spend as much as $10 billion annually on Anthropic's AI models, according to two sources familiar with the matter who spoke to The New York Times.

The figure represents a striking contradiction to Meta CEO Mark Zuckerberg's public positioning. In a 6,500-word essay published this month, Zuckerberg criticized leading AI labs for attempting to consolidate power while warning of a future "filled with doom," though he did not name Anthropic or its CEO Dario Amodei directly. Meta has simultaneously run advertising campaigns positioning itself as betting on people rather than automating work away.

How the spending unfolded

Meta's use of Anthropic's products surged at the start of 2024. Engineers began heavily adopting Claude Code, Anthropic's coding tool, and by April employees were competing on internal leaderboards to maximize their use of Anthropic tokens—a practice known as tokenmaxxing. Meta executives discussed the $10 billion projection around that time.

As token costs climbed, Meta removed the leaderboards. By June, the company informed employees it was on track to spend billions on AI use that year and would implement better spending management systems. Meta has since reduced some Anthropic spending but continues to pay hundreds of millions of dollars monthly to the company, according to two sources.

Meta used Anthropic's models to power and test Hatch, its internal name for a consumer agent Zuckerberg has described as the company's next breakthrough. The agent is designed to work around the clock to help users with goals, health, relationships, and finances. However, the public version will run on Meta's own latest model rather than Anthropic's technology.

The strategic calculation

Meta is developing Watermelon, an internal model intended to match Anthropic's strongest offerings. The company paused pretraining in July before resuming, delaying release until at least October, according to four sources. Meta has also updated Muse Code, its own coding tool launched in August, with employees increasingly shifting to it from Anthropic's products.

The timing matters because Anthropic is preparing an initial public offering that could value it at $2 trillion and may seek to raise as much as $100 billion. Nat Friedman, Meta's head of AI product, has told some employees that using only Meta's own tools or OpenAI's could reduce Anthropic's revenue before the offering.

In June, Anthropic proposed buying up to $10 billion in computing power from Meta's data centers over two years, though neither company has announced a deal. The two separate $10 billion figures—one for Meta buying models, another for Anthropic buying compute—emerged from the same period but remain unconfirmed as actual spending.

Why it matters

The financial entanglement between Meta and Anthropic illustrates how competition in AI operates differently than in traditional tech markets. Major players routinely purchase services from direct competitors while building rival products—Google and Amazon have committed $73 billion to Anthropic while developing their own models. Meta already pays Microsoft hundreds of millions annually to rent AI models. This interdependence creates unusual dynamics where companies simultaneously collaborate as customer and vendor while competing for market position and talent. Meta's ability to reduce its Anthropic spending before the IPO could directly impact the startup's valuation, adding a strategic dimension to routine procurement decisions.

Meta largely failed to recruit AI researchers from Anthropic despite offering nine-figure compensation packages to poach talent from rivals last summer, according to three sources familiar with the companies. Both Meta and Anthropic declined to comment on the figures.

The New York Times reporters Eli Tan and Kalley Huang first reported these details. The Times noted it has sued OpenAI and Microsoft over copyright infringement of news content, claims both companies deny.