Meta plans to spend as much as $10 billion annually on Anthropic's AI models.

      Meta has emerged as one of Anthropic's biggest clients, even as Mark Zuckerberg publicly criticizes the company. Earlier this year, Meta estimated internally that it might spend as much as $10 billion annually on Anthropic’s models, according to a report from The New York Times that cited two anonymous sources discussing confidential information. Anthropic projected in July that its yearly revenue would exceed $65 billion, making a $10 billion contribution from a single customer a substantial portion of that total. Five informed individuals indicated to the paper that Meta has become a significant user of Anthropic’s offerings. The figures were reported by Eli Tan and Kalley Huang. Both Meta and Anthropic declined to comment.

      The two companies are direct competitors. However, what has not been previously reported is the extent of their financial entanglement amidst competition. Meta already pays Microsoft hundreds of millions annually to use AI models, as reported by TNW earlier this month. Google and Amazon have collectively invested $73 billion in Anthropic while developing their competing models.

      Zuckerberg has publicly addressed the situation without naming Anthropic or its CEO Dario Amodei. In a 6,500-word essay published this month, he stated that major AI labs are seeking to centralize power while forecasting a bleak future. He warned that if these labs dominate, larger institutions will have the upper hand over individuals. This essay forms part of a broader public campaign, with Meta running advertisements emphasizing its commitment to people, while portraying unnamed competitors as the ones automating jobs.

      Despite these public statements, Meta has been significantly utilizing the rival lab's models throughout much of the year. The spending began to escalate at the start of the year as Meta engineers extensively adopted Claude Code, Anthropic’s coding tool. By April, employees were competing internally to maximize their use of Anthropic tokens, a practice known in the industry as tokenmaxxing. Meta executives had discussions about the $10 billion projection at that time, but as token costs rose, the company decided to remove the leaderboards. In June, Meta informed its employees that it was on track to spend billions on AI in the current year and would create a more efficient system to oversee those expenses. Microsoft made similar moves in August by establishing budgets at the division level.

      This summer, Meta reduced some of its spending with Anthropic but is still investing hundreds of millions of dollars per month into the company, according to two sources.

      Meta has utilized Anthropic’s models to develop and test a consumer agent referred to as Hatch internally, which Zuckerberg has characterized as Meta’s next breakthrough. He stated that Hatch is designed to operate continuously on users' behalf to aid in achieving their personal goals related to health, relationships, and finances. However, the public launch version will run on Meta’s latest model, as noted by one source. TNW reported last week that Meta is contemplating a subscription fee of $199.99 per month for the agent.

      Additionally, Meta is creating a model to rival Anthropic’s top product, which is called Watermelon internally. The company paused the pretraining phase in July but later resumed it, delaying the launch until at least October. Meta has not disclosed any release dates for either Hatch or Watermelon.

      As Meta moves toward a public listing, its leadership is aware of the implications of their spending. Anthropic is preparing for an initial public offering that could value the company at $2 trillion, and TNW reported this month that it may aim to raise up to $100 billion. Nat Friedman, Meta’s AI product head, has mentioned to some employees the potential consequences of this dynamic. If Meta exclusively uses its own coding tools or those from OpenAI, it could reduce Anthropic's revenue before their IPO, according to two sources.

      Meta recently updated Muse Code, the coding tool it introduced in August, and its employees are increasingly opting for that over Anthropic’s offerings. Microsoft’s AI leader stated in June that his company aims to eliminate its payments to Anthropic, suggesting that this is not a unique strategy devised by Meta. Tomasz Tunguz, an investor at Theory Ventures, commented, “It’s like watching dueling generals compete on the battlefield. Everybody’s gunning to try and destabilize that IPO.”

      In June, Anthropic proposed an offer to Meta, suggesting purchasing up to $10 billion in computing power from Meta’s data centers over a two-year period, a plan TNW reported when the discussions initially came to light. No deal has been publicly announced by either company.

      The two $10 billion figures represent opposing interests: one is what Meta projected it would pay Anthropic for its models, and the other is what Anthropic proposed to pay Meta for computing capacity. Both proposals surfaced in the same month, but neither has resulted in confirmed spending.

      Last summer, Meta sought to recruit AI researchers from competitors by offering nine-figure compensation packages, successfully hiring some notable individuals from OpenAI and Google. However, it largely struggled to attract talent from Anthropic