Internal projections at Meta indicated annual spending on Anthropic's offerings could climb as high as $10 billion. Mark Zuckerberg has been vocal in his criticism of AI labs like Anthropic for allegedly pursuing monopolistic power, yet he has failed to disclose a crucial detail: his company has quietly become one of the startup's largest paying customers. This scenario highlights the widespread "frenemy" dynamic that defines Silicon Valley's AI race.

According to media reports, Meta internally forecast this year that its annual expenditure on Anthropic's AI models might reach $10 billion, representing a substantial portion of Anthropic's estimated yearly revenue of $65 billion. Although Meta has trimmed some costs since the summer, it continues to spend hundreds of millions of dollars monthly on Anthropic's tools. This reliance is particularly significant for Anthropic, which is preparing for an initial public offering that could value the company at $2 trillion.

Meta's AI product chief, Nat Friedman, has informed some employees that if Meta were to switch exclusively to its own coding tools or OpenAI's products, it would negatively impact Anthropic's revenue performance ahead of its stock market debut.

Public Condemnation Versus Private Dependency

Earlier this month, Zuckerberg published a 6,500-word essay in which he, without naming names, accused AI leaders like Anthropic's CEO Dario Amodei of "portraying the future as a doomsday scenario" and characterized the rise of these AI labs as a threat of power consolidating into large institutions. At the same time, engineers across Meta's Facebook and Instagram platforms are heavily utilizing Anthropic's Claude Code coding tool.

Employee usage of Anthropic's services surged sharply at the start of the year, even spawning an internal competition dubbed "tokenmaxxing," where staff vied for the top spot on a leaderboard based on their consumption of Anthropic tokens. As token costs escalated dramatically, Meta halted the leaderboard in June and informed employees that AI spending for the year was projected to reach tens of billions of dollars, while announcing plans to establish a more robust expenditure management system.

Hatch Product Powered by Anthropic, Future Shift to In-House Tech

Meta's dependence on Anthropic extends beyond coding tools. The company's upcoming AI product, Hatch, has relied extensively on Anthropic's AI models during its development and testing phases. Zuckerberg has positioned Hatch as Meta's next breakthrough offering, describing it as a personal "agent" that "works on behalf of users around the clock to help achieve personal goals."

However, when Hatch is officially launched, it will be driven by Meta's own latest AI models rather than Anthropic's technology. Meanwhile, Meta is developing a new AI model codenamed "Watermelon," aiming to match the capabilities of Anthropic's frontier models. Meta paused the "pre-training" phase of this model in July, and although work has resumed, the timeline has slipped, with a release unlikely before October at the earliest. Meta has yet to announce official launch dates for either Hatch or Watermelon.

A Snapshot of Silicon Valley's Frenemy Relations

The complicated relationship between Meta and Anthropic is far from unique in Silicon Valley. Google and Amazon have committed $73 billion in investments to Anthropic while aggressively advancing their own AI model research. Microsoft, an early and significant backer of OpenAI, now emphasizes its independence from the company. Chip giant Nvidia has secured multi-billion-dollar deals with both Meta and Google, even as both firms develop their own chips.

Tomasz Tunguz, an investor at venture capital firm Theory Ventures, described the dynamic: "It's like two generals engaged in battle, each trying to disrupt the other's IPO process." In this context, Meta attempted to poach AI researchers from Anthropic and other competitors this summer with nine-figure compensation packages, but according to three sources familiar with the matter, those efforts largely failed.

Meta has also introduced its own in-house coding AI tool, Muse Code, which is seeing rising employee adoption and could gradually replace the use of Claude Code. In the same month, Anthropic reached out to Meta with a proposal to purchase up to $10 billion in data center computing power from Meta under a two-year agreement, but no deal has been announced between the two companies so far.