Meta's $10 billion Anthropic spend exposes AI's frenemy dilemma

Meta internally forecast spending up to $10 billion a year on Anthropic's AI models, roughly 15 percent of the startup's expected annual revenue, even as Chief Executive Mark Zuckerberg publicly accused AI labs of seeking monopoly power.

"It's like two generals fighting on a battlefield, each trying to disrupt the other's IPO process," Tomasz Tunguz, a venture investor at Theory Ventures, said.

Meta currently spends several hundred million dollars a month on Anthropic tools despite trimming some outlays this summer. The outlay is material to Anthropic, which is preparing an IPO that could value the company at $2 trillion and raise as much as $100 billion, according to The Wall Street Journal.

The dependency cuts both ways. Meta's AI product chief Nat Friedman told some employees that shifting to Meta's own coding tools or OpenAI products would hurt Anthropic's pre-IPO revenue, while Anthropic has separately offered to sell Meta up to $10 billion in data center compute under a two-year agreement that has not been finalized.

Zuckerberg's Critique vs. Meta's Token Habit

This month Zuckerberg published a 6,500-word essay criticizing AI leaders including Anthropic Chief Executive Dario Amodei for "painting the future as a doomsday scenario" and framed the rise of AI labs as a concentration of power in large institutions. He did not disclose that Meta had become one of Anthropic's largest paying customers.

Inside Meta, engineers on Facebook and Instagram have leaned heavily on Anthropic's Claude Code coding tool. Usage climbed sharply early this year, spawning an internal contest called "tokenmaxxing" in which employees competed on leaderboards over who consumed the most Anthropic tokens. Meta halted the leaderboard in June as token costs climbed, telling staff that AI spending would reach tens of billions of dollars this year and that it would build stricter expense controls.

Building a Way Out

Meta's reliance extends beyond coding. A forthcoming AI product called Hatch, which Zuckerberg has described as a personal "agent" that works around the clock to help users achieve goals, used Anthropic models extensively during development and testing. When Hatch launches, it will run on Meta's own latest models rather than Anthropic's.

Meta is also developing a model codenamed "Watermelon" aimed at matching Anthropic's frontier models. The company paused its pre-training in July, resumed the work, and now expects release no earlier than October. Meta has not announced launch dates for Hatch or Watermelon. In parallel, Meta rolled out an internal coding tool called Muse Code, whose usage is rising and could gradually replace Claude Code.

Frenemies Across Silicon Valley

The Meta-Anthropic relationship is one example of a broader pattern. Google and Amazon have committed $73 billion to Anthropic while pushing their own AI models. Microsoft, OpenAI's earliest major backer, now stresses its independence. Nvidia holds multi-billion-dollar deals with both Meta and Google even as those companies develop in-house chips.

Meta this summer tried to poach Anthropic researchers with nine-figure compensation packages, but three people familiar with the matter said the effort largely failed. Anthropic, meanwhile, has more than 500 open roles and pays staff software engineers base salaries of $320,000 to $405,000.

The $10 billion figure is material enough to shape analyst views of Meta's earnings, which face pressure from rising AI capital expenditure, while simultaneously supporting confidence in Anthropic's revenue trajectory ahead of its listing. Anthropic more than doubled revenue to $11.6 billion in the second quarter, and traders on Polymarket now give it 58 percent odds of staging 2026's largest IPO, ahead of SpaceX at 44 percent. The risk is concentration: if Meta's Watermelon model or Muse Code tool matures, Anthropic's forward revenue projections could lose a pillar.

This article is for informational purposes only and does not constitute investment advice.