Mark Zuckerberg, Meta’s chief executive, recently took aim at rival Anthropic.

Without naming the artificial intelligence start-up or its chief executive, Dario Amodei, Mr. Zuckerberg said leading A.I. labs were trying to consolidate power while they painted the future as “filled with doom.”

“If those labs lead,” he wrote in a 6,500-word essay published this month, “then the balance of power will favor larger institutions over individuals.”

But even as Mr. Zuckerberg publicly criticized Anthropic, his own company had privately helped its rival to rise in power.

Behind the scenes, Meta, which owns Facebook and Instagram, has become a heavy user of Anthropic’s A.I. products — so much so that it has turned into one of Anthropic’s largest customers, according to five people with knowledge of the companies.

At one point this year, Meta internally projected that it could spend as much as $10 billion annually on Anthropic’s A.I. models, said two of the people, who declined to be identified discussing private information. That would have formed a major chunk of Anthropic’s yearly revenue, which the start-up estimated in July would pass $65 billion.

Meta and Anthropic have long competed, but the extent to which they are financially intertwined has not been previously reported. Their situation offers a window into the friend-foe relationships permeating Silicon Valley’s A.I. race, in which tech titans can bad mouth rivals in public while giving them billions of dollars behind the scenes, even while saying they may pose an existential risk to humanity.

Among the frenemies of the A.I. start-ups are Google and Amazon, which have committed to investing $73 billion in Anthropic even as they develop their own A.I. models. Microsoft was one of OpenAI’s earliest investors, but the companies now highlight their independence from each other. The chipmaker Nvidia has multibillion-dollar partnerships with Meta and Google, even as those companies work on their own chips.

Meta’s relationship with Anthropic appears especially convoluted.

Meta is so reliant on Anthropic’s A.I. models that it has used them to power and test a coming A.I. product called Hatch, four people familiar with the matter said. Mr. Zuckerberg has described Hatch as Meta’s next breakthrough, a personal “agent” that works “24/7 on your behalf to help achieve your goals and improve your life, your health, your relationships, your finances.”

Meta cut back on some Anthropic spending this summer, but is still spending hundreds of millions of dollars a month on the start-up’s tools, two people with knowledge of the matter said.

Meta’s leaders are aware their spending could affect Anthropic as the start-up heads toward a blockbuster initial public offering that could value it at $2 trillion, two people with knowledge of the company said. Nat Friedman, Meta’s head of A.I. product, has told some employees that if Meta uses only its own coding tools or ones from OpenAI, it could lower Anthropic’s revenue before its public offering, the people said.

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

Meta and Anthropic declined to comment. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied the suit’s claims.)

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

Meta’s spending with Anthropic surged at the beginning of the year, when its engineers started heavily using Claude Code, Anthropic’s popular A.I. coding tool. By April, Meta employees were competing on internal leaderboards to see who could use the most Anthropic “tokens,” which are units of A.I. use roughly equal to a word fragment. The practice was called “tokenmaxxing.”

Around this time, Meta executives discussed the $10 billion spending projection on Anthropic. But as token costs skyrocketed, Meta removed the token leaderboards. In June, the company told employees that it was on track to spend billions on A.I. use this year and would create a better system to manage that spending.

That same month, Anthropic also approached Meta, offering to buy up to $10 billion in computing power from its data centers over two years, The Times previously reported. No deal has been announced.

Meta has since released an update to its coding A.I. tool, Muse Code, which its employees are increasingly using instead of Anthropic’s tools. And while Meta tested Hatch using Anthropic’s A.I., the product that will publicly debut will be powered by Meta’s latest A.I. model, one person with knowledge of the matter said.

Mr. Zuckerberg’s company is also developing a new A.I. model intended to be as powerful as Anthropic’s cutting-edge models. It’s unclear if that model, internally called Watermelon, will catch up to Anthropic’s.

In July, while developing Watermelon, Meta paused and later resumed a stage of A.I. development called “pretraining,” which delayed its release until at least October, four people with knowledge of the matter said. Meta has not announced when Hatch or Watermelon will be rolled out.

Erin Griffith and Sri Muppidi contributed reporting from San Francisco.