Meta Platforms CEO Mark Zuckerberg has recently turned his public criticism toward competitor Anthropic. Without explicitly naming the AI startup or its chief executive Dario Amodei, Zuckerberg suggested that leading AI labs are painting a future "filled with catastrophe" while simultaneously consolidating their own power. In a 6,500-word essay published this month, he wrote, "If these labs dominate the industry, the balance of power will tilt toward large institutions, leaving individuals at a disadvantage."

Yet, even as Zuckerberg publicly criticizes Anthropic, Meta has privately been bolstering its rival's growth. According to five sources familiar with the situation, Meta, the parent company of Facebook and Instagram, has quietly become one of Anthropic's most significant customers, using its AI products at a massive scale. During one period this year, Meta internally projected spending up to $10 billion annually on Anthropic's AI models, two insiders who requested anonymity said. That expenditure would account for a substantial portion of Anthropic's annual revenue, which the startup estimated in July would surpass $65 billion.

While Meta and Anthropic have long been competitors, the depth of their financial entanglement has never been previously reported. This dynamic illustrates the pervasive frenemy relationships in Silicon Valley's AI race, where tech giants can publicly denounce rivals while funneling billions to them behind the scenes, even as they claim those competitors could pose existential risks to humanity. Other AI startups share similar frenemy partnerships, including Google and Amazon, which have committed a combined $73 billion to Anthropic while developing their own AI models. Microsoft, an early investor in OpenAI, now emphasizes its independence from the company. Chipmaker Nvidia has multi-billion dollar partnerships with Meta and Google, both of which are also developing their own chips.

The relationship between Meta and Anthropic is particularly complex. Four insiders say Meta relies heavily on Anthropic's AI models, even using them to power and test its new internal product, Hatch. Zuckerberg has described Hatch as Meta's next major breakthrough, a personal agent that "works for you around the clock, helping you achieve your goals, and improving your life, health, relationships, and financial situation." This summer, Meta scaled back some of its purchases from Anthropic, but still spends hundreds of millions of dollars monthly on the startup's tools, according to two sources. Additionally, Meta management is aware that Anthropic is preparing for a landmark IPO, potentially valued at $2 trillion, and that Meta's spending could directly influence the company's public debut. Meta's AI product head, Nat Friedman, reportedly told some employees that if Meta used only its own coding tools or OpenAI's products, it could suppress Anthropic's revenue before its IPO.

Tomáš Tongooz, an investor at venture capital firm Theory Ventures, commented, "It's like watching generals from opposing armies fight on the battlefield. Everyone is trying to find ways to disrupt this IPO." Meta and Anthropic both declined to comment. Throughout most of last summer, Meta offered nine-figure compensation packages in an attempt to poach AI researchers from competitors. While it successfully recruited several prominent AI figures from OpenAI and Google, its efforts to lure talent from Anthropic were largely unsuccessful, according to three sources.

At the start of this year, Meta engineers began using Anthropic's popular AI coding tool, Claude Code, on a massive scale, causing Meta's spending with Anthropic to surge. By April, Meta had launched an internal leaderboard where employees competed on who could consume the most Anthropic tokens (the billing unit for AI, roughly corresponding to word fragments), a phenomenon internally dubbed "tokenmaxxing." Around that time, Meta management discussed procurement estimates reaching up to $10 billion annually. However, as token usage costs escalated, Meta removed the consumption leaderboard. In June, the company told employees that its AI tool spending would reach billions of dollars this year and that it would build better systems to control those costs. Media reports indicate that in the same month, Anthropic proactively approached Meta, proposing to sell up to $10 billion in computing power to Meta's data centers over two years. That deal has not been officially finalized.

Since then, Meta has launched its own coding AI tool, Muse Code, and employees have increasingly shifted to it, reducing their reliance on Anthropic's tools. One source says that while Meta used Anthropic models to test Hatch, the publicly released version will run on Meta's own latest AI model. Meta is also developing a new AI model codenamed "Watermelon," designed to compete with Anthropic's most advanced flagship model. It remains uncertain whether this internally named model can match Anthropic's performance. Four insiders reveal that during the development of "Watermelon" in July, Meta paused and then restarted pre-training, a critical phase in AI development, delaying the model's release by at least two months to October. Meta has not yet announced launch dates for Hatch or the "Watermelon" model.