Meta’s $10 Billion Anthropic Bill Can Move the IPO

Meta projected it could spend as much as $10 billion a year on Anthropic’s AI models, the New York Times reported Thursday. That bill would have been about 15 percent of Anthropic’s July revenue run rate, as bankers floated a $2 trillion listing.

Nat Friedman, Meta’s head of AI product, has already told staff that dropping Anthropic for Meta’s own tools, or OpenAI’s, could lower Anthropic’s revenue before that offering.

Meta’s $10 Billion Anthropic Bill Can Move the IPO

The Times, citing five people who know the companies, said Meta has become one of Anthropic’s largest customers even as Mark Zuckerberg, Meta’s chief executive, criticized leading labs this month without naming Anthropic or its CEO, Dario Amodei. Two of those people said Meta’s internal projection peaked at $10 billion a year. Anthropic told investors in July that its annualized revenue would pass $65 billion.

Ten billion dollars is not a rounding error on a listing of that size. It is also not the current run rate. Meta cut some Anthropic spending this summer and still pays hundreds of millions of dollars a month, two people told the Times. Both companies declined to comment.

Anthropic’s own May filing of a confidential draft with the SEC, and the $65 billion Series H round that closed at $965 billion on May 28, put those figures on a clock. Krishna Rao, Anthropic’s chief financial officer, said at the time that the money would “serve the historic demand we are experiencing” and “bring Claude to more of the places where work happens.” Bankers later told potential investors the IPO could raise more than $100 billion and value the five-year-old company at $2 trillion, the Times reported on August 21, which would top SpaceX’s June debut.

The 6,500-Word Attack on a Lab Meta Helps Fund

On August 10 Zuckerberg posted a 6,500-word essay on personal superintelligence titled “The Future is for Everyone.” He did not name Amodei. He described a future “filled with doom,” then argued that labs building AI for companies and governments would tilt power away from individuals if they pulled ahead.

It is surprising that the discourse from many developing AI is so filled with doom. I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity’s relevance would rush to build that future. Most other labs are focused on building AI for companies, governments, or other institutions, so if those labs lead, then the balance of power will favor larger institutions over individuals.

Mark Zuckerberg, chief executive, Meta, “The Future is for Everyone,” August 10, 2026

The same essay pitches a personal agent that “will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more,” with WhatsApp-style privacy and glasses as an interface. That is the public language Meta has used for Hatch, the agent Zuckerberg has called the company’s next breakthrough. Four people told the Times that Hatch has been powered and tested on Anthropic’s models.

The attack and the invoice are the same month in the same company. Zuckerberg’s letter asks the industry to spread superintelligence. Meta’s procurement file, at the high end of the projection, would have been a major slice of the lab he was warning about.

How Tokenmaxxing Pushed the Spend Toward $10 Billion

The bill did not appear as a strategy memo. It grew out of Claude Code, Anthropic’s coding tool, which Meta engineers started using heavily at the beginning of the year. By April, employees were competing on internal leaderboards to burn the most Anthropic “tokens,” units of model use roughly equal to a word fragment. They called it tokenmaxxing.

Anthropic had already told investors that coding was carrying a large share of its growth. In its February Series G post it said Claude Code run-rate above $2.5 billion, more than double the pace at the start of 2026, with weekly active users also doubling since January 1. Business subscriptions had quadrupled. Enterprise use was more than half of Claude Code revenue. The company said eight of the Fortune 10 were Claude customers, and that more than 500 customers now spent over $1 million a year, up from a dozen two years earlier. A SemiAnalysis figure Anthropic cited put Claude Code behind 4 percent of public GitHub commits worldwide.

The spend calendar inside Meta

1. Early 2026: Meta engineers pile onto Claude Code after the tool’s 2025 public launch.
2. April 2026: Token leaderboards go up; executives discuss a $10 billion annual projection.
3. May 28, 2026: Anthropic closes Series H with a $47 billion run rate already on the books.
4. June 2026: Token costs spike, the leaderboards come down, and Meta tells staff it is on track to spend billions on AI use this year and will build a tighter system to manage it.
5. This summer: Some Anthropic spending is cut; monthly outlays stay in the hundreds of millions.

The Information reported in late June that Meta had also told applied-AI engineers to limit Claude Code and OpenAI’s Codex, warning that rival output could seep into Meta’s training data and cause “serious escalations with partner companies.” The Times later said staff were moving onto Muse Code, Meta’s own coding tool. The two reports describe the same bind: the best coding assistant on the desk was the one Meta was trying to replace.

Last Summer’s Talent Raid Stopped at Anthropic’s Door

Meta spent last summer trying to hire researchers with nine-figure packages after it stood up Meta Superintelligence Labs, with Scale AI’s Alexandr Wang as chief AI officer and Friedman, the former GitHub chief executive, on products. It hired well-known people from OpenAI and Google. Three people told the Times it largely failed to recruit from Anthropic.

That miss is why the invoices exist. A lab that will not sell you its people will still sell you tokens, and Meta bought them at a volume that turned a vendor relationship into a line that finance had to notice. Friedman now runs the product side of the same lab that could not staff its way past Claude.

Tomasz Tunguz, an investor at Theory Ventures, told the Times the listing has become a battlefield of its own. “It’s like watching dueling generals compete on the battlefield,” he said. “Everybody’s gunning to try and destabilize that I.P.O.”

What Hatch Used Before Meta Could Replace Claude

Hatch is the consumer bet that makes the vendor file awkward. Zuckerberg has described it as a personal agent that works around the clock on goals, health, relationships, and money. Four people said Meta used Anthropic’s models to power and test it. One person said the version that will debut publicly will run on Meta’s latest model, not Claude.

The swap depends on a system Meta has not shipped. Internally the company is training Watermelon, a model meant to match Anthropic’s frontier systems. In July, while that work was underway, Meta paused and later resumed pretraining, four people said, which pushed a release to at least October. Meta has not announced dates for Hatch or Watermelon.

Until those models are in production, Claude is scaffolding. Paying a rival to prototype the agent you plan to run on your own weights is a way to ship on someone else’s schedule. It is also a way to keep writing checks into the quarter when that rival is talking to public-market investors.

Friedman Walked Staff Through an IPO Lever

Two people told the Times that Meta’s leaders know the spending can move Anthropic as it heads toward the offering. Friedman made the point in plain language. If Meta used only its own coding tools, or OpenAI’s, it could lower Anthropic’s revenue before the IPO, he told some employees.

That is not a theory from outside the building. It is a manager describing a budget choice as a market event. A cut that shows up as cost control in Menlo Park shows up as a customer-concentration risk in an S-1. A hold shows up as revenue Anthropic can take into banker meetings. Muse Code is the operational version of that choice: move the work in-house, and the invoice shrinks on someone else’s listing calendar.

The projection was never a signed $10 billion contract. It was a path the token boards were pointing at before finance pulled them down. The path that remains, hundreds of millions of dollars a month, is still large enough that a further cut would be visible in Anthropic’s run rate. So would a rebound if Watermelon slips again.

A $10 Billion Compute Offer Still Has No Deal

The money does not run only toward San Francisco. In June, the same month Meta warned staff about AI bills, Anthropic approached Meta about buying up to $10 billion in computing power from Meta data centers over two years, the Times previously reported. Bloomberg later described the talks as early. No deal has been announced.

If those talks ever close, both companies can book the same dollars on opposite sides of the ledger, Meta as a landlord and Anthropic as a customer, while Meta remains a buyer of Claude. That two-way invoice is the part of the Times story that the market chatter keeps returning to, because it turns a vendor fight into a pair of offsetting receipts.

The other checks already on Anthropic’s books

• Google and Amazon: The Times said the two have committed $73 billion to Anthropic while they train their own models, a split that matches public pieces of Amazon’s extra $20 billion and Google’s larger contingent stake.
• Amazon compute: Anthropic’s Series H post said AWS remains its primary cloud and training partner, with agreements for up to five gigawatts of new Amazon capacity.
• Google and Broadcom: The same post cited five gigawatts of next-generation TPU capacity, on top of Claude already running on Google Cloud and Microsoft Azure.
• Nvidia and in-house chips: Nvidia still holds multibillion-dollar partnerships with Meta and Google as those companies design their own accelerators, the same split now visible in the OpenAI Jalapeño chip still needs Nvidia story.

Google is funding a lab it also competes with, while it separately absorbs Alphabet’s 2027 Nvidia server prices. Microsoft has spent years as OpenAI’s backer and now stresses distance. The Meta-Anthropic file is the same pattern with a sharper edge, because the buyer is also the critic, and because the critic’s product chief has already named the IPO as a reason to keep or kill the spend.

Hatch still has no public launch date. Watermelon is not due before October. Meta and Anthropic declined to comment.

Disclaimer: This article is news reporting and analysis of company spending, product plans, and published fundraising figures. It is informational only and is not investment advice, a recommendation to buy or sell any security, or a forecast of Anthropic’s offering price or Meta’s results. Readers who are considering any investment decision tied to these companies should consult a licensed financial adviser or other qualified professional who can review their own situation. Figures, deal talks, and product dates reflect the sources cited as of August 28, 2026, and may change as the companies update investors or ship new models.