Meta’s Anthropic Compute Talks Expose Cracks in Amazon and Google’s Bet

Meta Platforms is discussing whether to lease Anthropic up to $10 billion in computing power over two years, a deal first reported by the New York Times on July 17 and confirmed by CNBC and Reuters. Anthropic pitched the idea in June. Nothing is signed, either side can walk away early, and Meta shares closed down more than 2% the day the talks became public, part of a wider tech selloff that Friday.

A bigger number sits behind that headline figure. Amazon and Google have already committed roughly $76 billion combined to keep Anthropic’s compute loyal to them. Anthropic is negotiating around them anyway, and that tension, not the $10 billion itself, is the real story here.

The Terms Still on the Table

Start with what is actually known. Anthropic proposed the arrangement in June. Under the structure being discussed, Anthropic would pay Meta in monthly installments over two years, with both companies able to exit early if the economics stop working.

Meta does not currently run a business selling computing power to outside customers, which is part of why the talks have reportedly gotten complicated. Meta did not respond to a request for comment from Reuters, and Anthropic declined to comment. Shares of Meta pared earlier losses slightly once the report broke, before closing down more than 2% that day.

• What we know: Anthropic proposed the lease in June; the figure discussed is up to $10 billion over two years; payments would run monthly with early exit rights on both sides; Meta shares fell more than 2% the day the story broke.
• What’s unconfirmed: whether a final contract gets signed; which data centers or chips would be involved; whether Meta turns this into a standing cloud product; both companies’ official comment beyond a declined or absent response.

Size only makes sense next to what Anthropic has already locked up elsewhere. In May, it signed a deal to pay Elon Musk’s SpaceX roughly $45 billion across three years, about $1.25 billion a month, for the full output of SpaceX’s Colossus 1 data center in Memphis, Tennessee. Against that, a $10 billion, two year arrangement with Meta looks almost modest.

Amazon and Google Already Paid to Avoid This

This is where the story gets interesting for anyone who isn’t Meta or Anthropic. Amazon has invested up to $25 billion in Anthropic on top of the $8 billion it put in earlier, taking its cumulative stake to roughly $33 billion. In return, Anthropic committed to spend more than $100 billion on AWS technology over ten years and locked in up to 5 gigawatts of Trainium and Graviton chip capacity.

Google moved in the same window. Its own commitment reportedly reached roughly $40 billion on top of an earlier $3 billion stake, pushing its cumulative bet to around $43 billion and lifting Anthropic’s valuation to about $350 billion. That deal secured access to up to a million of Google’s TPUs, including its newer Ironwood chips.

Add it up and Amazon and Google have put close to $76 billion into Anthropic specifically to be its infrastructure backbone. Anthropic went to SpaceX two months later anyway, then opened talks with Meta the month after that.

None of that money bought Amazon or Google exclusivity. It bought priority, and priority is clearly not the same thing once a company is burning through capital fast enough to need a fourth supplier.

Why Does Anthropic Need Even More Compute?

Anthropic’s revenue is growing faster than almost any hardware buildout can track. Its annualized revenue run rate jumped from about $9 billion at the end of 2025 to more than $30 billion by early April 2026, and the company is widely described as heading toward a public listing. That growth has come with strain: reporting has pointed to GPU bottlenecks tight enough that Anthropic has capped some users’ access to its premium Claude tiers.

Anthropic has been open about the strategy behind spreading itself across so many vendors. The company has said it trains and runs Claude across AWS Trainium, Google TPUs and Nvidia GPUs, matching each workload to whichever chip suits it best rather than depending on one supplier’s roadmap.

• AWS Trainium and Graviton chips, under the ten year, $100 billion plus AWS commitment
• Google TPUs, including seventh generation Ironwood accelerators
• Nvidia GPUs at SpaceX’s Colossus 1 site in Memphis
• Potentially Meta’s own Nvidia clusters, if the lease talks close

A fourth supplier is not redundancy for its own sake. It is leverage. Every additional vendor Anthropic can credibly walk toward weakens the pricing power of the ones it already has contracts with.

Meta Hires the Competition’s Cloud Talent

Meta’s interest is not abstract either. At its May shareholder meeting, chief executive Mark Zuckerberg said entering cloud computing was definitely on the table, and that firms were approaching Meta almost every week about buying access to its AI models or spare capacity. He had made a similar point months earlier, saying companies regularly ask if we have compute they could buy from us at some premium.

Meta has also brought in a former senior AWS executive, Dave Brown, according to one industry report, a hire that fits a company trying to build cloud muscle it has never needed before.

The timing sits awkwardly against Meta’s other 2026 headline. The company cut about 8,000 jobs, roughly a tenth of its workforce, while closing another 6,000 open roles, at almost the same moment it raised its AI infrastructure spending target. Selling excess compute to a rival lab would help Meta explain that combination to investors who have already punished the stock once this year over the spending itself.

The Capex Math Behind Meta’s Offer

None of this happens without Meta’s own numbers. The company told investors in January it expected 2026 capital expenditure of $115 billion to $145 billion. Its Q1 2026 earnings call raised that to $125 billion to $145 billion, citing higher component costs and future year data center needs. That is nearly double the $72.2 billion Meta actually spent in 2025.

Meta can absorb that. The company posted $12.4 billion in free cash flow last quarter alone, well ahead of what analysts had expected. S&P Global’s research desk noted the quarter’s lighter than forecast capital spending reflected the timing of AI infrastructure investments rather than a strategic pullback.

Some of the pressure driving that guidance higher is showing up outside Meta’s own building plans too, with memory prices forecast to climb as much as 50% this quarter, squeezing every buyer of AI hardware at once. Meta’s own Superintelligence Labs still needs a growing share of whatever capacity Meta builds, which is exactly why a formal leasing business, rather than a one off favor to a friendly lab, is the more likely long term shape of this.

A New Landlord for CoreWeave and Nebius

Reuters framed the strategic logic plainly: a deal like this would let Meta diversify beyond advertising while competing directly with specialist GPU providers such as CoreWeave and Nebius. Those firms built their businesses renting Nvidia capacity to AI labs that don’t want to build their own data centers. A hyperscaler with Meta’s ad revenue cushion entering that market on the supply side changes the competitive math for everyone already in it.

Not every hyperscaler is moving the same direction. Microsoft has gone the other way, reportedly canceling data center leases as analysts at TD Cowen flagged possible oversupply. Chief executive Satya Nadella has acknowledged there will be an overbuild industry wide, even while saying Microsoft expects to be leasing a lot of capacity again in 2027 and 2028. Amazon, for its part, has faced its own investor pushback over the scale of its AI spending even as it writes checks to Anthropic.

Anthropic is still headed toward a public listing. Meta still has no cloud product to point to. And the $10 billion figure that both companies declined to confirm remains, for now, exactly that: a number in somebody else’s report.