Meta-Anthropic Compute Talks Put a $10 Billion Ceiling on an Unproven Cloud Deal

• Anthropic proposed paying Meta up to $10 billion over two years for AI computing capacity, but either company could leave early and the talks may not produce an agreement.
• No chip count, power capacity, delivery schedule or unit price has been reported, making the headline value a poor measure of the proposed deal's economics.
• Meta would be selling capacity while committing hundreds of billions of dollars to infrastructure and outside cloud services; Anthropic, meanwhile, already has several other large capacity channels.

Anthropic's proposal could give Meta a first major customer for a cloud operation. It does not yet establish that Meta has a cloud business—or that the two companies have found a durable way to price and deliver the capacity.

The $10 billion figure is a ceiling, not a contract

Anthropic proposed the arrangement in June, and Meta is considering it, according to an account based on three people with knowledge of the discussions. The proposal calls for monthly payments over two years, with a total value of as much as $10 billion. Either company could opt out early.

The qualifications are more informative than the maximum value. The terms remain in flux, the discussions are at an early stage and the companies may not reach an agreement. Anthropic and Meta declined to comment. A separately sourced account of the talks said they were complicated by Meta not yet having a business that sells computing power. Another person familiar with the matter described the discussions as very preliminary.

The available reporting does not identify the number or type of chips, the power allocation, the locations, when capacity would become available or whether payments would be tied to reserved or consumed compute. Without those terms, $10 billion cannot be translated into a unit price or compared reliably with another infrastructure contract.

That caveat matters for the most obvious comparison. Anthropic's May agreement with SpaceX was reported at $45 billion over three years and also allows either side to exit early. But Anthropic specified in its announcement that it would receive all capacity at the Colossus 1 data center within a month: more than 300 megawatts and over 220,000 Nvidia GPUs. No equivalent capacity measure has been reported for the Meta proposal. Comparing the contracts by dollars per month would therefore compare payment schedules subject to early exit without a common capacity denominator.

Meta would be a cloud customer and a compute seller

Meta has publicly discussed the possibility of selling infrastructure. Mark Zuckerberg said at the company's May shareholder meeting that entering cloud computing was “definitely on the table” and that companies approached Meta almost every week about buying access to its models or spare compute.

There is a clear diversification case. In the first quarter of 2026, Meta generated $55.02 billion of its $56.31 billion in revenue from advertising—about 98%, calculated from its quarterly filing. A compute customer could add a new revenue stream to a company still overwhelmingly financed by ads.

But the same filing makes “spare compute” an incomplete description of Meta's position. The company forecasts $125 billion to $145 billion of capital expenditures in 2026 to support AI and its core business. It spent $19.84 billion in the first quarter, including finance-lease principal payments, and said higher infrastructure costs were a major contributor to rising expenses.

As of March 31, Meta also had $237.67 billion of non-cancelable contractual commitments, mostly for third-party cloud capacity, servers, network infrastructure, data centers and Reality Labs hardware. Of that total, $42.25 billion was due in 2026 and $47.65 billion in 2027. Separately, it disclosed $182.88 billion of leases that had not yet begun, covering data centers, colocation and some network infrastructure with start dates extending from the remainder of 2026 through 2036. Those figures describe different obligation categories and should not be added together.

Meta is already a major buyer from a would-be rival to any new cloud operation. In March, it committed about $21 billion under an existing CoreWeave agreement for reserved computing capacity extending into 2032, subject to delivery and availability requirements, according to CoreWeave's filing. That agreement permits termination for cause; the proposed Anthropic arrangement reportedly gives either company broader rights to leave early.

These disclosures do not prove that Meta has overbuilt, nor does its CoreWeave purchase prove that it lacks sellable capacity. Chips, power, locations, delivery dates and contract restrictions can make one pool of compute available while another remains scarce. They do show why gross contract value is insufficient: Meta would need revenue from outside customers to cover power, networking, depreciation, support and the cost of capacity it owns or rents.

Anthropic already has many ways to buy capacity

For Anthropic, Meta would be an additional supplier rather than the foundation of its infrastructure strategy. The company says it trains and runs Claude across AWS Trainium, Google TPUs and Nvidia GPUs. In the same company announcement, it listed an agreement with Amazon for up to five gigawatts, including nearly one gigawatt of new capacity by the end of 2026; a five-gigawatt Google and Broadcom agreement beginning to come online in 2027; a Microsoft and Nvidia partnership that includes $30 billion of Azure capacity; and a $50 billion U.S. infrastructure investment with Fluidstack.

Those arrangements differ in timing, hardware, geography and financial structure, so their headline numbers are not like-for-like measures. Collectively, however, they limit the significance of Meta as a single source. Anthropic is assembling a portfolio across hyperscalers, chip platforms and newer infrastructure operators, not replacing one exclusive provider with another.

Anthropic presented the SpaceX capacity alongside immediate product changes: doubled five-hour Claude Code rate limits for several paid plans, removal of peak-hour reductions for Pro and Max accounts, and higher API rate limits for Claude Opus models. The company attributed those increases to the SpaceX partnership together with other recent compute deals, not to one provider alone.

Meta would also enter a crowded market. The independently sourced account identified CoreWeave and Nebius as potential neocloud competitors, while Anthropic's own supplier list includes the established AWS, Google Cloud and Azure platforms. One large, flexible lease could demonstrate demand for Meta's capacity, but it would not show that Meta can match those providers' delivery, service and economics for a broader customer base.

The next test is repeatability, not the headline value

A signed agreement would be the first meaningful threshold. To evaluate it, the missing evidence would include the capacity and hardware Meta promises, the delivery dates and locations, the workloads Meta will support, the basis for monthly charges, service guarantees and the conditions that permit either party to exit.

The operating test would come later. Meta would need to show that serving Anthropic does not constrain its own AI work and that revenue exceeds the full cost of delivering the service. Anthropic would need to show that the arrangement adds sustained usable capacity rather than another cancellable reservation.

The decisive evidence of a cloud business would be a second customer and repeatable terms. For now, the reporting establishes only that Anthropic made a proposal and Meta is considering it. The capacity, commitment and economics needed to turn those talks into a business remain undisclosed.