The New York Times reported on July 17, 2026, that Meta has entered early talks to lend compute resources to Anthropic. The deal could be worth up to $10 billion over two years, and Anthropic reportedly made the proposal in June. Reuters also confirmed the key terms through its own sources. However, neither company has announced the deal, and there remains a possibility that the talks will fall through.
More significant than the $10 billion figure itself is whether Meta—whose main revenue comes from Facebook and Instagram advertising—can sell externally the AI infrastructure it built for its own use. Anthropic already uses multiple clouds and semiconductor providers in combination, yet is still seeking additional compute resources. If the deal goes through, Anthropic's demand would convert Meta's facilities into external revenue.
The Reported $10 Billion Is Not a Finalized Contract Value
The reported $10 billion represents an upper limit on what could be paid over two years. According to Reuters, Anthropic would pay Meta monthly, and the terms could still change. Both companies are also said to have provisions allowing early termination of the contract. Meta did not immediately respond to the news agency's inquiry, and Anthropic declined to comment.
It is therefore premature to call Anthropic a Meta customer. The talks are at an early stage, and Meta does not currently operate a business selling compute resources. The data centers to be used for supply, the scale of power capacity, the type and number of semiconductors, and the start date of usage have not been reported either. There is no basis for treating the $10 billion as revenue or as a backlog figure.
At the shareholder meeting on May 27, Mark Zuckerbergraised entering the cloud business as an option. According to remarks reported by Reuters, companies seeking API access to Meta's models or wanting to purchase surplus compute have been approaching Meta "almost every week." However, the reason Meta was not yet selling at that time was that it expected to use its planned compute internally. Zuckerberg's explanation was conditional—selling would happen only if Meta ended up overbuilding in the future.
The current talks move this option one step closer to an actual deal. Even so, it cannot yet be said that Meta has accumulated facilities with no current use. If the deal is finalized, Meta would be able to consider both internal demand and external customers as destinations for its compute resources.
An Annual Cap of $5 Billion Against $125–145 Billion in Capital Expenditure
In 2025, Meta spent $72.22 billion on capital expenditures, including principal repayments on finance leases. The projected amount for 2026 is $125–145 billion. The upper end is roughly double the previous year's actual spending. In the first quarter of 2026 alone, Meta invested $19.8 billion to expand servers, data centers, and networking.
Assuming the $10 billion is received evenly over two years, the annual cap would be $5 billion. This corresponds to just 3.4–4.0% of the planned 2026 capital expenditure. Securing a major customer could demonstrate demand for external sales, but a single deal is not large enough to broadly cover construction costs. Moreover, actual payments could fall below the cap, and there is no guarantee they would be distributed evenly.
Meta's longer-term obligations are even larger. As of the end of March, non-cancelable contractual commitments totaled $237.67 billion, much of which consisted of contracts for third-party cloud capacity and equipment ranging from servers to data centers. Of this, $42.25 billion is due in 2026 and $47.65 billion in 2027. Lease obligations for data centers and colocation facilities not yet operational added another $182.88 billion. Because these figures differ in nature, they cannot simply be added together.
While building its own facilities, Meta is also purchasing external cloud capacity. As of the end of March, it had a conditional obligation to purchase up to $147.2 billion worth of cloud capacity over five years, under a contract that would reduce the amount if the supplier could resell the capacity to other companies. The increase in first-quarter infrastructure costs was also attributed to a combination of depreciation, data center operating costs, and spending on third-party cloud.
This purchasing approach reflects the reality that compute supply and demand cannot currently be treated as fixed. Meta itself has warned in its 10-Q that facilities exceeding actual demand could lead to impairments and negatively affect its financial condition. If the sale to Anthropic is finalized, Meta would gain an outlet to reduce the risk of overinvestment without cutting back on capital expenditure.
Three Compute Foundations Supporting Anthropic, and Meta's Blank Fields
In April 2026, Anthropic disclosed that its annualized revenue had surpassed $30 billion, a sharp increase from roughly $9 billion at the end of 2025. According to the company, this surge in usage strained the reliability and performance of offerings ranging from the free tier to Team plans at peak times. Contracts for compute resources serve both to prepare for future model training and to procure the stability needed for immediate services like Claude Code and the API.
| Supplier | Announced Scale | Timing/Conditions of Operation |
|---|---|---|
| Amazon | Up to 5GW, over $100 billion over 10 years | Nearly 1GW combined of Trainium2/3 by end of 2026. Anthropic uses over 1 million Trainium2 chips |
| Google/Broadcom | 5GW | Next-generation TPUs operational from 2027 |
| SpaceX | Over 300MW, over 220,000 NVIDIA GPUs | Full capacity of Colossus 1 utilized within one month. Reflected in raised usage limits for Claude Code and the API |
| Meta | Up to $10 billion over 2 years | Early talks. Power, semiconductors, locations, and start date undisclosed |
What this comparison shows is that Anthropic is not consolidating around a single cloud. The company distributes workloads across AWS Trainium, Google TPU, and NVIDIA GPUs. While Amazon remains its primary training and cloud provider, SpaceX's capacity was also used immediately after contracting to raise usage limits. Even while waiting for Google's 5GW to come online in 2027, demand on the product side continues to grow.
If the deal is finalized, Anthropic's sources of supply would increase further. However, equal figures in MW or GW do not necessarily mean equivalent compute resources. Depending on whether the workload is training or inference, which semiconductors are used, and how networking and storage are configured, the work that can be processed and the associated costs vary even at the same power scale. With the technical terms of Meta's proposal still blank, it is impossible to evaluate when and for what processing Anthropic could actually use it.
Can Supply Terms Be Finalized Before a Cloud Business Exists?
Reuters reported that the talks are complicated by the fact that Meta has no existing business selling compute resources. Owning facilities is not the same as providing a service that customers can rely on continuously. The contract would need to determine uptime, compensation for outages, separation of data and customer environments, and the scope of support. Pricing structures and handling of early termination are also essential.
The form of provision remains undecided as well. Whether Meta would lend raw compute capacity—combining semiconductors and networking—or operate Anthropic's workloads on its own infrastructure would determine the organizational structure and profitability required. A cloud offered broadly to developers and a business wholesaling capacity to select large customers are also different things. This single deal alone is not enough to conclude that Meta is moving toward becoming a comprehensive cloud provider on par with AWS, Google Cloud, or Microsoft Azure.
Even so, if the deal is finalized, Meta would gain a second avenue for recouping its massive AI investments. Anthropic, while keeping AWS as its primary provider, would be able to choose Meta as well depending on the use case. What should be verified next is not the headline figure of $10 billion, but rather the minimum purchase amount, the supply start date, the MW or number of semiconductors involved, and which party bears operational responsibility. Only once these details are disclosed can it be judged that the talks have progressed into a sustainable compute business.