Meta is in early talks to rent out computing capacity to Anthropic under a two-year agreement that could be worth as much as $10 billion, The New York Times reported on July 20. If the talks close, Anthropic would gain access to server infrastructure owned by Meta and pay for it in monthly instalments, while both sides would keep the right to walk away early. Neither company has confirmed the discussions.
The reported figure is a ceiling, not a signed contract. The details that matter to any cloud buyer — the facilities, the chip models, the volume of capacity — have not been disclosed. What the report sketches is a shift in how the largest AI labs source the raw horsepower behind their models. For years Anthropic has paid hyperscalers and specialist operators for GPUs and accelerators. A Meta lease would add a new kind of supplier: a fellow AI company that built more data center capacity than it needs.
What the reported deal looks like
The arrangement described in the report is a capacity lease, not the launch of a public cloud. Anthropic would pay month to month and could terminate the pact ahead of schedule. That structure mirrors how Anthropic already runs some of its infrastructure. It agreed to pay SpaceX about $1.25 billion a month through May 2029 for access to the Colossus computing cluster, and it signed a 20-year data center lease with TeraWulf that is expected to generate roughly $19 billion in contracted revenue over its initial term.
Those numbers show how much compute Anthropic is willing to lock in to keep Claude and its other models running. The company trains and serves its models across several accelerator platforms, including AWS Trainium chips, Google TPUs, and Nvidia GPUs. It has also secured multiple gigawatts of Google and Broadcom TPU capacity that is scheduled to begin coming online in 2027. A Meta deal would sit alongside that mix rather than replace it.
The early-stage nature of the talks matters. The $10 billion label represents the maximum potential value, not a committed amount. Neither Meta nor Anthropic has said which facilities would be involved, what processors they would run, or how Anthropic's workloads would be managed. Until those points are settled, the story is a signal of intent more than a booked contract.
Anthropic's widening infrastructure web
Anthropic has moved fast to spread its infrastructure across suppliers. Its work with AWS includes long-term Trainium capacity and engineering collaboration with AWS's Annapurna Labs on chip optimisation and the Neuron software stack. With Google it has reserved TPU capacity at a scale measured in gigawatts. With SpaceX and TeraWulf it has taken on long leases and steady monthly bills.
That web matters because model training and inference are bounded by physical capacity. A lab that controls more compute can ship larger models and serve more users. A lease from Meta would let Anthropic add capacity without building its own campuses or waiting for a hyperscaler to expand its footprint.
The talks also signal that compute has become a tradable commodity among AI developers. Labs that once competed only on models now trade the underlying hardware capacity. This is a change from the early days of the AI boom, when most startups simply rented from AWS, Azure, or Google Cloud and treated the bill as a cost of doing business.
Meta's data centers, now for hire
Meta has spent heavily to build its own infrastructure. The company expects capital expenditure of between $125 billion and $145 billion in 2026, a jump from roughly $72 billion in 2025, driven largely by data centers and AI hardware. Its global estate runs on CPUs, Nvidia GPUs, and its own MTIA accelerators. It is also expanding its Richland Parish campus in Louisiana toward 5 gigawatts of capacity.
Chief executive Mark Zuckerberg told shareholders in May that other companies regularly ask Meta to sell them computing capacity at a premium, and that entering the cloud-computing market was under review. Meta could serve outside customers, he said, once it builds more infrastructure than it needs internally. A report from The Wall Street Journal described an internal effort called Meta Compute, though Meta has not announced a public cloud service under that name.
Meta is also hiring from the hyperscaler playbook. Dave Brown, a senior Amazon Web Services executive who spent nearly two decades at Amazon, is joining to work on data center expansion under infrastructure chief Santosh Janardhan. His hire suggests Meta is building the operational muscle a cloud provider needs: procurement, capacity planning, and customer-facing engineering.
The reported Anthropic talks would put Meta on both sides of the infrastructure market at once. It already buys capacity from specialists such as CoreWeave and Nebius while it builds its own sites. Leasing some of its own spare capacity to another AI lab would make it both a buyer and a seller of compute.
A cloud provider in all but name
A completed deal would not turn Meta into a hyperscaler overnight. Established clouds bundle processors with storage, databases, networking, identity controls, security tools, billing, and support. Meta has not announced that kind of external portfolio. Reuters noted that a commercial compute business would place Meta in competition with specialists such as CoreWeave and Nebius, while AWS, Microsoft Azure, and Google Cloud offer far broader catalogs.
The talks leave open questions that neither company has answered. It is unclear whether Meta would hand Anthropic dedicated clusters, management software, or bare capacity. There is no word on pricing per unit of compute, expected utilisation, or service terms. Without those details, the arrangement cannot be lined up against enterprise offers from the big three clouds.
The power question behind the capacity
Capacity on paper depends on electrons in the wall. Data center builds across the United States are running into power constraints, and states are starting to write rules around them. New York recently became the first state to pause permits for large hyperscale data centers while it reviews power and siting policy, a move we covered in New York hits pause on hyperscale data centers. S&P Global's July 2026 legislation roundup tracks similar measures moving from local zoning fights toward federal debate.
That context shapes any Meta capacity plan. A lease to Anthropic is only worth something if the power is there to run the racks. Meta's own Richland Parish expansion and its spending guidance show it is betting that the power will be secured. Competitors such as OpenAI have struck their own multibillion-dollar compute agreements with Oracle and CoreWeave, a sign that the scramble for both chips and energy is now the central contest in the AI business.
For readers tracking the category, our Cloud & Edge Computing desk has followed how inference is pulling compute toward metro and edge sites in AI inference pulls compute back to the edge as metro data centers surge.
What comes next
None of this is settled. The talks are preliminary, the $10 billion is a maximum rather than a signed figure, and either side can exit early. What is clear is the direction: the largest AI developers are turning their data centers into assets they can monetise, not just cost centers they fill to train the next model. For a sector that once treated cloud as a utility to rent, that is a notable turn.
More reporting is available from Cloud Computing News, which first surfaced the terms, and from Reuters' technology desk, which covered the competitive angle.