Anthropic weighs possible multi-year compute lease from Meta in proposed $10 billion arrangement

Anthropic, the AI startup behind the Claude models, is discussing a potentially large arrangement to lease computing power from Meta, in a deal described as worth about $10 billion. The discussion was first proposed in June, and the current negotiations are described as early-stage, with no certainty that the parties will reach a final agreement.

The reported structure is framed around Anthropic obtaining access to Meta’s computing resources rather than buying hardware directly. In practical terms, a compute lease typically means an AI company pays for access to specialized data-center capacity, such as GPUs used to train and run large machine-learning models, under terms that can span months or years.

For Meta, the appeal of such deals is straightforward: AI companies with fast-growing training and deployment needs can provide demand for data-center capacity while potentially improving utilization and revenue per machine. For Anthropic, leasing capacity can reduce the time and capital required to assemble and operate large-scale infrastructure, especially when model development schedules are tight.

The report characterizes the talks as not yet decisive, meaning key terms are still likely unresolved. That includes, based on what is publicly described, the eventual duration of any agreement, how compute would be metered, and what performance or availability commitments would apply.

This kind of compute procurement is increasingly common in the race to build and deploy frontier AI systems. Large model labs often balance two priorities: securing enough high-performance chips to keep training and scaling efforts moving, and doing so at predictable cost. Hardware acquisition and capacity expansion can be slow, while access to already-operating data-center clusters may offer a faster path to scale.

Meta’s role in these dynamics reflects its position as a major operator of AI-focused infrastructure. Through its own AI research and product work, the company has invested heavily in data-center capabilities, and it has also faced an industry-wide question: how quickly can compute capacity be converted into reliable output for customers or partners beyond its internal use.

Still, the latest report does not provide enough detail to say what, if any, agreement is likely. It does not disclose which Meta systems or regions would be included, the expected timeline for contract signing, or the commercial terms that would define pricing and service levels. Until those points are clarified in a final contract or a more specific filing, the financial scale described in the report should be treated as a proposal, not an outcome.

What to watch next is whether the talks progress to a confirmed commercial arrangement, and whether Meta or Anthropic provides additional detail on scope, duration, and performance targets. Another key indicator will be whether compute leasing becomes a broader pattern in Meta’s partnerships, particularly if multiple AI labs pursue similar access models as they scale model training and inference workloads.