Meta to Lease Compute Power to Anthropic in ~$10B Deal

The proposed arrangement is structurally simple but strategically odd: Meta would rent raw computing capacity from its own data centers to Anthropic in a deal reportedly worth up to $10 billion over two years, with Anthropic paying in monthly installments and both sides free to walk away early ^[1]. Notably, Anthropic proposed the deal to Meta in June 2026, not the other way around, and the talks reportedly became complicated precisely because Meta has no existing business for selling compute to outside customers ^[1]. This is a company improvising a wholesale product on the fly, in response to inbound demand rather than a mature go-to-market plan.

What makes it genuinely unusual is that Meta's Llama models compete directly with Anthropic's Claude, so Meta would simultaneously be a frontier AI rival and Anthropic's infrastructure landlord. That raises live questions about conflicts of interest and about how customer workloads and intellectual property would be walled off from Meta's own research ^[1]. Datacenter Dynamics frames this cross-competitor leasing as increasingly routine in an industry where the compute shortage now outweighs old rivalries ^[7].

For Anthropic, the Meta talks are one more layer in an aggressive compute-accumulation strategy driven by scarcity. The lab already holds a roughly $100 billion, ten-year partnership with AWS, and in May 2026 signed a separate ~$45 billion, three-year deal with SpaceX for the full capacity of the Colossus 1 data center in Memphis, paying about $1.25 billion per month through 2029 ^[3]. Against those commitments, a $10 billion, two-year Meta lease is comparatively small - but it signals that Anthropic still cannot source enough accelerators from any single provider to train its frontier models ^[3].

The annualized economics reveal how differently these deals are structured. Meta's rumored terms pencil out to roughly $5 billion per year, versus about $10 billion per year for the AWS partnership and around $15 billion per year for the SpaceX arrangement ^[3]. Anthropic is effectively building a diversified supply portfolio across a hyperscaler, a rocket company, and now a social-media giant, betting that compute demand will keep outrunning supply industry-wide ^[4]. That hunger is the demand-side engine making Meta's whole leasing thesis viable in the first place.

On Meta's side, the deal is inseparable from its capital spending problem. Meta raised its 2026 AI capex guidance to $125-145 billion, and Zuckerberg has repeatedly said that if the company concludes it has overbuilt, renting out spare capacity is an option it holds ^[6]. The talks with Anthropic are paired with a concrete organizational build-out: Meta is hiring departing AWS SVP Dave Brown, who spent nearly 19 years building AWS compute and machine-learning services, to report to infrastructure chief Santosh Janardhan and stand up a 'Meta Compute' offering ^[5]. Poaching a career AWS infrastructure executive is a strong signal that Meta intends to compete with the neoclouds and hyperscalers, not just offload idle GPUs.

The upside math is why analysts are paying attention. SemiAnalysis argues that allocating even 200MW of compute to an external customer at premium rates could drive on the order of $10 billion a year in high-margin revenue ^[4]. Morgan Stanley has estimated that leasing 250MW at $40 per watt could add roughly $2.97 to Meta's 2028 EPS - about 8% upside - with internal infrastructure utilization pegged near 65%, leaving meaningful slack to monetize ^[2]. In that frame, a compute-leasing arm could re-rate Meta's narrative from 'AI overspender' to 'AI infrastructure cash-flow machine' ^[2].

The report split the market cleanly. Jefferies' Brent Thill argues that fears of Meta overbuilding are premature since demand still outstrips supply, so a leasing business simply raises utilization and cash flow ^[2]. D.A. Davidson's Gil Luria takes the opposite view, reading the move as evidence Meta is deprioritizing its own frontier work in favor of near-term monetization while its superintelligence lab lags rivals ^[2]. Meta's stock captured that ambivalence, sliding more than 5% intraday before recovering to close down roughly 2-3% ^[2].

Outside Wall Street, the reaction skews more skeptical. Retail-investor forums largely read the deal as bearish evidence of an AI infrastructure bubble - the story that Meta overbought compute, could not find enough internal demand, and is now recouping capex by renting to one of the few labs still buying at scale. That 'circular revenue' worry, where a handful of AI labs appear to be the main customers propping up infrastructure spend, dominated the loudest threads. More neutral voices pushed back, comparing the arrangement to a standard hyperscaler playbook of monetizing spare capacity - AWS's own origin story - and noting that a real technology boom and a financial bubble can coexist, much like 19th-century railroad overbuilding.