Anthropic signed a $35 billion cloud computing agreement with Lambda, the Nvidia-backed infrastructure provider, tied to a roughly 350 megawatt artificial intelligence data center in Texas. The deal represents the largest single-vendor compute commitment disclosed by a foundation model company and marks the first time a tier-one AI lab has anchored its training infrastructure to a non-hyperscaler provider at datacenter scale.
Lambda will build and operate the facility. Anthropic secures dedicated capacity equivalent to roughly 12,000 H100-class GPUs at sustained load, based on standard rack density and power efficiency curves for liquid-cooled tensor clusters. The Texas site is expected to begin phased deployment in Q4 2025, with full capacity online by mid-2026. Lambda has not disclosed whether the deal includes equity warrants or revenue-share provisions, but the structure appears to be a long-term capacity reservation rather than a pay-as-you-go arrangement. Anthropic declined to comment on contractual terms.
The move changes the leverage map for frontier labs. Anthropic's compute budget through 2024 ran almost entirely on Google Cloud and Amazon Web Services, with estimated annualized spending near $2.8 billion based on disclosed training runs for Claude 3.5 Sonnet. By locking Lambda at this scale, Anthropic trades spot-market flexibility for cost certainty and removes the structural risk that a hyperscaler could throttle access or reprice capacity during a model launch window. Lambda gains a creditworthy anchor tenant and can now raise project finance against a $35 billion multi-year contract, which dramatically lowers its cost of capital for the Texas build. Nvidia benefits indirectly: Lambda's architecture is GPU-native, and the deal effectively pre-commits a large share of future Blackwell and Rubin shipments.
For allocators, the second-order effects matter more than the headline number. Anthropic is signaling that inference economics at scale require owned or dedicated infrastructure, not rented cloud. The company's inference margin on Claude has been under pressure since late 2024, when enterprise customers began negotiating volume discounts that compressed per-token pricing by an estimated 40-50% year-over-year. A dedicated facility allows Anthropic to amortize capex across training and inference workloads, potentially improving gross margin by 15-20 percentage points relative to hyperscaler pricing for equivalent sustained throughput. This also suggests Anthropic expects its inference demand to remain predictable and large enough to justify the capital lock-in, which is a bet that Claude retains enterprise share against OpenAI and DeepSeek through at least 2028.
Watch three follow-on events. First, whether OpenAI or Google DeepMind announce similar dedicated deals with CoreWeave, Crusoe, or other GPU-specialist providers in the next six to nine months—if they do, the hyperscaler oligopoly in AI compute is breaking faster than the public cloud thesis assumes. Second, whether Lambda's cost of debt for the Texas project comes in below 6.5%, which would confirm that project finance markets now view long-term AI compute contracts as investment-grade equivalent. Third, Anthropic's next fundraising round, likely in late 2025 or early 2026, where the Lambda deal will either be framed as disciplined capital allocation or as a $35 billion liability that constrains optionality.
Lambda's equity value just became measurable. The company was last privately valued near $1.5 billion in 2023. A $35 billion contract with a creditworthy counterparty and Nvidia hardware allocation rights makes Lambda a plausible IPO candidate by 2026, particularly if it can sign one or two more anchor tenants at even a fraction of the Anthropic scale.