Anthropic has agreed to spend $35 billion on computing capacity from Nvidia-backed Lambda, the latest in a series of enormous data-center deals that are expanding the physical infrastructure behind AI – even as regulators question how much proposed demand will actually materialize.
The agreement covers roughly 350 megawatts of capacity at a Texas data center being developed by bitcoin miner Hut 8 in Nueces County, according to Reuters.
The arrangement has an unusual structure. Nvidia is an investor in Lambda and will supply the chips used in the facility. The Wall Street Journal reported that Nvidia will also hold the lease on the data center, while Lambda will provide the resulting computing capacity to Anthropic.
The agreement follows a $45 billion, six-year deal Anthropic struck last week with another Nvidia-backed cloud company, Nscale, for computing capacity at a planned West Virginia data-center campus. Beginning in late 2027, Anthropic is expected to use 460 megawatts of Nvidia Vera Rubin processors there.
Together, the two agreements represent $80 billion in recently reported computing commitments from Anthropic alone.
But the data-center rush is beginning to collide with constraints on the electrical grid – and questions about whether every proposed project is supported by real or “ghost demand.”
Texas has paused new data-center grid connections pending an audit while regulators audit proposed projects. Requests from data centers and other large users to connect to the Texas grid have soared from about 48 gigawatts in 2023 to more than 474 gigawatts, Reuters reported.
Across parts of the Midwest, Mid-Atlantic and South, requests from very large electricity users, mostly data centers, now exceed 700 gigawatts – more than 10 times estimates of current U.S. data-center power consumption.
Regulators suspect some requests are duplicates or involve speculative projects that lack customers or financing. Utilities have already reduced projected demand after requiring developers to make upfront financial commitments.