Anthropic's $45B Nscale deal: your capacity arrives in 2027

AI compute capacity is now bought the way utilities buy transmission: years ahead, in gigawatt-scale blocks, at prices that dwarf the revenue they support. Bloomberg reported on August 26 that Anthropic agreed to pay Nscale roughly $45 billion over six years for capacity at Nscale's West Virginia campus. The number to underline is not the $45 billion. It is the date the machines switch on: late 2027.

What actually happened

Per Bloomberg's reporting, the deal covers about 460 megawatts — a first building in a campus planned to reach roughly 1.35 gigawatts, with an on-site power plant and an estimated $71 billion total build cost. Nscale will run Nvidia's Vera Rubin systems, which begin coming online late next year. TechCrunch's writeup puts it in sequence: Anthropic has stacked up something on the order of $60 billion in compute commitments in eight months, alongside deals with Volta, AMD, Amazon, and Google.

Two details give the deal its shape. Microsoft held a letter of intent on that West Virginia site and walked away over the summer — Anthropic took the slot. And Nscale, a two-year-old London company that bought Anyscale in July, is preparing a US listing as soon as September. A $45 billion anchor tenant is a useful thing to have in an S-1.

Why it matters for your business

Read the timeline as a forecast, because that is what it is. Anthropic just told the market it expects to need 460 megawatts it does not currently have, and that the fix does not arrive until late 2027. Nobody signs a six-year, $45 billion contract for capacity they think will be cheap and abundant next spring.

So price your next eighteen months against tight supply, not falling costs. Concretely: if your margin only works at a token price that has to drop, that is not a plan. Don't build a workflow whose economics depend on frontier inference getting 60% cheaper by mid-2027 — build it so the expensive model handles the step that actually needs it and everything else runs on something you can host yourself. Keep contracts short. Instrument per-workflow token cost now so a repricing shows up as a line item instead of a surprise.

The other half is counterparty risk. Your AI vendor's capacity is increasingly rented from two-year-old companies financing data centers against contracts they haven't delivered on yet. That is not a reason to panic. It is a reason to make sure the prompt, the eval set, and the routing logic live in your repository — so swapping the model behind them is a config change and a weekend, not a rebuild.

Sources: Bloomberg, TechCrunch.