The deal locks in up to 401 megawatts of dedicated AI capacity at the Hawesville, Kentucky campus, with operations not expected to begin until late 2027.
Anthropic has signed a 20-year lease with TeraWulf for up to 401 megawatts of dedicated AI compute at the operator's Hawesville, Kentucky campus. The contract is projected to deliver roughly $19 billion in cumulative revenue to TeraWulf, with initial operations expected in late 2027 and full capacity in early 2028.
A 20-year term is what makes this different from Anthropic's typical cloud rental. The lab's normal relationship with AWS, GCP, or Microsoft Azure lets it add or remove capacity within a quarter; a dedicated lease locks in power, cooling, and floor space for the full life of the deal. The structural bet underneath the headline is that owning long-dated, power-anchored infrastructure will be a durable advantage as AI demand compounds.
TeraWulf is a former bitcoin miner that has spent the past two years repositioning itself as an AI data center operator. Its new "Justified Data Campus" at Hawesville puts the deployment in the band of what the three hyperscalers typically reserve for their biggest internal customers. The announcement lands inside a broader category forming around power-anchored AI data centers, sometimes called "neoclouds," which sell dedicated GPU capacity to frontier AI labs. Shares of TeraWulf jumped as much as 18% intraday on the news, and peer operators CoreWeave, IREN, Hut 8, and Riot Platforms all traded higher. The market is reading the announcement as a category vote, not a single-customer deal.
That $19 billion is worth parsing carefully. The figure is projected cumulative revenue to TeraWulf over the contract, not Anthropic's cash outlay. The actual structure of the lease, including base rent, escalators, and how GPU infrastructure costs flow through, will determine what Anthropic pays year by year, against a demand curve that, to date, has only ever moved upward.
That demand curve is where the deal's real risk sits. A 20-year AI compute commitment prices in sustained training demand at the frontier and continued growth in inference workloads. The first is genuinely contested. Open-weight models from DeepSeek, Alibaba's Qwen team, and Meta's Llama line are closing the gap on frontier closed-weight performance, which could compress the compute required to train the next generation of leading systems. The second depends on whether inference becomes the dominant share of total AI compute, a question no outside observer can resolve. A lease that does not mature until 2046 bets that AI demand grows for longer than open-weight competition or efficiency gains can compress it.
TeraWulf's same-day announcement also disclosed the sale of the company's majority interest in a roughly 168-megawatt Texas AI data center joint venture to Fluidstack, a deliberate recycling of capital into wholly owned projects like Hawesville, where the Anthropic rent flows directly to the operator rather than being split with a partner.
Watch the late-2027 power-on date. Closing conditions, GPU procurement at scale, and the Kentucky grid's ability to deliver 401 megawatts continuously will determine whether the lease converts to revenue on schedule, and whether the rest of the neocloud sector's valuation holds against it.