Anthropic rents CoreWeave
Anthropic agreed to rent AI compute capacity from CoreWeave to power Claude, a sign that model providers are increasingly contracting specialist infrastructure instead of relying solely on hyperscalers. The deal helped fuel a market rally for CoreWeave and highlights how compute capacity is being financed and contracted more like utility infrastructure. (bloomberg.com) (markets.financialcontent.com)
Anthropic didn’t buy a giant new data center this week. It signed a multiyear deal to rent computing capacity from CoreWeave so it can build and run Claude, with the first capacity scheduled to come online later in 2026. (coreweave.com) (reuters.com) That sounds mundane until you remember what is being rented. Training and serving a model like Claude means securing huge blocks of Nvidia chips inside power-hungry data centers, then keeping those chips fed with networking, cooling, and electricity around the clock. (bloomberg.com) (sec.gov) CoreWeave is not Amazon Web Services or Microsoft Azure. It is a specialist cloud company built around graphics processing units, which are the chips used for most modern artificial intelligence training and inference jobs. (sec.gov) (investors.coreweave.com) Anthropic is also not betting on one supplier. On April 6, 2026, it said it had expanded its partnership with Google and Broadcom for multiple gigawatts of next-generation Tensor Processing Unit capacity starting in 2027, so the CoreWeave deal sits alongside a broader effort to lock up compute wherever it can find it. (anthropic.com) (googlecloudpresscorner.com) That is the shift hiding inside the headline. Model companies used to talk as if cloud computing were something you could buy on demand like hotel rooms, but frontier artificial intelligence now needs reserved capacity years ahead, more like leasing factory space or power from a utility. (anthropic.com) (coreweave.com) Investors immediately read the Anthropic contract as proof that CoreWeave is becoming that kind of supplier. CoreWeave shares jumped more than 10% on April 10 after the announcement, according to CNBC and SiliconANGLE, following another huge deal disclosed a day earlier with Meta. (cnbc.com) (siliconangle.com) (reuters.com) That Meta deal was enormous on its own. Reuters reported on April 9 that Meta expanded its CoreWeave agreement by $21 billion for cloud computing capacity, pushing the relationship further out to 2032. (reuters.com) The market likes these contracts because CoreWeave has been trying to solve a very old infrastructure problem with a very new product: borrow heavily upfront, build capacity fast, then fill it with long-term customers before the interest bill catches up. Its March 2025 stock-market filing showed revenue exploding from $16 million in 2022 to $229 million in 2023 and $1.9 billion in 2024. (sec.gov) By February 26, 2026, CoreWeave said it had reached more than $5 billion in annual revenue for 2025 and had record backlog, which is the pile of signed business it expects to deliver later. That backlog is what makes lenders and equity investors treat future chip time almost like contracted utility output. (investors.coreweave.com) (markets.financialcontent.com) There is a second reason this deal landed so well. CoreWeave’s 2025 filing showed heavy customer concentration, with Microsoft accounting for 62% of 2024 revenue, so every big contract from Anthropic or Meta makes the business look less like a one-customer trade and more like a broader market for rented artificial intelligence capacity. (sec.gov) (markets.financialcontent.com) So the real story is not that Anthropic found another vendor. It is that the biggest artificial intelligence labs are starting to secure chips, power, and data-center space through a patchwork of long-dated contracts, and companies like CoreWeave are trying to turn that scramble into a new kind of utility business built around rented compute. (anthropic.com) (coreweave.com)