Anthropic locks TPU capacity
Anthropic is locking in utility-scale TPU capacity with Google and Broadcom, planning about 3.5 gigawatts of TPU-based compute across new U.S. data centres to guarantee training headroom. This deal deepens Google’s role as a merchant supplier of TPUs and extends Broadcom’s TPU and networking supply relationship into a long-term commercial arrangement through 2031, underlining that model labs are buying multi-year capacity as much as raw silicon. (nationaltoday.com) (eenewseurope.com)
Anthropic just did the artificial intelligence version of reserving power plants years before flipping the switch: it signed for about 3.5 gigawatts of Google tensor processing unit capacity, with the new compute expected to start coming online in 2027. (anthropic.com) (cnbc.com) That number is so large that Broadcom’s filing described it in power terms, not chip counts, because the bottleneck now is electricity, buildings, cooling systems, and network gear as much as the chips themselves. (sec.gov) (datacenterdynamics.com) The three companies each play a different role in the stack. Google designs the tensor processing units, Broadcom helps develop and supply the tensor processing units and rack components, and Anthropic is the customer locking in the future capacity for Claude. (reuters.com) (anthropic.com) This was not a one-off purchase order. Broadcom said its long-term agreement with Google to develop and supply future generations of custom artificial intelligence chips, plus a supply assurance deal for networking and other rack components, runs through 2031. (sec.gov) (reuters.com) Anthropic is not waiting for all of that capacity to appear at once. Broadcom chief executive Hock Tan said Broadcom was already providing 1 gigawatt of Google tensor processing unit compute for Anthropic in 2026, and that demand for 2027 was expected to jump above 3 gigawatts. (cnbc.com) The company says the vast majority of the new compute will sit in the United States, which ties this deal to Anthropic’s November 2025 pledge to invest $50 billion in American computing infrastructure. (anthropic.com) The background here is that model labs used to talk like software companies and buy compute like renters. In 2026 they are acting more like utilities and airlines, signing multi-year capacity deals because the scarce thing is guaranteed access at the exact moment a new model needs to train. (sec.gov) (cnbc.com) Anthropic’s own numbers explain the urgency. On April 6 it said its annualized revenue run rate had passed $30 billion, up from about $9 billion at the end of 2025, and that more than 1,000 business customers were now spending over $1 million a year, double the count from February. (anthropic.com) (bloomberg.com) The deal also says something about Google. For years, Google’s tensor processing units were mostly seen as in-house weapons for Google services and Google Cloud, but this arrangement makes Google look more like a merchant supplier selling huge blocks of custom silicon capacity to outside model builders. (anthropic.com) (cnbc.com) It does not mean Anthropic is abandoning other hardware. Anthropic said it still trains and runs Claude across Amazon Web Services Trainium, Google tensor processing units, and Nvidia graphics processing units, and it said Amazon remains its primary cloud provider and training partner on Project Rainier. (anthropic.com) One quiet line in Broadcom’s filing keeps this from being a blank check: Anthropic’s use of the expanded capacity depends on its continued commercial success, and the companies are still discussing operational and financial partners for the rollout. (sec.gov)
