Anthropic's big compute bet

Anthropic just did something that looks less like a software contract and more like booking a private power plant. On April 6, 2026, the company said it signed a deal with Google and Broadcom for multiple gigawatts of next-generation Tensor Processing Unit capacity, with service expected to start in 2027. (anthropic.com) Anthropic did not describe the agreement as a normal cloud rental. It described it as a long-term expansion of compute infrastructure for Claude, its family of artificial intelligence models, at a scale usually discussed in electricity markets, not app subscriptions. (anthropic.com) The reported figure attached to the deal is 3.5 gigawatts. A gigawatt is a unit of power large enough that utilities use it to describe city-scale demand, which tells you this is not about buying a few more servers for a busy quarter. (thenextweb.com) The chips at the center of the agreement are Tensor Processing Units, which are Google’s in-house chips built for training and running artificial intelligence models. Instead of relying only on general-purpose graphics chips from outside suppliers, Google can pair its own chips with its own cloud and sell that full stack to a company like Anthropic. (anthropic.com) (tradingkey.com) Broadcom’s role is the hardware plumbing behind that stack. Broadcom has long worked with Google on custom chip design and networking components, so Anthropic’s deal ties together the chip designer, the cloud operator, and the model company in one supply chain. (tradingkey.com) That structure matters because the artificial intelligence race has changed shape. Two years ago, the loudest competition was over model demos and benchmark scores, but in 2026 the harder contest is getting enough chips, networking gear, data center space, and electricity to keep those models running for paying customers. (bloomberg.com) (anthropic.com) Anthropic’s own growth numbers explain why it is locking in supply so far ahead. Bloomberg reported on April 6, 2026, that Anthropic’s revenue run rate had topped $30 billion, up from more than $9 billion at the end of 2025. (bloomberg.com 1) (bloomberg.com 2) That jump was not driven by a handful of experiments. Bloomberg said Anthropic now has more than 1,000 business customers spending over $1 million a year, which suggests the company is selling Claude into real budgets inside large organizations rather than just attracting curiosity from developers. (bloomberg.com) The company had already been moving in this direction. In October 2025, Anthropic said it was expanding its use of Google Cloud Tensor Processing Units and services, and it said at the time that it served more than 300,000 business customers. (anthropic.com) So this new agreement is not a sudden pivot. It is the next step after Anthropic spent months proving that Google’s chips could support Claude at commercial scale and that enterprise demand was rising fast enough to justify reserving future capacity years in advance. (anthropic.com 1) (anthropic.com 2) The comparison with OpenAI is part of why the announcement landed so hard. TradingKey and Bloomberg both pointed to reports that Anthropic’s current revenue run rate may now edge past OpenAI’s, which would have sounded unlikely to many investors a year ago. (tradingkey.com) (bloomberg.com) There is an important caveat in all run-rate comparisons. Run rate is not booked annual revenue; it is a projection built from recent performance, so it can rise fast in a boom and fall fast if customer usage slows. (bloomberg.com) Even with that caveat, the signal is clear. Anthropic is acting like a company that believes future winners in artificial intelligence will be decided partly by who secures electricity and chips first, then figures out features on top of that foundation. (anthropic.com) (tradingkey.com) Google also gets something valuable out of the arrangement. Every large commitment to Tensor Processing Units is a public argument that Google’s custom chips are now credible alternatives to the graphics processors that have dominated artificial intelligence training, and that helps Google sell more cloud infrastructure to other model builders. (bloomberg.com) (tradingkey.com) Broadcom benefits too, because custom artificial intelligence chips are one of the few hardware businesses where demand is now being booked years ahead. The more companies sign long-duration compute deals, the more valuable Broadcom becomes as the builder behind the branded cloud platforms. (tradingkey.com) The deeper story is that artificial intelligence companies are starting to look like industrial companies. They still ship software, but they now negotiate for transformers, networking gear, chip packaging, and power delivery with the urgency that older industries once reserved for oil, steel, and rail capacity. (thenextweb.com) (tradingkey.com) Anthropic’s bet is simple enough to state. If Claude keeps winning enterprise customers at anything close to its current pace, the company will need far more compute than the spot market can reliably provide, so it is trying to reserve the future before the future gets more expensive. (anthropic.com) (bloomberg.com)