Model firms eye custom silicon

Anthropic is reportedly considering its own artificial intelligence chips even though it has not picked a design or built a dedicated chip team yet. That is a sharp move for a company that, until now, has mostly been known for selling the Claude model, not for making hardware. (cnbc.com) The trigger is scale. Anthropic said Claude’s annualized revenue run rate has climbed past $30 billion in 2026, up from about $9 billion at the end of 2025, which means every efficiency gain in computing now lands on a much bigger business. (thenextweb.com) An artificial intelligence chip is the engine that does the math behind training and running a model. Buying those engines from Nvidia, Google, or Amazon is fast, but it also means your costs, supply, and product timing depend on somebody else’s factory schedule. (usnews.com) That is why custom silicon keeps coming up. A company can shape a chip around the exact jobs its model does most, the same way a delivery fleet might swap a general-purpose van for one built only for its own routes. (thenextweb.com) Anthropic is not starting from zero on infrastructure. On April 7, it expanded a deal with Google and Broadcom for multiple gigawatts of next-generation tensor processing unit capacity scheduled to come online in 2027. (techcrunch.com) A tensor processing unit is Google’s in-house artificial intelligence chip, built for the matrix math that large models use constantly. By signing for future capacity now, Anthropic is trying to lock in power before demand gets even tighter. (outlookbusiness.com) OpenAI is answering this from the other direction. In a memo to investors reported this week, OpenAI argued that its edge is not a custom chip announcement but sheer infrastructure scale, saying it has 1.9 gigawatts of computing capacity versus Anthropic’s 1.4 gigawatts. (bloomberg.com) OpenAI’s memo also said it plans to reach 30 gigawatts of compute by 2030, while it expects Anthropic to be around 7 to 8 gigawatts by the end of 2027. That tells you the argument investors are hearing now: not just whose model is smarter, but who can keep enough machines running to serve millions of users. (cnbc.com) At the same time, OpenAI has paused its Stargate data center project in the United Kingdom, citing high energy costs and regulatory uncertainty. Even the companies with the biggest ambitions are finding that artificial intelligence expansion is now limited by power prices, permits, and grid access as much as by software. (bloomberg.com) So the race is getting more vertical. One company is exploring chips, another is selling investors on existing capacity, and both are being pulled toward the same conclusion: the artificial intelligence business now runs on whoever controls more of the hardware stack, from the silicon at the bottom to the chatbot at the top. (the-decoder.com)