Anthropic exploring its own chips

Anthropic is reportedly exploring its own artificial intelligence chips even though Amazon Web Services is its primary training partner and Google just signed a new compute deal with it. That tells you how tight the market for the chips behind large models still is in April 2026. (cnbc.com) (anthropic.com 1) (anthropic.com 2) Right now, most frontier artificial intelligence systems run on a small set of very expensive accelerators, with Nvidia still setting the pace for much of the market. If you cannot get enough chips, you cannot train bigger models fast enough or serve millions of users cheaply enough. (cnbc.com) Anthropic’s plan is still early. Reuters reported on April 9 that the company has not committed to a final design and could still decide to keep buying chips instead of building them. (usnews.com) The timing lines up with a demand spike inside Anthropic itself. The company said this week that its annualized revenue run rate rose to $30 billion, up from about $9 billion at the end of 2025, and it also signed for multiple gigawatts of new Google Tensor Processing Unit capacity starting in 2027. (anthropic.com) (thenextweb.com) A custom chip is not a general-purpose truck. It is more like building a delivery van around one route, one cargo size, and one warehouse system so each trip wastes less fuel, space, and time. (cloud.google.com) Google has spent years doing exactly that with its Tensor Processing Units, which are chips built specifically for machine learning workloads. Its latest Ironwood system scales to 9,216 chips in one pod, which shows why the biggest model companies increasingly treat hardware design as part of the product, not just a supplier choice. (blog.google) Amazon is pushing the same idea from another angle. In November 2024, Anthropic said Amazon Web Services would become its primary cloud and training partner, and that it would use Amazon Trainium and Inferentia chips for future foundation models. (anthropic.com) (aboutamazon.com) Meta has already gone further down the road to in-house silicon. In March 2026, it said it was developing and deploying four new generations of Meta Training and Inference Accelerator chips within two years to run recommendation and generative artificial intelligence workloads more efficiently. (about.fb.com) OpenAI is on the same path too. CNBC reported in October 2025 that OpenAI and Broadcom were jointly building custom accelerators for deployment starting in 2026, which means Anthropic would not be inventing a new playbook so much as joining an arms race already underway. (cnbc.com) The hard part is that designing a chip is only half the job. You also need software tools, networking, memory packaging, manufacturing slots, and enough scale to justify years of engineering before the first board reaches a data center. (cloud.google.com) (about.fb.com) So Anthropic’s chip effort, if it happens, would not mean a fast break from Nvidia or from its cloud partners. It would mean Anthropic wants a seat at the hardware table, where cost, supply, and bargaining power are now as important to an artificial intelligence lab as model quality itself. (cnbc.com) (anthropic.com)