Anthropic Mulls Own Chips

Anthropic is reportedly thinking about building its own artificial intelligence chips instead of renting so much computing power from Nvidia, and Reuters said on April 9 that the idea is still early enough that the company could walk away from it. (reuters.com) That sounds strange until you remember what an artificial intelligence chip does: it is the engine that trains a model like Claude and then runs it every time a user asks a question. Nvidia sells most of those engines today, which gives one supplier enormous leverage over price, delivery, and access. (reuters.com) Anthropic is not starting from zero on hardware. In October 2025, the company said it already used three chip families at once: Google’s tensor processing units, Amazon’s Trainium chips, and Nvidia’s graphics processing units. (anthropic.com) A chip strategy like that is the artificial intelligence version of buying electricity from three power plants instead of one. Anthropic said that mix was meant to keep capacity flowing as demand for Claude kept rising across large companies and startups. (anthropic.com) Amazon is especially deep in the picture. Anthropic said in November 2023 that Amazon would invest another $4 billion, bringing Amazon’s total investment to $8 billion, and making Amazon Web Services its primary cloud and training partner. (anthropic.com) That partnership got even bigger in late 2025, when Amazon said Project Rainier had brought nearly 500,000 Trainium2 chips online and that Anthropic was expected to scale past 1 million chips by the end of 2025. (aboutamazon.com) Google is also tied into Anthropic’s compute stack. On April 6, 2026, Anthropic said it was expanding work with Google Cloud and Broadcom, and repeated that Claude already trains and runs across Amazon Trainium, Google tensor processing units, and Nvidia graphics processing units. (anthropic.com) So the reported chip plan is not a clean break from partners. It looks more like Anthropic trying to own one more layer of the machine, the way large cloud companies design some of their own servers even when they still buy parts from Intel, AMD, or Nvidia. (reuters.com) Other artificial intelligence companies have already gone down that road because custom chips can be tuned for one job instead of every job. Google built tensor processing units for its own workloads years ago, and Amazon built Trainium to cut the cost of training large models inside Amazon Web Services. (blog.google, aboutamazon.com) The catch is that designing a chip is only the first mile. A company still needs software tools, manufacturing partners, packaging, networking, and data centers, which is why Reuters said Anthropic may decide it is easier to keep buying than to become a full chip designer. (reuters.com) That is why this report landed just days after Anthropic announced more outside compute deals, not fewer. A company can chase independence and still lock in supply wherever it can get it, especially when the race for chips is now as important as the race for models. (anthropic.com, reuters.com)