Anthropic weighs designing its own chips

Anthropic is studying whether to design its own artificial intelligence chips instead of relying only on chips it buys from other companies, according to a Reuters report published on April 9, 2026. The talks are still early enough that Anthropic could decide not to build anything at all. (reuters.com) That sounds like a company deciding whether to bake its own bread after years of waiting for deliveries from one crowded bakery. In artificial intelligence, the bakery is the market for high-end accelerators, the specialized chips used to train models and answer user prompts at scale. (reuters.com) Anthropic already depends on several outside suppliers because one chip family is no longer enough. The company said this week that it trains and runs Claude on Amazon Web Services Trainium chips, Google Tensor Processing Units, and Nvidia graphics processing units. (anthropic.com) Amazon has been pushing Anthropic toward that multi-chip world for more than two years. In September 2023, Amazon said Anthropic had selected Amazon Web Services as its primary cloud provider and would train and deploy future models on Trainium and Inferentia chips. (aboutamazon.com) Google is now in the picture at enormous scale too. Anthropic said in October 2025 that it was expanding its use of Google Cloud Tensor Processing Units in a deal worth tens of billions of dollars that was expected to bring well over a gigawatt of capacity online in 2026. (anthropic.com) So the question is no longer whether Anthropic can get access to chips. The question is whether renting capacity from Amazon, Google, and Nvidia is enough when the biggest model labs are turning chip choices into part of their core product strategy. (anthropic.com) (reuters.com) Other companies have already moved in that direction. Meta said in March 2026 that its Meta Training and Inference Accelerator program now includes four new generations of custom chips planned within two years, and OpenAI has been reported to be working with Broadcom and Taiwan Semiconductor Manufacturing Company on its own custom accelerators. (about.fb.com) (datacenterdynamics.com) Designing a chip does not mean building a chip factory. A lab like Anthropic would still need partners for design software, manufacturing, advanced packaging, and memory, which is why custom silicon usually starts as a long partnership with companies like Broadcom or Taiwan Semiconductor Manufacturing Company rather than a clean break from the supply chain. (reuters.com) (anthropic.com) The payoff is control. A custom chip can be tuned for one job, like training giant models faster or serving millions of chatbot replies more cheaply, instead of paying for a general-purpose chip that does many things well but none exactly the way one lab wants. (cloud.google.com) (about.fb.com) Anthropic is large enough now for that math to matter. Reuters reported in February 2026 that the company raised $30 billion at a $380 billion valuation, which means even small improvements in chip cost or availability could move very large numbers. (reuters.com) If Anthropic goes ahead, the shift will be less about one new chip than about one more artificial intelligence lab acting like a cloud company. The business of making models would keep sliding downward into the hardware stack, where the fight is no longer just over smarter software, but over who controls the machines that make the software possible. (reuters.com) (anthropic.com)