Anthropic eyes own chips

Anthropic is reportedly looking at designing its own artificial intelligence chips, which is unusual because model labs usually buy compute from cloud companies instead of trying to build the silicon themselves. Anthropic already runs Claude on Nvidia graphics processing units, Google tensor processing units, and Amazon Trainium chips, so moving one step deeper into chip design would push it closer to the infrastructure business. (bloomberg.com) (anthropic.com) An artificial intelligence chip is the engine inside a data center server that does the matrix math behind training and answering prompts. Nvidia became the default supplier because its graphics processing units arrived with the software tools and networking pieces that let thousands of chips work together as one machine. (nvidia.com) (developer.nvidia.com) That dependence has become expensive. Anthropic said five days ago that it is expanding its partnership with Google and Broadcom, and in November 2025 it committed to invest $50 billion in United States computing infrastructure, which shows how much capital is now tied up in securing enough hardware. (anthropic.com) Anthropic has spent the last two years spreading its bets across other people’s chips instead of relying on one supplier. In late 2024 it said its compute strategy used three platforms at once: Google tensor processing units, Amazon Trainium, and Nvidia graphics processing units. (anthropic.com) Amazon is central to that plan. Amazon said in 2023 and 2024 that it would invest a total of $8 billion in Anthropic, make Amazon Web Services its primary cloud and training partner, and work with Anthropic on future generations of Trainium accelerators through Annapurna Labs. (aboutamazon.com) (anthropic.com) Google is central too. Anthropic said this week that it is deepening work with Google Cloud and Broadcom, and Bloomberg reported on April 6 that Broadcom confirmed a deal to ship Google tensor processing unit chips to Anthropic as Claude demand accelerated. (anthropic.com) (bloomberg.com) Microsoft and Nvidia have also moved closer. Anthropic said four months ago that it would scale Claude on Microsoft Azure powered by Nvidia and commit to purchase $30 billion of Azure compute capacity, which means even a company trying to diversify is still writing giant checks for outside hardware. (anthropic.com) The pressure is not just training new models once. Nvidia said in 2025 that newer “reasoning” systems can require far more compute at inference time, with long-thinking workloads using up to 100 times the compute of a single inference pass, so every popular chatbot answer now burns through more silicon than the old autocomplete style did. (developer.nvidia.com) If Anthropic does design its own chip, it probably would not look like a consumer gadget with an Anthropic logo on the box. The more likely path is a custom accelerator built with a manufacturing and design partner like Broadcom, the same way cloud companies use in-house chips to tune cost, power use, and speed for their own workloads. (anthropic.com) (aboutamazon.com) That would make Anthropic less like a pure model lab and more like the giant cloud firms it depends on today. The company is already tied into Project Rainier at Amazon, which Amazon said reached nearly half a million Trainium2 chips and is expected to scale Anthropic to more than one million chips by the end of 2025, so the next fight is no longer just who has the best model but who controls the machine room underneath it. (aboutamazon.com)