Anthropic Eyes Custom Chips

Anthropic is reportedly considering designing its own artificial intelligence chips even though it just locked in huge outside supply from Google and Amazon, which tells you how hard it has become to get enough computing power once an AI company gets big fast. (cnbc.com) The timing is tied to a revenue jump that is unusually steep even by AI standards: Anthropic said this week its annualized revenue run rate passed $30 billion, up from about $9 billion at the end of 2025. (anthropic.com) That kind of growth turns chips into the main raw material of the business. A model like Claude needs vast numbers of processors both to train new versions and to answer millions of user requests after launch. (aws.amazon.com) Right now Anthropic does not mainly run on its own hardware. Reuters says it uses Google tensor processing units, which are Google’s in-house AI chips, and Amazon chips as well, and the custom-chip idea is still early enough that the company has not committed to a design or assembled a dedicated team. (cnbc.com) Anthropic has spent the past year tying itself closely to Amazon. In November 2024, Anthropic said Amazon Web Services became its primary cloud and training partner, and Amazon’s total investment in the company rose to $8 billion. (anthropic.com) It also deepened ties with Google this week. Anthropic said on April 6 that it expanded its partnership with Google and Broadcom for multiple gigawatts of tensor processing unit capacity, which is the kind of long-term reservation companies sign when they are worried about future shortages. (anthropic.com) So the story is not that Anthropic is walking away from partners. It is that even after securing outside supply through deals measured in gigawatts, it may still want a chip built around Claude’s exact workload, the same way a delivery company eventually wonders whether buying trucks forever is more expensive than designing its own fleet. (anthropic.com) (cnbc.com) There is a recent template for this. Google spent years building tensor processing units to reduce dependence on Nvidia, and Broadcom said this week it signed a long-term agreement through 2031 to help Google develop future generations of those chips. (msn.com) Amazon is following the same playbook with Trainium, its in-house accelerator family, which Amazon says is built for cost-efficient training and inference, meaning both teaching a model and running it after it is trained. (aws.amazon.com) Anthropic is now big enough that a few percentage points of savings on each query can turn into enormous money. The company said more than 1,000 business customers are already spending at least $1 million a year, more than double the figure it disclosed in February. (anthropic.com) The catch is that chip design is slow, expensive, and easy to get wrong. Reuters says Anthropic may still decide not to build anything and keep buying from others, which is why the clearest reading of this move is not “new chip soon” but “AI labs are starting to look more like utilities, where owning the pipes can matter as much as owning the product.” (cnbc.com)