Anthropic Uses Claude to Design Its Own AI Chips, Investors Told
A team of 15 Anthropic engineers is reportedly using Claude to design the company's custom chips, and says it cuts time versus traditional processes at Nvidia and AMD.
Anthropic is using its own Claude model to help design custom AI chips. A team of 15 engineers applies Claude to chip design and engineering work, and the company described the effort to prospective investors in recent meetings, according to sources familiar with the matter.
The engineers believe Claude significantly shortens the work compared with their own estimates of the traditional design process at large chipmakers such as Nvidia and AMD. These are the team's internal assessments, not the result of an independent comparison. Neither the size of the speedup nor the design stages Claude handles has been disclosed publicly.
Part of a multi-chip strategy
Anthropic announced in August 2026 that it intended to build its own accelerators. It is assembling an in-house team to create specialized silicon that would help train and run Claude faster and more efficiently.
The company does not plan to drop third-party hardware. It currently relies on accelerators from Amazon, Google, Nvidia and AMD, and the custom chips are meant to become one more element in a multi-chip strategy.
How capable are models at chip design?
A benchmark from ChipAgents and Nvidia, which measures how well models can handle chip development, offers a sobering counterpoint. Claude Fable 5.1 posted the best score among participants, but reached only 16.6%. That suggests fully automated chip design remains far off, even if models can speed up parts of the workflow.
Anthropic is not alone. OpenAI has also used its models to develop Jalapeño, its own inference chip built with Broadcom. Earlier model generations helped design and bring up the chip, while newer ones helped optimize it and write software for it. OpenAI published first test results for Jalapeño in August.
For the industry, the story points to a feedback loop: AI labs build models, then use those models to build the hardware that runs the next generation. If the claimed time savings hold up, it could lower the barrier for AI companies to develop custom silicon and reduce their dependence on a handful of suppliers.
Apps worth a look
SUMO: the number puzzleTap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Dialo: learn languages by speakingTalk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
ediviewer: read EDI files like documentsOpen EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
More from this day
See all →
Y Combinator's Garry Tan spends up to $200K a year on AI tokens
CODE380's D·Vision SDR lets Ukraine watch Russian drone feeds
EGG UP Launches WEARLOG to Track Rented Garments for TV Stylists
