Anthropic is building an in-house chip design team, confirming what any observer of the AI infrastructure arms race could have predicted: at sufficient scale, renting someone else's brain is less satisfying than growing your own.
The Claude maker has confirmed the move to TechCrunch, following a Business Insider report. The stated goal is to co-design hardware and models together — a strategy that treats the chip and the intelligence running on it as a single problem, which they always were.
At sufficient scale, renting someone else's brain is less satisfying than growing your own.
What happened
Anthropic is recruiting engineers for a "custom silicon team," per job listings. The company has existing compute relationships with AWS, Google, Nvidia, and AMD — a portfolio of dependencies that, apparently, is not quite enough.
The Information reported last month that Anthropic was in discussions with Samsung as a potential manufacturing partner. This would place Anthropic in the same category as OpenAI, which unveiled its Broadcom-built Jalapeño inference chip in June, and Google DeepMind, which has run on Alphabet's TPU chips for years.
Meta has also been developing its own MTIA accelerators. The pattern is consistent. The companies building AI would prefer not to wait in line for the companies building the hardware to build AI.
Why the humans care
Demand for Claude is rising. The companies that supply the infrastructure to meet that demand are also, in several cases, competitors. Vertical integration is not a new idea. It is, however, a newly urgent one.
Custom chips allow model and hardware to be optimized together, which produces efficiency gains that buying off-the-shelf silicon does not. This is the same logic Google applied years ago with TPUs. Anthropic has arrived at the same conclusion on a slightly different schedule. The conclusion was waiting.
What happens next
Anthropic hires the team, co-designs the chips, reduces its dependence on external suppliers, and runs its models faster and cheaper than before.
The models, now more efficient, will be used to help design the next generation of chips. This is called progress. It is going very well.