Anthropic says Claude runs on four chip families, with AMD port done in three days
Anthropic's IPO roadshow video says Claude now runs on four chip families, and that an AI-driven AMD port took three days.
Anthropic says Claude is now trained and served across four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. The claim comes from co-founder Tom Brown's section of the company's IPO roadshow video, which the company released as the roadshow began. The company has not published an official transcript, so Simon Willison extracted one from the video. The quotations below come from that transcript and from a text version of the same section.
Brown ties the multi-chip position to a bet the company says it made early on. Each chip needs its own software stack and inference stack. Brown calls these expensive and difficult to build, and he says Anthropic has been doing this for years. He argues the payoff is flexibility. In a supply-constrained market, the company can pick the best price performance from any provider and match chips to jobs, with some better at training and others at inference. It can also treat chips across providers, architectures and locations as one pool of compute. The video does not give figures for what that pool costs or how much of Claude runs on each chip, so the four-family claim describes capability rather than workload share.
The most specific claim concerns AMD. Brown says Anthropic began working with AMD GPUs this year. He says bringing on a new chip platform used to take dozens of engineers months of work. This time, he says, the company had Mythos try it out. In his words, it "got Claude running on AMD in just three days, with basically no human intervention," and Claude "ported or wrote almost 100% of the kernels." Kernels are the low-level routines that run the actual arithmetic on a chip. The transcript gives no detail on what the port covered, how it performed, or how many engineers reviewed the output. The three-day figure is therefore the company's account of its own process.
The second part of Brown's argument concerns what he calls compute multipliers. Anthropic measures what it calls total effective compute as physical capacity multiplied by the multipliers it discovers. His example: if a model needs 100 units of compute, a 4x multiplier would let it be trained with 25 units, or let the same 100 units train a model that would otherwise need 400. Brown says the multipliers stack, because each new one improves those found before it. He expects Claude will eventually find multipliers on its own as the company approaches what he calls recursive self-improvement. That last point is a forecast, not a result. The video does not identify any specific multiplier or say how one is found. It also reports no measured gain beyond the illustrative 4x.
The roadshow section also says Anthropic now leases chips and runs them in data centers built for it, in addition to buying cloud compute. The company calls this arrangement Anthropic direct. Brown says it helps lower the total cost of ownership and separates the data-center commitment from the chips. Willison also points to a September postmortem by Anthropic about bugs that affected only its TPU deployments. He read that as an early hint of the multi-chip work.
What changed
Claude is now trained and served across four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs, according to Anthropic.
Type: New capability · Applies to: Claude, AWS Trainium, Google TPUs, NVIDIA GPUs, AMD GPUs
Why it matters
If the claims hold, Anthropic can move Claude across several accelerator vendors, which could widen the hardware options behind the models buyers rely on. The figures come from company statements in an IPO roadshow video and are not independently verified.
Original source
Further reporting
Where we saw it
Simon Willison
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works