Anthropic says Mythos ported Claude to AMD GPUs in 3 days
Co-founder Tom Brown told IPO investors the work once took dozens of engineers months, and a slide footnote limits the claim to one model on MI355X GPUs.
Summary

In brief
- Anthropic says Mythos got Claude running on AMD GPUs in 3 days, almost unaided
- A slide footnote limits the claim to prompt processing for one model on MI355X GPUs
- AMD committed in July to invest up to $5 billion in Anthropic
Anthropic’s Mythos model got Claude running on AMD GPUs in three days with “basically no human intervention,” Chief Compute Officer Tom Brown said in the company’s IPO roadshow video, released on November 10.
Software is a large part of Nvidia’s hold on the chip market: models built around its CUDA platform, the programming layer for its GPUs, are costly to move to rival hardware. Brown, an Anthropic co-founder, gave the AMD port as an example of how quickly the company can now bring up new chips.
In the past, bringing on a new chip platform took dozens of engineers months of work, Brown said. Anthropic started working with AMD GPUs for the first time this year and had Mythos, the restricted-access version of its most capable model, try the job. Claude ported or wrote almost all of the kernels, the low-level routines that run a model’s math on a chip, he said.
A slide shown during the segment sets about 3 days with Mythos against 6 to 12 months with what it calls the traditional approach. Its footnote narrows the claim. The 3 days is the approximate time Claude Mythos Preview took to bring up prompt processing for one Anthropic model on AMD’s MI355X GPUs at a useful level of performance, and the longer estimate rests on Anthropic’s own past experience. The figures have not been independently verified.
“We train and serve Claude now across all four major chip families,” Brown said, and a slide names them as AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Each needs its own software stack, which he called expensive and difficult to build, but he said the approach lets Anthropic seek the best price for performance while chips are scarce.
Brown said the company treats its chips across providers, architectures and locations as a single interchangeable pool, and that a dollar of compute goes further at Anthropic than at any other frontier lab. He also described Anthropic Direct, under which the company leases chips and runs them in data centers built for it.
AMD is also an Anthropic investor. In a July 22 announcement, the chipmaker committed to an equity investment of up to $5 billion, and Anthropic agreed to deploy up to 2 gigawatts of AMD’s MI450 series GPUs from the first half of 2027, building on its use of the MI355X. Anthropic first described running Claude on AWS Trainium, Nvidia GPUs and Google TPUs in a September 2025 postmortem.
Anthropic is seeking up to about $64.4 billion in its IPO. It expects to set the final share price on November 18 and to begin trading on Nasdaq on November 19, subject to regulatory clearance.