THE CRUNCH

Anthropic has told IPO investors that its Claude Mythos Preview model handled the software work needed to run one of its AI models on AMD's MI355X GPUs in about three days, with "basically no human intervention". Co-founder and chief compute officer Tom Brown said the same job previously took six to 12 months and a team of dozens of engineers. The model ported or wrote "almost 100% of the kernels", the low-level code that binds an AI model to a particular chipmaker's hardware.

The three-day figure comes with a footnote: it covers prompt processing for one Anthropic model on the MI355X at what the company called a useful level of performance, not a full production deployment. Until this week, versions of the claim came from AMD's stage, including Brown's own July keynote appearance; the IPO roadshow is the first time Anthropic has made the case to its own investors, and its prospectus repeats the point in writing, though without a number: "We can onboard new chip platforms in a fraction of the time it once took."

The claim matters because that kernel-writing work has long been part of Nvidia's CUDA moat, the mature programming platform that keeps AI code tied to its chips. AMD's equivalent, ROCm, requires CUDA code to be converted, historically via a tool called HIPify. AMD's vice president of AI software, Anush Elangovan, told EE Times in April that "Claude is better than HIPify because it has web search built in".

Anthropic already runs its models on Nvidia and AMD GPUs, Google's TPUs and AWS Trainium, managed as "one fungible pool of compute", a decision Brown called "an unconventional bet". It began working with AMD GPUs for the first time this year and has an agreement for up to 2 gigawatts. Under the July deal, AMD also uses Claude to optimise workloads for its Instinct GPUs and to speed up ROCm development.

WHAT HAPPENS NEXT

Anthropic's IPO is due next week, when investors will get a closer look at the multi-chip strategy the claim is meant to support.