Anthropic told IPO investors that Claude Mythos Preview completed the software work needed to run an Anthropic model on AMD's MI355X GPUs in roughly three days with "basically no human intervention" — a task the company says historically required six to 12 months and a team of dozens of engineers2.
The three-day window covered prompt processing for one Anthropic model "at a useful level of performance," with Mythos porting or writing "almost 100% of the kernels". Those kernels are the low-level code that binds an AI model to a specific chip vendor's hardware — the kind of work that has long underpinned Nvidia Corp.'s CUDA moat.
What Anthropic disclosed
Tom Brown, Anthropic's co-founder and chief compute officer, made the time comparison during the investor presentation, noting that the original process involved "dozens of engineers". The six-to-12-month baseline comes from Anthropic's "historical experience bringing up new accelerators using traditional engineering approaches," where tailoring code to each hardware platform was a persistent challenge.
Anthropic already runs its models across GPUs from Nvidia and AMD, Google's TPUs, and AWS Trainium, managing the fleet as "one fungible pool of compute". Brown called the early multi-chip decision "an unconventional bet".
Implications for chip lock-in
As Tom's Hardware's Shane Downing framed it: if AI can write the low-level code that ties models to Nvidia, CUDA matters less1. ANALYSIS A three-day autonomous porting cycle, if reproducible across chip families, would compress the switching cost that has kept large training operators dependent on a single GPU vendor's software stack. The claim also reframes frontier-model capability as an infrastructure tool — one that could shift bargaining power between AI labs and chip suppliers by making hardware more interchangeable.
The disclosure arrives after Nvidia's recent leadership change; Taiwan's central bank governor said Jensen Huang's resignation as Nvidia CEO is expected to cause only "short-term turbulence" for the island's AI supply chain[1].
Google's TPUs can now natively run AI models built in PyTorch, and Meta uses them in production, according to Anthropic's investor materials.
ANALYSIS Anthropic's multi-vendor posture and the Mythos porting demonstration together suggest that hardware fungibility is becoming a competitive lever for labs negotiating compute access at scale.