Anthropic Says Claude Mythos Got Claude Running on AMD GPUs in About 3 Days
From the Anthropic Road Show video: 'This has also made us really good at bringing up new chips. As an example, this year, we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work. This time we had Mythos try it out. It got Claude running on AMD in just 3 days with basically no human intervention. Claude ported or wrote almost 100% of the kernels.'
Context
If a model can bring up new chips in days instead of months, AI labs become far less tied to Nvidia and its CUDA software.
Anthropic launched the roadshow for its planned IPO on Nov 10, including a video pitch to investors. Claude Mythos is Anthropic's most capable model line. Most AI labs train and serve models on Nvidia GPUs, whose CUDA software stack is often described as Nvidia's biggest advantage over rivals such as AMD.
Deep dive
Andrew Curran posted a quote he attributes to Anthropic's IPO roadshow video. In it, Anthropic says that this year it started working with AMD GPUs for the first time, and that in the past, bringing on a new chip platform took dozens of engineers months of work. This time, the quote says, Anthropic had Mythos try it out, and it got Claude running on AMD in just 3 days with basically no human intervention, with Claude porting or writing almost 100% of the kernels.
The chart in the post compares about 3 days with Mythos against about 6 to 12 months with the traditional approach. Its footnote says the 3 day figure reflects the approximate time for Claude Mythos Preview to bring up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance, and that the 6 to 12 month estimate is based on the company's past experience bringing up new accelerators.
The claim spread widely on X. Marcelo P. Lima wrote that the Nvidia CUDA moat is gone and that Anthropic sees all compute as fungible, while Jukan shared the same chart and asked why Anthropic chose to emphasize it.
The footnote limits the claim to prompt processing for one model, not full training or serving across Anthropic's fleet. The figures come from Anthropic's own marketing for its IPO and have not been independently verified.
What to expect next
Watch for Anthropic to say how much of its serving actually runs on AMD hardware, and whether the same approach extends to training and to other chips such as Google TPUs or Amazon Trainium.
AMD and Nvidia investors will read this as a test of how durable the CUDA advantage is. Other labs may now try to show that their own models can port kernels to new hardware, which would make claims like this easier to compare.