Mythos put Claude on AMD in about 3 days, not 6-12 months

Anthropic co-founder Tom Brown says Claude Mythos got Claude running on AMD GPUs in about three days “with basically no human intervention”. He made the claim in an IPO roadshow video posted on Anthropic’s IPO site on November 11 and said the same job used to take dozens of engineers months. The fine print limits the measured result to prompt processing for one Anthropic model on AMD MI355X GPUs, and both the three-day figure and the ”~6-12 months” baseline next to it are Anthropic’s own numbers, taken from marketing for its share sale.

The claim appears in the video’s “Approach to compute” chapter, as Madrobot reported. Brown, Anthropic’s Chief Compute Officer, says the company only started using AMD GPUs this year and that bringing up a new chip platform used to be slow, manual work. In his account, “Claude ported or wrote almost 100% of the kernels”. Kernels are the small, heavily tuned programs that run a model’s maths on a specific chip, and every chip family needs its own set.

What the footnote covers

A footnote in small type sits under the on-screen chart that sets three days against ”~6-12 months with traditional approach”. It says the three days 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”. Prompt processing is the stage where a model reads its input, before it starts generating an answer. So the claim covers only that first stage for one model. It does not say that a whole model, or any training, moved to AMD in three days.

The baseline is Anthropic’s own as well. The footnote says the 6 to 12 months “is based on the Company’s historical experience bringing up new accelerators using traditional engineering approaches”. Anthropic has published no technical details that would let anyone else check either number.

Which AMD chips

The test ran on the MI355X, a chip Anthropic was already using before the two companies announced their partnership on July 22. That deal covers something else: up to 2 gigawatts of the next-generation MI450 series, delivered as AMD Helios racks with MI455X GPUs, with the first gigawatt from the first half of 2027. The three days tells you nothing yet about the hardware the big contract is for. The July deal also included up to $5 billion of AMD investment in Anthropic and a plan to use Claude to speed up development of ROCm, AMD’s answer to Nvidia’s CUDA software.

AMD is also buying $2.5 billion of Anthropic shares alongside the IPO. The final price is expected on November 18, and the shares are due to start trading on Nasdaq under ANT the next day.

Why Nvidia comes into it

Nvidia’s strongest defence is software. Most AI software is written for CUDA, and moving to rival hardware means rewriting and re-tuning the kernels underneath. That cost keeps many labs on Nvidia. If a model can do most of that porting in days, switching suppliers gets much cheaper, and labs would likely gain leverage over the companies that sell them chips. Others are attacking the same lock-in by different routes. Google and Meta are building TorchTPU so that PyTorch code written for Nvidia runs on Google’s chips, and DeepSeek gave Huawei’s Ascend chips their own open source software.

Brown never names Nvidia or CUDA in the video, and Nvidia is still one of Anthropic’s biggest suppliers. It is buying $10 billion of Anthropic shares in the same placement as AMD. Anthropic’s S-1 also commits it to spend up to $120.5 billion on SpaceX.AI’s Nvidia GPU compute over terms running into 2029, a figure published in October. Cheaper porting gives Anthropic more choice about where to run its workloads. It does not change those contracts.

“One fungible pool of compute”

Brown says Anthropic now trains and serves Claude “across all four major chip families”. The video’s graphics show these as Amazon’s Trainium, Google’s TPUs, and GPUs from Nvidia and AMD. He calls building a separate software stack for each one an early “unconventional bet”, made when many people thought it wasn’t worth the effort. The payoff, he says, is that Anthropic can match each chip to the job it does best and “treat all of our chips across providers, architectures and locations as one fungible pool of compute”. The IPO materials describe that pool being shared out across research, model development, customer services and Anthropic’s internal use of AI. The investor fact sheet names the multi-chip architecture and research breakthroughs as the foundations of the approach.

Those breakthroughs are what Brown calls “compute multipliers”: ways to get a smarter model out of the same hardware. He expects Claude to take over finding them too, an idea that worries some leading researchers. For now it is a goal, and the materials offer no evidence that it is already happening:

We believe that this can accelerate even further as we approach recursive self-improvement. Claude will be able to discover compute multipliers entirely by itself, allowing the number of compute multipliers to grow faster and faster.