Back to news
researchAnthropicAMDNVIDIA2026-11-11

Anthropic IPO materials show Claude adapted a model to AMD GPUs in roughly three days

A page of Anthropic's IPO roadshow materials shows its Claude Mythos Preview model worked out how to run another model on AMD MI355X GPUs in about three days, against a conventional adaptation timeline of six to twelve months.

A page in Anthropic's initial public offering roadshow materials disclosed that as high-performance AI models increasingly gain the ability to adapt themselves to different computing platforms, graphics processors are increasingly becoming interchangeable general-purpose compute resources.

According to the disclosure, Anthropic's Claude Mythos Preview model worked out in roughly three days how to get another model running on AMD MI355X graphics processors. By contrast, the conventional work of adapting inference to a new computing platform has typically taken six to twelve months. That implies Anthropic can now treat graphics processors from different vendors as relatively general compute resources, with the technical barrier to switching platforms substantially lowered, which has led some to argue that Nvidia's long-celebrated CUDA ecosystem moat may have lost its original significance.

Declaring the end of CUDA's advantage now would still be premature. CUDA is not merely a set of software tools; behind it sits a very large developer ecosystem, rich function libraries and development tools, and years of accumulated optimization, and a great many developers still build and maintain applications around it. Over recent years Nvidia has also worked to reinforce CUDA's edge, for instance by providing graphics processors and training resources free to universities and researchers, and by binding the NCCL communication library used to connect multiple graphics processors closely to proprietary hardware such as NVLink and InfiniBand, raising multi-card collective efficiency while also raising the cost of migrating to other platforms.

AnthropicClaudeAMD GPUCUDA推理适配