Claude runs AMD's new GPU in a weekend, breaking CUDA's wall

AMD sent Anthropic a rack with MI355X chips. The team thought it would take weeks to get it running. Instead, they let Claude connect to the machine and said 'make it work.' When the weekend ended, the performance chart was already climbing—and still optimizing itself. Anthropic co-founder Tom Brown said only one engineer was involved, and AMD's offer to help was politely declined.

Anthropic announced it will deploy up to 2GW of Instinct GPUs in AMD Helios systems, with the first 1GW set to start in the first half of 2027. The two companies will also use Claude to optimize AMD workloads and speed up ROCm software development. AMD launched ROCm.AI, a toolbox that lets agents set up environments, deploy models, read logs, and troubleshoot on their own.

AMD changed how it writes chip manuals. Each new GPU now comes with an AI-readable instruction set. Agents use Hyperloom to test different setups and create custom code. In a demo, it boosted MiniMax M3's output speed by 38%. AMD is also working with leading AI labs to make models 'naturally speak AMD programming.'

CUDA's strength came from years of human expertise. Agents can now find and fix bottlenecks in parallel, turning years of work into tasks that run overnight. Chinese engineers are already at the forefront of this shift, and their know-how is being turned into skills that any agent can use.