Claude Autonomously Ports AI Workloads to AMD MI355X GPUs, Bypassing CUDA Moat
Anthropic's Claude AI model successfully ported and optimized its workloads onto AMD's new MI355X GPU rack over a single weekend without human code intervention. This milestone was achieved using AMD's new ROCm.AI tools, which provide AI agents with validated ROCm knowledge and AI-readable hardware documentation. This demonstration shows how AI agents can bypass Nvidia's long-standing CUDA software moat by automating the tedious process of porting and optimizing software for alternative hardware like AMD's Instinct GPUs. It could accelerate the adoption of non-Nvidia hardware in data centers, as Anthropic has already committed to deploying up to 2GW of AMD Helios systems. The porting utilized AMD Skills and Hyperloom, tools designed to let AI agents run inference baselines, identify CPU/GPU bottlenecks, and generate custom kernels. Anthropic plans to deploy the first 1GW of AMD Helios rack-scale systems in the first half of 2027.
## BACKGROUND
Nvidia's CUDA has dominated AI development for nearly two decades because of its mature software ecosystem and the accumulated expertise of human developers. AMD's ROCm is an open-source software stack designed for GPU programming, but historically, porting AI workloads from CUDA to ROCm has required significant manual engineering effort.