Anthropic commits to up to two gigawatts of AMD Instinct MI455X in Helios racks, the largest non-NVIDIA frontier commitment yet
AMD and Anthropic outlined a partnership to deploy up to 2 GW of Instinct MI455X GPUs in Helios rack-scale systems. AMD claims Helios delivers up to 30% more inference tokens per dollar than the competition. A frontier lab placing two gigawatts outside NVIDIA is the first commitment at a scale that changes anyone's planning assumptions.
Second-source announcements have been a recurring genre for three years and have mostly meant evaluation clusters. Two gigawatts is not an evaluation cluster. It is a capacity commitment on the order of a frontier lab's entire serving footprint, which means the software risk — kernels, compilers, the accumulated CUDA dependency of an entire research org — has been judged acceptable.
The tokens-per-dollar framing tells you which workload this is for. Inference economics, not training throughput, is where the claim is staked, and inference is where the volume and the margin pressure now live. A serving fleet is also a far more forgiving place to absorb a toolchain migration than a training run that has to complete.
What to watch is whether utilisation figures follow. A gigawatt commitment announced is not a gigawatt commitment running, and the gap between the two is where every previous second-source story has quietly ended. If Anthropic reports real serving share on Instinct by year end, the single-vendor era is genuinely over.
AMD Investor Relations — AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era → · GlobeNewswire — AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era →