AI 基礎設施
Anthropic to Expand Compute Capacity by 2 GW With AMD Helios and Use Claude to Optimize ROCm
Anthropic plans to deploy its first AMD Instinct MI450-series GPUs starting in the first half of 2027, with a total capacity of up to 2 GW. The partnership goes beyond hardware procurement: the companies will also use Claude to optimize GPU workloads and the ROCm software stack.
AMD and Anthropic have announced an expanded infrastructure partnership under which Anthropic will deploy up to 2 GW of Instinct MI450-series GPUs in Helios rack-scale systems, with the first 1 GW scheduled to come online in the first half of 2027. The systems will use MI455X GPUs, EPYC CPUs code-named Venice, Pensando networking, and ROCm software for both training and inference. This will scale Anthropic’s current use of MI355X GPUs to full rack and data center deployments.
More technically significant is the companies’ multi-year engineering collaboration. Anthropic will use Claude to help tune Instinct workloads and accelerate ROCm development, while AMD will deploy Claude across its engineering and product teams. If this feedback loop proves effective, the model could help analyze kernels, generate or modify operators, and identify performance bottlenecks, with hardware telemetry and benchmark results used for validation. This could shorten the software maturation period required to move a new GPU from merely running workloads to achieving high utilization. For teams dependent on the CUDA ecosystem, the real competitive battleground is therefore not peak FLOPS, but whether compilers, communication libraries, core operators, distributed training systems, and inference frameworks can operate reliably.
AMD has also committed to investing up to $5 billion in Anthropic. However, both the investment and the “up to 2 GW” capacity are forward-looking arrangements and should not be treated as capacity that has already been delivered. The companies have not disclosed the GPU count, memory configurations, interconnect topology, performance-per-watt figures, or benchmark data for Claude workloads. Engineering teams should next watch the MI455X production timeline, Helios failure rates and network efficiency in large-scale clusters, and whether widely used stacks such as PyTorch, vLLM, and SGLang can deliver reproducible performance on ROCm.