AMD launches Helios AI platform and invests up to $5 billion in Anthropic

Advanced Micro Devices (AMD) will invest up to $5 billion in Anthropic and supply the AI lab with up to 2 gigawatts of Instinct graphics processors, the two companies said. The commitment anchored AMD’s Advancing AI 2026 event in San Francisco on July 22.

The deal pairs an equity bet with a hardware commitment. Anthropic will run AMD’s new MI450-series accelerators in Helios racks, with the first gigawatt of capacity coming online in the first half of 2027. For AMD, it secures a marquee frontier-lab customer while Nvidia still dominates the market for AI training silicon.

Helios is the system AMD is selling into that demand: a liquid-cooled, rack-scale unit that combines 72 Instinct MI455X GPUs and 18 6th Gen EPYC “Venice” CPUs. AMD says it delivers up to 30% more inference tokens per dollar than the leading competing system, a vendor figure that has not been independently verified. The MI455X carries 432GB of HBM4 memory, and the Venice EPYC chips are the first x86 server processors in volume production on TSMC’s 2-nanometer node.

The event drew commitments beyond Anthropic. OpenAI is optimizing workloads on the MI455X with Helios deployments starting in the fourth quarter of 2026, Oracle plans a 50,000-GPU cloud cluster in the third quarter, and Microsoft and Meta detailed their own plans.

The two firms also agreed to a multiyear engineering tie-up in which Anthropic’s Claude models will help build out AMD’s ROCm software stack. “Access to compute is central to keeping Claude at the frontier and meeting demand from our customers,” said Anthropic co-founder Tom Brown. AMD Chief Executive Lisa Su said the next phase of AI would span frontier models, agents and physical AI.

The first gigawatt of Anthropic capacity is not due until 2027 — leaving room for slippage before the revenue lands.