← Back to feed

AMD and Anthropic announced a strategic partnership on July 22 that gives AMD its first meaningful share of the compute stack running Claude. AMD will invest up to $5 billion in Anthropic as a strategic equity holder, and Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in Helios rack-scale solutions. The first gigawatt enters deployment in the first half of 2027.

The deal has two sides that reinforce each other. AMD gets hardware orders large enough to validate Helios at scale and a commitment from Anthropic to help optimize workloads for AMD’s Instinct GPUs and accelerate ROCm software development. Anthropic gets $5 billion in equity capital and a fourth hardware vendor capable of running frontier training at gigawatt scale.

What Runs Where

Anthropic’s compute portfolio is now genuinely multi-vendor. Each agreement reflects a different trade-off:

• Google + Broadcom: Largest deal, TPU-based, multi-year, with Google financing $5B of Anthropic’s Texas data center. Primarily training.
• Amazon: 5GW of Trainium chips; Anthropic is AWS’s signature AI partner.
• CoreWeave + SpaceX (Colossus): NVIDIA GPU deployments, fast turnaround capacity.
• AMD (new): 2GW of MI455X in Helios rack-scale solutions, H1 2027 start.

Anthropic’s Chief Compute Officer Tom Brown described the rationale as mapping “the right workloads to the right hardware.” At the frontier, different hardware architectures have different price-performance profiles across training, inference, and fine-tuning. A multi-vendor stack reduces dependency risk and creates negotiating leverage with each vendor.

Why AMD Made the Bet

The $5 billion equity commitment mirrors NVIDIA’s playbook of investing in customers to guarantee hardware adoption at scale. NVIDIA put $2 billion into Nebius and has taken equity stakes across the neocloud ecosystem; AMD is doing the same at a larger number with a higher-profile counterparty.

For AMD, landing Anthropic as a 2GW customer is the strongest production-scale validation Helios can get. The previous Helios reference deployment — a GLM-5.2 benchmark run on MI355X hardware — posted 2,626 tokens per second at roughly half the per-token cost of Blackwell equivalents for inference. That showed AMD could compete on inference economics. Training at 2GW scale will test whether the architecture holds at the workloads that matter most for frontier development.

The hardware stack for this deployment: AMD Instinct MI455X GPUs (part of the MI450 Series), AMD EPYC “Venice” CPUs, AMD Pensando networking, and ROCm open software. Anthropic already runs MI355X in production, so the MI455X upgrade is an in-family migration rather than a cold start.

Claude Accelerates ROCm

A secondary commitment in the announcement is commercially significant for the open GPU software ecosystem. Anthropic will use Claude to optimize workloads for AMD Instinct GPUs and accelerate ROCm development. AMD’s software stack has historically been the primary reason developers defaulted to NVIDIA’s CUDA. A frontier AI lab actively investing in ROCm tooling changes the incentive structure for the open-software GPU ecosystem.

AMD will also broadly adopt Claude across its engineering and product development teams — a form of enterprise revenue that reinforces the compute relationship with a software contract.

Key Numbers

• AMD equity investment: up to $5B (strategic, in the future)
• Compute deployment: up to 2GW AMD Instinct MI455X in Helios rack-scale systems
• First GW deployment: H1 2027
• Hardware stack: MI455X + EPYC “Venice” + Pensando networking + ROCm
• Prior MI355X result: 2,626 tokens/second on GLM-5.2, 2x cheaper than Blackwell inference
• Anthropic compute suppliers now: Google/Broadcom, Amazon, CoreWeave, SpaceX, AMD