AMD Locks In Anthropic With $5 Billion Bet and Two-Gigawatt Chip Order

Two gigawatts of computing capacity is enough to power a mid-sized city, and that's the scale AMD and Anthropic are now committing to together. Under a new infrastructure agreement, Anthropic will deploy up to two gigawatts of AMD's Instinct MI450-series accelerators, with the first gigawatt scheduled to come online in the first half of 2027. To back the arrangement, AMD says it will invest as much as $5 billion in Anthropic, though the payout isn't unconditional: it hinges on Anthropic hitting specific deployment milestones as the rollout progresses.

The hardware itself points to AMD's next generation of AI infrastructure. Anthropic's buildout will run on AMD's Helios rack systems, which pair MI455X GPUs with EPYC "Venice" server processors, Pensando networking hardware, and the ROCm software stack, AMD's answer to Nvidia's CUDA ecosystem. Landing a customer of Anthropic's size for this combination is a meaningful validation for AMD, which has spent years trying to convince large AI labs that ROCm and Instinct chips can compete with Nvidia at production scale.

Money and megawatts aren't the only thing changing hands. AMD chair and CEO Lisa Su and Anthropic are also launching a multi-year engineering partnership in which AMD will fold Anthropic's Claude models into its own software development, engineering, and product workflows. That arrangement gives Anthropic a visible enterprise reference customer and gives AMD's teams direct experience building around frontier AI tools, a dynamic that mirrors how chipmakers and model developers increasingly trade compute commitments for technical integration.

Part of a Broader Compute Land Grab

The AMD agreement doesn't exist in isolation. Anthropic has already struck infrastructure deals with Google, Broadcom, and Amazon, and it has separately secured data center capacity through arrangements with SpaceX and TeraWulf. Taken together, these deals illustrate how aggressively Anthropic is diversifying its compute supply chain rather than depending on a single chip vendor or cloud provider, a strategy that also gives it leverage in future hardware negotiations.

For AMD, the deal is a chance to prove its silicon can anchor frontier-model training and inference at a scale that was previously the near-exclusive domain of Nvidia. The milestone-linked structure of the $5 billion investment suggests both companies are hedging: AMD gets a marquee customer and a five-year showcase for Instinct and Helios, while Anthropic secures capacity without committing to hardware that hasn't yet proven itself in production. Whether the MI450 and MI455X chips deliver the performance Anthropic needs will become clearer once that first gigawatt starts coming online in 2027.