AMD will supply Anthropic with up to 2 gigawatts of Instinct MI450-series GPU capacity and plans to invest as much as $5 billion in the Claude maker, the companies said in a joint announcement Wednesday.
The deal gives AMD another marquee customer for its next AI accelerator platform and gives Anthropic another large compute source for training and serving Claude. The first gigawatt is expected to be available in the first half of 2027, using AMD’s Helios rack-scale systems, according to the announcement.
The arrangement also confirms something AMD had not previously said in public: Anthropic is already using AMD Instinct MI355X GPUs. Until now, Anthropic’s disclosed chip suppliers included Nvidia, Amazon, and Google.
What AMD is selling Anthropic
Anthropic’s planned deployment uses Instinct MI455X GPUs with Epyc “Venice” CPUs, Pensando networking, and AMD’s ROCm software stack inside Helios racks, according to AMD and Anthropic. Each rack contains 72 accelerators, 31 TB of HBM4 memory, and up to 2.9 FP4 exaFLOPS of inference compute.
Those numbers are vendor claims, not independent benchmarks. They do show the shape of AMD’s pitch: sell the whole rack, not just a card, and make the interconnect, CPU, GPU, and software story look less like a parts bin.
The companies also said they have started a multi-year engineering collaboration. Anthropic will use Claude to help tune workloads for Instinct GPUs and speed ROCm development, while AMD will deploy Claude across its engineering and product groups.
Tom Brown, Anthropic’s co-founder and chief compute officer, said in the companies’ announcement that using multiple hardware types helps Anthropic assign workloads to suitable systems.
AMD is buying into its customer
The equity piece is notable because AMD is not only selling chips to Anthropic. It is also taking a stake in the buyer, with an investment of up to $5 billion.
That is the reverse of AMD’s recent OpenAI arrangement, under which OpenAI received warrants for up to roughly 10% of AMD’s stock. In the Anthropic deal, AMD is the investor.
AMD’s first gigawatt for Anthropic is not first in line for Helios. OpenAI’s first gigawatt of AMD capacity is scheduled for the second half of 2026, while Oracle is due to begin deploying 50,000 MI450 GPUs this quarter, according to earlier company announcements cited in the reporting.
The leak trail was short
SemiAnalysis reported Sunday that a YAML configuration file in a public GitHub repository belonging to an AMD senior director listed Anthropic as a customer with the maximum 30 priority boost points. SemiAnalysis said that put Anthropic in the same priority tier as hyperscale customers such as Meta.
The companies’ announcement now makes the customer relationship official, and states that the Helios deployment builds on Anthropic’s current use of MI355X GPUs.
Anthropic’s compute supply is increasingly spread across rival chip ecosystems. The company already uses Nvidia GPUs, Amazon Trainium chips under Project Rainier, Google TPU capacity supplied by Broadcom from 2027, and Nvidia compute rented from SpaceX’s Colossus 1 data center, according to the reporting. AMD now gets a larger seat at that table, provided it can deliver the racks, power envelope, and ROCm maturity on the schedule it has advertised.
This story draws on original reporting from Tom's Hardware.
