AMD to invest up to $5 billion as Anthropic plans 2GW GPU rollout

AMD plans to invest up to $5 billion in Anthropic, and Anthropic plans to deploy up to 2 gigawatts of AMD Instinct GPU capacity for training and running its Claude models. AMD said the first 1-gigawatt phase is scheduled to begin in the first half of 2027, giving the chipmaker another high-profile AI lab customer as it tries to narrow Nvidia’s lead in accelerator deployments.

The companies did not disclose the structure of the investment, Anthropic’s valuation, purchase commitments, delivery schedule beyond the first phase, or the expected cost of the systems. AMD described the arrangement as a strategic partnership, tying its capital to future deployment of its AI infrastructure by one of the better-funded frontier model companies.

According to AMD, Anthropic will use AMD Instinct MI450 GPUs in Helios server systems. The systems will combine MI455X GPUs, AMD EPYC “Venice” processors and AMD networking technology. Anthropic already uses AMD’s MI355X GPUs, according to the companies, so the deal extends an existing hardware relationship rather than starting one from zero.

Claude becomes part of AMD’s software pitch

The agreement also includes a multi-year effort to apply Claude to AMD’s own software stack and GPU workloads. AMD said Anthropic will work with it on improving ROCm, the software platform AMD needs developers to treat as a credible alternative to Nvidia’s CUDA ecosystem.

AMD also said it will use Claude internally across its development teams. The companies did not disclose how broadly Claude will be deployed inside AMD, whether the internal usage is tied to spending commitments, or what specific engineering work the model will support.

Anthropic co-founder Tom Brown said working with multiple hardware providers lets the company match workloads with the hardware best suited to them. That framing is useful for AI labs that want bargaining power and supply diversity, but it also reflects a more basic constraint: frontier model training and inference require more power, more chips and more data center capacity than any single supplier can provide on demand.

AMD adds another AI lab deal

For AMD, the Anthropic agreement follows similar AI infrastructure partnerships with OpenAI and Meta. The pattern is clear: AMD is using long-term supply relationships and capital commitments to secure demand for multiple generations of Instinct GPUs.

That approach gives AMD named customers in a market where Nvidia has set the default architecture for large-scale AI training. The harder test is execution: shipping enough accelerators, pairing them with competitive networking and software, and persuading model builders that ROCm can support demanding production workloads at scale.

The deal also adds to criticism of circular financing in AI infrastructure. Chip and cloud suppliers invest in AI companies, and those AI companies then spend heavily on compute from the same suppliers or their partners. The arrangement can accelerate deployment and lock in customers, but it leaves open whether AI labs can generate enough revenue to pay for these compute commitments without continued vendor financing.

AMD and Anthropic have put a large number on the table, up to $5 billion, and an even larger power target, up to 2 gigawatts. What they have not provided is the commercial detail needed to judge how much of that turns into shipped infrastructure, recognized revenue or sustainable model economics.

This story draws on original reporting from The Decoder.