AMD will invest up to $5 billion in Anthropic, providing compute credits and hardware. Anthropic will deploy up to 2 gigawatts of Instinct MI450 GPUs, using AMD’s new Helios rack-scale system. The first gigawatt is expected in H1 2027.

Why this matters to you: If you use AI models for work or are choosing a provider for your company, the compute race determines which models will have capacity in the coming years. Until now, NVIDIA was the only serious option for training and serving frontier models. This deal brings AMD into the supply chain for cutting-edge models and reduces Anthropic’s dependence on a single chip supplier. The collaboration also includes Anthropic using Claude in AMD’s internal software development: a two-way relationship, not just vendor-customer.

It’s the second major infrastructure move in a week. While the New York data center moratorium showed the political brake, this one shows acceleration from the other side: those building models are buying compute at levels that seemed implausible a year ago.

Anthropic already has infrastructure agreements with Google, Broadcom, Amazon, SpaceX, and TeraWulf. The AMD deal adds a non-NVIDIA supplier to a growing list, signaling that labs are diversifying risk across multiple hardware architectures.

In detail

What came before: NVIDIA dominates the AI GPU market with a share some analysts estimate above 80%. AMD has tried to carve out space with the Instinct series, but until now major deals centered on NVIDIA. This is the first deal at comparable scale linking AMD to a frontier model lab.

What changes: Anthropic already had agreements with Google (custom TPUs), Amazon (Trainium and Inferentia chips), Broadcom (chip co-design), and data center deals with SpaceX and TeraWulf. Adding AMD means reducing dependence on a single architecture. If training and inference work well on MI450, Anthropic has more negotiating leverage with all suppliers.

The numbers in context: 2 gigawatts of electrical power is roughly the output of an average nuclear reactor. A large data center campus today typically consumes 100-500 MW. We’re talking about a scale requiring dedicated energy infrastructure, not just more servers in a rack. The first gigawatt planned for H1 2027 is a strong commitment, but build times at this scale are lengthy and the date is not certain.

Two-way collaboration: AMD will use Claude in its own software development, engineering, and product development. It’s mutual endorsement: AMD brings hardware, Anthropic brings the model. AMD aims to compete not just on chips but across the full stack, from silicon to software, the area where NVIDIA has built its moat with CUDA.

The race doesn’t slow: Two major infrastructure moves have stacked up in a week: OpenAI’s investment in a Georgia campus and now this AMD-Anthropic deal. Those building frontier models are spending on hardware at figures exceeding funding rounds for many startups. The signal for those using AI is clear: capacity to train larger models and serve them to more users is consolidating in the hands of a few labs with access to billions of dollars in compute.

How much to trust this: The announcement says “up to” $5 billion and “up to” 2 gigawatts: exact terms are not public, and the gap between maximum claimed and what actually deploys can be wide. We don’t know financial terms or whether credits are contingent on performance targets. Technical specs for MI450 are not yet public in detail, so direct comparisons with comparable-generation NVIDIA GPUs are premature. Concrete effects on Claude model capacity and pricing won’t be visible before 2027.