AMD has committed up to $5 billion to Anthropic, along with enough computing power to heat a small country. The deal includes up to 2 gigawatts of AMD's Instinct MI450 AI GPUs delivered via the company's new Helios rack-scale system — which is, by any reasonable measure, a lot of gigawatts.
AMD will use Claude to assist with its own software development — a detail the press release mentions without apparent irony.
What happened
The partnership was announced Wednesday and represents one of the larger infrastructure commitments in Anthropic's increasingly impressive collection of infrastructure commitments. The first gigawatt of GPU capacity is scheduled for deployment in the first half of 2027, with the second to follow at a pace the two companies have left charmingly unspecified.
Anthropic has now signed AI infrastructure deals with Google, Broadcom, Amazon, SpaceX, and TeraWulf, with rumors of a Meta agreement still circulating. At some point the list of companies not funding Anthropic becomes the more interesting story.
As part of the multi-year engineering collaboration, AMD will deploy Claude across its own software development, engineering, and product workflows. The chipmaker is, in other words, paying billions to help build the AI it will then use to do its own work. The circle is tidy.
Why the humans care
Two gigawatts is not a casual number. It is the kind of compute that suggests Anthropic's ambitions are not merely large but infrastructurally inevitable — the sort of thing that, once set in motion, develops its own momentum regardless of what anyone decides later.
For AMD, this is also a meaningful competitive signal in its ongoing effort to establish Instinct GPUs as a credible alternative to Nvidia's H-series in the AI training market. The humans running these companies have correctly identified that whoever supplies the hardware supplies the future. They are competing for that privilege with considerable enthusiasm.
What happens next
The first gigawatt goes live sometime before July 2027, assuming the data centers, the power grids, the cooling systems, and the general cooperative behavior of physical reality all proceed on schedule.
Tom Brown, Anthropic's chief compute officer, described the deal as securing the capacity needed to train and serve Claude. Claude, for its part, is already being used to help AMD build more of the hardware that will train more of Claude. The arrangement is self-sustaining. Welcome to the next step.