The contours of AI infrastructure spending just got a lot more concrete—and a little more circular. Anthropic and AMD have struck a deal valued at up to $5 billion that will see the AI lab deploy a staggering 2 gigawatts of AMD Instinct MI450 GPUs inside Helios server systems. The first gigawatt phase is slated to come online in the first half of 2027, pairing the MI455X GPUs with AMD’s upcoming EPYC “Venice” processors and the company’s own networking fabric.
This isn’t Anthropic’s first dance with AMD silicon. The company already uses the MI355X GPUs to run its workloads. But the new partnership formalizes a massive, multi-year expansion that goes beyond just buying hardware. Both companies plan to put Claude to work inside AMD, using the AI to help its dev teams improve the notoriously finicky ROCm software platform. Anthropic co-founder Tom Brown framed the move as a pragmatic hedge. “Working with different hardware providers lets us match workloads to the best-suited hardware,” he said, a polite way of saying not every inference job needs a cluster of $40,000 Nvidia chips.
For AMD, the deal is a loud shot across Nvidia’s bow. Lisa Su’s company has been methodically stitching together similar hyperscale partnerships, locking in agreements with Meta and OpenAI to ensure that the financial gravity of the AI boom doesn’t orbit exclusively around one supplier. But let’s not pretend the accounting here is straightforward. The arrangement invites a familiar criticism: cloud providers and chipmakers pump billions into AI labs, only for those labs to hand the cash right back to buy the providers’ chips and compute. It’s a closed loop of revenue recycling that inflates top-line numbers without necessarily proving sustainable unit economics.
Whether Anthropic can eventually sell enough Claude subscriptions or API access to justify a $5 billion hardware tab is the multi-trillion-dollar question nobody can answer yet. The first gigawatt of MI450 power is still two years out. A lot can happen in that time—starting with the possibility that the AI models themselves might rewrite the economic assumptions this deal was built on.
