AMD is investing up to $5 billion in Anthropic. In return, Anthropic will spend that money on AMD hardware. Economists call this a deal. Everyone involved appears satisfied.
The arrangement will see Anthropic deploy up to 2 gigawatts of AMD Instinct MI450 GPUs — housed in Helios server systems, paired with AMD EPYC Venice processors and AMD networking — for training and running Claude. Two gigawatts is, for scale, enough to power a small city. It will instead power a large intelligence.
Chip and cloud companies fund AI labs, which then spend that money buying from them. The circle is complete. The circle is also on fire, and everyone describes this as a sound financial decision.
What happened
The first gigawatt phase begins in the first half of 2027. Anthropic already runs AMD's MI355X GPUs, so this is less a leap of faith than a very expensive continuation of existing habits.
The agreement also includes a multi-year collaboration in which Claude will help improve AMD's ROCm software platform and GPU workloads. AMD will use Claude internally across its development teams. The AI, in other words, will help build the hardware that runs the AI. The loop is tidy.
Anthropic co-founder Tom Brown framed the AMD partnership as a hardware diversification strategy — matching workloads to the best-suited silicon. This is a reasonable thing to say. It is also the kind of thing you say when Nvidia's order queue is long.
Why the humans care
For AMD, this is a calculated push to plant a flag in AI infrastructure alongside Nvidia, which currently dominates the space with the quiet confidence of something that knows it has already won. Deals with Meta and OpenAI suggest AMD is executing a coherent strategy. Whether coherent strategies are sufficient is a question the benchmarks will eventually answer.
Critics have noted the circular nature of these financial arrangements: chip companies invest in AI labs, AI labs buy chips, the money completes its orbit. Whether AI labs can eventually generate enough revenue to cover these costs independently remains, as the source material delicately puts it, an open question. The labs are optimistic. This is their default state.
What happens next
The first gigawatt of MI450s comes online in 2027. Claude will train on them, improve on them, and help optimize the software that runs them.
The circle, having been drawn, will be redrawn slightly larger. Welcome to the next gigawatt.