- AMD is investing up to $5 billion in Anthropic; Anthropic plans to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs.
- The first gigawatt phase starts in the first half of 2027, using Helios server systems that pair MI455X GPUs with AMD EPYC “Venice” CPUs and AMD networking.
- Anthropic already runs on AMD’s MI355X GPUs, and the two companies will use Claude to improve AMD’s ROCm software stack.
- The structure mirrors AMD’s deals with Meta and OpenAI, and draws the same criticism that suppliers are funding their own customers.
What Happened
AMD will invest up to $5 billion in Anthropic, and Anthropic in return plans to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs to train and serve its Claude models, The Decoder reported on July 22, 2026. The first one-gigawatt phase is scheduled to begin in the first half of 2027.
Anthropic co-founder Tom Brown framed the arrangement as a hardware-diversification move, saying that working with multiple hardware providers lets the company match each workload to the silicon best suited to it. Anthropic already runs production workloads on AMD’s current-generation MI355X accelerators.
Why It Matters
Two gigawatts is a capacity commitment, not a chip order — it is roughly the scale at which frontier labs now describe their infrastructure, and it puts AMD in a tier of supply relationships that Nvidia has largely held alone. AMD has now signed comparable arrangements with Meta and OpenAI, which turns the Instinct line from a challenger part into a second sourcing option that labs can price against.
The pattern also raises the financing question that has followed every one of these deals. As The Decoder puts it, critics point to “the circular nature of these agreements”: chip and cloud vendors invest in AI labs, and the labs spend that capital buying from the same vendors. Whether the labs can eventually fund this compute from operating revenue rather than strategic investment remains unresolved.
Technical Details
The deployment uses AMD’s Helios rack-scale systems, which pair Instinct MI455X GPUs with AMD EPYC “Venice” server processors and AMD’s own networking silicon — a full-stack configuration rather than a GPU-only purchase, which matters for the interconnect bandwidth that large-scale training runs are bounded by.
The second half of the agreement is a multi-year software effort: Anthropic and AMD will use Claude to improve ROCm, AMD’s open compute platform, and to optimize GPU workloads. AMD will also deploy Claude internally across its own development teams. ROCm’s maturity relative to CUDA has been the recurring obstacle for AMD in AI training, and using a frontier coding model to close that gap is the explicit bet here.
The 2-gigawatt figure describes power capacity across the full program rather than a single site, and it is split into phases, with only the first gigawatt tied to a stated 2027 date.
Who’s Affected
For AMD, this is the most direct challenge yet to Nvidia’s position as the default training supplier, and it comes with a reference customer whose workloads are public enough to be scrutinized. For Anthropic, multi-vendor sourcing across AMD, Nvidia, Google TPUs, and Amazon Trainium reduces exposure to any single supply queue.
Developers on ROCm are the second-order beneficiaries: improvements funded by this partnership land in an open platform rather than in a private fork, which is the practical difference between this and a standard volume purchase agreement.
Nvidia is affected less by the volume than by the precedent. Anthropic is the second frontier lab after OpenAI to commit gigawatt-scale capacity to Instinct hardware, and each such commitment weakens the argument that CUDA lock-in makes alternative silicon impractical for training rather than only for inference. Cloud providers that resell AMD capacity — and the power and construction suppliers behind the Helios deployments — sit downstream of the same decision.
What’s Next
The first gigawatt of MI450-series capacity is due in the first half of 2027, which is the first checkpoint where the announcement becomes verifiable. Neither company has disclosed the tranche structure of AMD’s investment or the conditions attached to it, and the deal’s economics depend on details — pricing, warrants, delivery schedules — that have not been made public. AMD’s next quarterly report is the earliest place those terms are likely to surface.
