Anthropic is evaluating AMD hardware for training and running its Claude language models, according to a report from research firm SemiAnalysis that cited AMD’s public source code. Any commercial relationship is still at an early stage, with testing and software optimization underway rather than a full deployment.
GitHub code lists Anthropic as an evaluation customer
Public GitHub source code referenced in the report shows Anthropic listed as an AMD evaluation customer. It was also assigned priority points at the same level as hyperscale data center operators including Meta, a detail that suggests AMD is treating the account as strategically important during the evaluation process.
If the two sides move forward, the report said the arrangement could center on Anthropic buying or leasing MI300-series hardware or later AMD AI accelerators. AMD’s role would not stop at supplying chips. The company would also need to provide customized technical support to help Anthropic tune model performance.
Software and networking progress remains a key condition
SemiAnalysis said Anthropic is still in the evaluation phase, and any wider adoption will depend on AMD’s progress in software fixes and networking coordination.
The report pointed to continuous integration capacity limits as one of AMD’s main constraints. That issue is said to affect Pollara network interface cards under Kubernetes orchestration, an area described as critical for the global rollout of large-scale machine learning inference. SemiAnalysis also said AMD needs to improve its ROCm software ecosystem, or the timeline for a final contract could be affected.
AMD’s annual flagship event, Advancing AI 2026, is scheduled for July 22-23 at the Moscone Center in San Francisco. Whether the company formally announces any relationship with Anthropic there is drawing attention.
Anthropic still relies heavily on NVIDIA infrastructure
While evaluating AMD, Anthropic already has a deep technical and capital relationship with NVIDIA. The report said the two sides previously signed compute agreements worth tens of billions of dollars and used hundreds of thousands of H100 and Blackwell chips for deep optimization.
Anthropic also has an indirect infrastructure relationship involving SpaceX through the leasing of the Colossus super data center in Memphis. According to the report, that facility offers more than 220,000 NVIDIA chips and over 300MW of power capacity.
Based on that setup, Anthropic still operates primarily within the NVIDIA ecosystem at this stage, while the SpaceX-linked arrangement is focused on data center hardware capacity. If AMD eventually wins a broader role, it would mark a notable step in its effort to compete for top-tier AI customers.