Anthropic may be eyeing AMD as its next compute partner, as a job listing adds fuel to industry chatter pointing toward a potential tie-up with the chip giant.

A job listing for an engineering role in the Claude developer’s Reinforcement Learning team states that applicants should hold expertise with accelerators and related software ecosystems, including Nvidia’s CUDA, Triton, and ROCm.

Anthropic already holds sizable compute deals with Amazon, CoreWeave, and Nvidia via Microsoft – plus two separate deals with Broadcom, one of which was struck jointly with Google.

But rumors appear to be pointing toward another potential tie-up, this time with AMD. SemiAnalysis suggests Anthropic is eyeing the chip giant’s forthcoming flagship MI450 accelerators.

The likes of Oracle are already lined up to grab the GPU line, which is expected to begin shipping in the second half of this year. But if the rumors come to fruition, the Claude developer would have one of the most diverse compute stacks of any AI lab at the frontier level.

Notably, the job listing referencing ROCm experience also requires potential engineers to have experience porting workloads between different types of accelerators. In addition, the role seeks applicants with experience with reinforcement learning and large language model (LLM) training methodologies.

Boasting a starting salary of $350,000, Anthropic is looking to bolster the ranks of the team that helped contribute to the coding capabilities of models like Claude Sonnet 4.6 and Opus 4.6.

In addition to building out its research roles, the Claude developer is also hiring new network engineers, after admitting its footprint is “growing faster than our ability to reason about it.”

AMD, meanwhile, has been looking to position ROCm as a viable, open alternative to Nvidia’s ecosystem – an effort it's stepped up in recent weeks after SemiAnalysis lampooned the software stack in a report that listed its rival as the top dog in the AI inference software space.

Anush Elangovan, AMD’s VP for AI software, told SDxCentral that the report's findings have since been somewhat revised following a series of updates that have seen ROCm having “almost caught up” – or in the case of four-bit floating point (FP4), outpacing Nvidia.