Anthropic says it used an internal frontier model called Mythos to bring Claude onto AMD GPUs in three days, a task it says would normally occupy dozens of engineers for six to twelve months. The model reportedly wrote or ported nearly all the required compute kernels with little human intervention.
The company now says it trains and runs Claude across four major chip families, treating different vendors and locations as a flexible compute pool. If the AMD result generalizes, AI-generated low-level software could weaken one source of Nvidia’s advantage: the mature CUDA ecosystem that makes alternative hardware expensive to adopt.
The numbers come from an Anthropic investor presentation and have not been independently verified. Porting one workload also does not prove equal cost, reliability or performance across every model and chip. Still, the claim points to a consequential use of coding models inside AI infrastructure itself. Easier portability would let large labs choose hardware by price, supply and workload rather than committing as strongly to a single vendor’s software stack.