Anthropic's Claude Mythos has accomplished what would normally demand dozens of engineers and six to twelve months — nearly 100% of the low-level kernel code for AMD's MI355X chips — in a mere three days, with almost no human intervention. After porting, the model's runtime performance trails Nvidia's CUDA-optimized counterpart by only 8–12%, a gap already narrow enough to satisfy commercial requirements. This breakthrough dramatically lowers the barrier to cross-chip adaptation, and Anthropic has secured a computing power supply partnership with AMD totaling up to 2 gigawatts. That said, the advance thus far covers mainly the inference prefill stage; Nvidia retains a clear edge in areas such as multi-GPU collaboration for ultra-large-scale distributed training. The CUDA ecosystem has hardly been toppled outright, but a crack has been opened in its software moat — and the computing power market now looks poised to shift toward a more open and diversified landscape.
