Anthropic says its AI adapted to AMD GPU in 3 days, shaking Nvidia's CUDA lead

Nvidia's long-standing lead in the CUDA ecosystem is facing growing challenges. A slide from Anthropic's IPO roadshow shows that as high-performance AI models get better at adapting to different computing platforms on their own, GPUs are turning into interchangeable general-purpose computing resources. Anthropic's Claude Mythos Preview model took only about 3 days to figure out how to run another model on AMD's MI355X GPU. Traditionally, this kind of work takes 6 to 12 months, so the barrier to switching platforms has dropped sharply. Still, it may be too early to say CUDA's edge is over. CUDA comes with a huge developer community, rich libraries, and years of fine-tuning. Nvidia also gives free GPUs and training to universities, and ties its NCCL communication library closely to proprietary hardware like NVLink and InfiniBand, boosting multi-GPU performance while raising the cost of switching away.