Anthropic claims in an investor presentation that Mythos ported or wrote nearly all kernels needed to bring Claude to usable performance on AMD GPUs in three days. The displayed figures concern prompt processing for one model on MI355X, compared with six to twelve months for traditional accelerator onboarding. The company says it treats NVIDIA, AMD, Google TPU and AWS Trainium capacity as a shared compute pool. AI may lower switching costs, but these figures are not independently verified and do not mean an entire training and inference system migrated in three days.