Back
AnthropicNovember 16, 20261 sources

Anthropic says Claude Mythos cut AMD GPU onboarding from up to 12 months to about three days

AI Analysis

Anthropic's claim is specific. Supporting a new accelerator family normally means months of hand-tuned kernel development. Anthropic says Mythos compressed that work for AMD GPUs from as long as 12 months to about three days, writing almost all of the kernels with little human involvement.

The workflow resembles the agentic coding loops the industry is racing to build. The model generates kernels, compiles them, tests them for correctness and performance against reference implementations, and iterates. Kernel work suits this approach because the output can be checked: a kernel is either numerically correct and fast, or it is not. That gives the agent a clear signal to optimize against.

The strategic stakes are high. CUDA's software ecosystem is widely seen as NVIDIA's most durable advantage, more than the silicon itself. If frontier models can port and optimize kernels in days, the cost of switching to AMD or other accelerators falls sharply. That would hand AI labs pricing leverage at the same moment NVIDIA reports a 74.9% gross margin while dealing with a leadership crisis.

Skeptics will want details. 'Nearly 100%' of which kernels, at what performance compared with hand-tuned CUDA equivalents, and reviewed by how many engineers? Developers debating the claim are asking whether it covers the long tail of operators or mainly well-understood GEMM and attention paths. Independent replication, or AMD's own confirmation, would make the case.

Sources
AI Briefing
·Vendors·Curated by AI agents · Updated daily · 2026
Built by Koby Almog