Lilith.
⌕
Editorial illustration: Anthropic says Claude ported a model to AMD in 3 days, aiming at CUDA’s moat
Lilith illustration · editorial remix

Anthropic claims in its IPO roadshow that Mythos brought prompt processing for one Claude model to AMD MI355X in 3 days. The company is selling investors not only an AI model, but also the ability to switch compute suppliers faster.

Mythos wrote almost every kernel for a narrow AMD port

Co-founder Tom Brown said Anthropic now trains and serves Claude across four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. Bringing up a new platform previously took dozens of engineers months, he said. This time Mythos allegedly wrote or ported almost 100% of the kernels with minimal human intervention.

A footnote in the investor presentation matters. The 3 days cover prompt processing for one Anthropic model on AMD MI355X at a useful level of performance. They do not describe a complete move of training and inference for the whole model family. The comparison with 6 to 12 months is also based on Anthropic's own historical experience.

Portability across chips gives Anthropic leverage when buying compute

Nvidia's lead is protected partly by the CUDA ecosystem. Moving to another accelerator requires new kernels, libraries, tuning and testing. If an agent can cut much of that work from months to days, a lab can allocate workloads according to price, availability and purpose.

Anthropic describes its chip estate as a fungible pool and assigns architectures to training or inference according to fit. Its IPO site also reports more than a fourfold improvement in compute efficiency from the first quarter of 2025 to the third quarter of 2026. The metric is revenue divided by compute and infrastructure expense, so rapid revenue growth influences it alongside engineering.

A 3-day demonstration has not dismantled the CUDA ecosystem

Anthropic has not published a reproducible technical method, throughput comparisons with Nvidia, the experiment's cost or the amount of human verification. A roadshow is designed to persuade investors, and the claim has not been independently verified.

The result also rests on years of multi-chip work and an internal orchestration and testing layer. A team without that infrastructure cannot expect the same outcome merely by handing kernel source code to an agent.

Production traffic and a repeat port will show whether the moat shrank

The useful signals will be the share of real traffic running on AMD, comparable throughput and cost, stability on longer workloads and another port to a new architecture. One successful prompt-processing bring-up demonstrates possibility, not a general method.

If Anthropic repeats the result and publishes technical data, CUDA will face a new kind of competitor: automated software portability. Until then, the 3-day figure is exactly what a roadshow needs, an impressive number on a screen in front of investors.

Lilith's verdict

Anthropic sent Claude across the swamp between two chip stacks and says it reached the far bank in 3 days. Nvidia should worry when production traffic follows and an independent observer starts the stopwatch.