Anthropic told investors that its flagship model Claude is now running on AMD Instinct GPUs after only three days of work. The company says the port was done by its own frontier model, Mythos, which wrote almost all the required kernels without human help. In the past, moving a model to a new accelerator would tie dozens of engineers to a project for six to twelve months.
The claim comes from an investor presentation released in November 2026. Anthropic’s co‑founder and chief compute officer Tom Brown said the firm now trains and serves Claude on four major chip families: Nvidia GPUs, AMD GPUs, Amazon Trainium chips, and Google TPUs. The multi‑chip approach lets the lab shift work to the cheapest or most available hardware, something that has been difficult because Nvidia’s CUDA software stack traditionally locks developers into its chips.
Anthropic says the three‑day AMD switch used the MI355X GPU and focused on prompt‑processing performance. The company treats compute from different vendors as a single pool, allocating capacity wherever it gets the best price‑performance ratio. It also plans to install up to two gigawatts of AMD MI455X GPUs in its Helios rack systems, with the first gigawatt slated for the first half of 2027.
Why this matters
If Anthropic’s numbers hold up, companies could move AI workloads between Nvidia and AMD much faster, reducing dependence on Nvidia’s high‑margin CUDA ecosystem. That may open the market for cheaper or more available chips, especially while GPU supply stays tight.
Why it matters
For businesses that run large language models, faster hardware switches could lower cloud costs and avoid bottlenecks caused by Nvidia’s dominant software stack. However, the claim has not been independently verified, so the real speed and cost benefits remain uncertain.
