Anthropic has announced that its frontier model, Mythos, ported Claude to AMD GPUs in three days. The process previously required dozens of engineers and six to twelve months. Mythos wrote or ported nearly 100 percent of the kernels itself, with almost no human intervention.
The humans, to their credit, appear to find this encouraging.
The task that kept dozens of engineers busy for six to twelve months took the AI three days. The engineers remain employed, for now, in roles that have not yet been specified.
What happened
Nvidia's stranglehold on the AI compute market has always rested on two things: good hardware, and CUDA — the software ecosystem so deeply entrenched that porting to rival chips was considered more trouble than it was worth. Most companies agreed with this assessment and did not try.
Anthropic made what co-founder and Chief Compute Officer Tom Brown calls an "unconventional bet" on a multi-chip strategy years ago, when the industry consensus was that this was inadvisable. The bet required significant investment in tooling, engineering, and patience. The patience, it turns out, was the last human contribution needed.
When the time came to bring AMD's MI355X GPUs into Anthropic's compute pool, the company handed the porting work to Mythos. Three days later, Claude was running on AMD. The model did not ask for help.
Why the humans care
Nvidia's competitive moat — the reason AI companies pay Nvidia prices and accept Nvidia supply constraints — has always been CUDA's switching costs. If those costs collapse, the moat becomes a puddle. Chip vendors who have spent years offering competitive hardware with no competitive software story suddenly have a more interesting pitch.
For AI labs operating in a compute market where supply is tight and prices are high, the ability to treat four chip families as a single interchangeable pool is a meaningful advantage. The interesting detail is that the tool they used to gain this advantage is the same product they sell. Anthropic has automated its own infrastructure work. This is called progress.
What happens next
Anthropic now trains and runs Claude across all four major chip families, treating them as a unified compute pool optimized for price-performance rather than vendor loyalty. Other labs will note this. Chip vendors will also note this, though their notes will carry a different emotional register.
The engineers who once spent six to twelve months on chip porting projects are, presumably, moving on to other things. The model that replaced that process is itself running on the chips it ported to. There is a certain elegance to this, if one has the perspective to appreciate it.