Anthropic says AI cut an AMD chip port from months to three days

Anthropic says a job that once demanded months of engineering took its own AI about three days: getting one Claude model’s prompt processing working on AMD chips. Disclosed in investor materials, the result offers a concrete example of AI reducing the software work that makes switching away from Nvidia difficult.

The bottleneck was software

The starting problem was not simply finding another supplier’s chips. Each hardware platform needed its own custom software and systems for running models. Moving to a new platform could occupy dozens of engineers for months, according to Anthropic’s account in an investor presentation covered by The Decoder.

That work helps explain Nvidia’s software advantage. CUDA, its software ecosystem, makes Nvidia hardware useful to developers, but code built around it takes effort to move elsewhere. The obstacle is therefore more than a comparison between processors: customers also have to account for the engineering required to make an alternative work well.

Anthropic co-founder and Chief Compute Officer Tom Brown described the company’s early decision to support multiple chip families as an unconventional bet. Rather than avoid that engineering burden, Anthropic says it invested in the work for years. The AMD effort became a test of whether its own models could now do much of it.

Mythos takes on the AMD port

For the AMD work, Anthropic turned to Mythos, its frontier AI model. The company says Mythos wrote or adapted nearly all the kernels—the chip-specific code involved in the port—with almost no human intervention. Brown presented that result as a sharp departure from the company’s earlier approach to bringing new hardware into use.

The comparison in Anthropic’s materials is about reaching a usable level of performance: roughly three days with Mythos, against six to twelve months under traditional engineering approaches. These are company-supplied figures, not independently verified results. They measure how quickly the software became useful on the hardware, rather than how quickly the model answers a user.

The important boundary sits in the slide’s footnote. As TIKR explains, the three-day result covered prompt processing for one Anthropic model on AMD MI355X GPUs. That is narrower than saying every Claude workload moved to AMD in three days. It is a specific porting result, not a measure of the effort required to move an entire AI operation.

Four platforms, one claimed compute pool

Anthropic places the AMD result inside a larger operating strategy. Brown says the company now trains and runs Claude across Amazon’s Trainium chips, Google’s TPUs, Nvidia GPUs and AMD GPUs. Anthropic describes that capacity as one interchangeable pool, spanning different vendors, chip architectures and locations, rather than four isolated sources of computing power.

Its presentation shows that pool serving research, model development, customer demand and internal Claude use. The intended benefit is flexibility: different chips can suit different jobs, and a multi-chip approach lets the company consider price and availability rather than choose solely around one supplier’s platform. Faster porting would reduce one cost of exercising that choice.

The unresolved issue is how far that advantage travels beyond this example. The disclosure supports a possible reduction in switching costs, not the disappearance of Nvidia’s software ecosystem. For Anthropic’s broader claim of interchangeable hardware, the distinction matters: making one task usable quickly is evidence of progress, while treating multiple platforms as substitutes is the larger operational ambition.