Anthropic ported Claude to AMD chips in three days using its Mythos model

Anthropic has launched its IPO roadshow, and its investor video revealed something it had not previously published: how the company actually operates internally at the hardware level. The most interesting part is the computing section, presented by co-founder Tom Brown. For anyone building infrastructure for LLMs, it offers a rare look inside.

Four chip families instead of one

Brown says Claude is now trained and served across all four major chip families: AWS Trainium, Google TPU, NVIDIA GPUs and AMD GPUs. He says this was made possible by an unconventional bet placed long ago. At the time, many believed running models on different chips was not worth the effort.

There were reasons for that skepticism. Each chip requires its own software stack and inference stack. That is expensive and complicated. If you have ever ported code from CUDA to another platform, you know what it entails: kernels—the low-level functions that actually perform matrix computations on specific hardware—have to be rewritten almost from scratch.

Anthropic has been doing this for years and, according to Brown, has been “incredibly successful at it.” There was already an indirect hint of its use of multiple chip platforms in September. In an analysis of three recent outages, the company described bugs that affected only TPU deployments. In other words, the same model really does run on different platforms, and those platforms can experience different bugs.

Mythos got Claude running on AMD in three days

The most striking figure concerns AMD. This year, Anthropic began working with this manufacturer's GPUs for the first time. Previously, adding a new chip platform took dozens of engineers months of work. This time, the task went to the Mythos model.

According to Brown, it got Claude running on AMD in three days with almost no human intervention. Claude ported or wrote almost 100% of the kernels.

For practitioners, that is an important signal. Porting kernels was considered precisely the kind of “expensive and tedious” work that keeps engineers tied to a single vendor. If an LLM really can handle most of that work, the cost of breaking free from a single GPU supplier falls. That also provides leverage in hardware pricing negotiations.

Compute multipliers and the discussion of self-improvement

Brown identifies the speed of discovering “compute multipliers” as a second source of advantage. Anthropic focuses on the resulting effective compute: physical computing capacity multiplied by the multipliers it discovers.

Here is the example from the video. Suppose training a model of a certain capability requires 100 units of compute. A 4x multiplier means the same model can be trained using 25 units. Alternatively, you could spend all 100 and get a smarter model that would previously have required 400.

According to Brown, the multipliers compound: each new one amplifies the effects of the previous ones, and that kind of advantage is difficult to replicate. Then comes the boldest claim. The company believes the process could accelerate as it approaches recursive self-improvement, when Claude will be able to discover multipliers entirely on its own.

What to keep in mind

First, these are investor materials, not an independent report. Figures such as “three days” and “almost 100% of the kernels” come from one of the founders and have not been independently verified. Details are missing, too: exactly which kernels were involved, how close performance is to NVIDIA, and how many hours of human review were still required.

Second, recursive self-improvement remains a statement of future ambition here. It is not a result anyone has measured. It should be treated as the company's strategy and part of its IPO pitch.

Third, there is no official transcript of the video. Simon Willison produced a transcript himself using MacWhisper and published it, so quotations should be checked against it.

The practical takeaway for teams building their own inference infrastructure: if you have been putting off support for a second vendor because of the cost of porting kernels, it is time to revisit that estimate. Try giving a capable model a small kernel with tests and calculate how many engineering hours it actually saves. Even if your result is more modest than three days, the threshold for what is considered “expensive and not worth the effort” has already shifted.