Anthropic roadshow: Its AI model put Claude on AMD chips in 3 days

Anthropic says its Mythos model got Claude running on AMD GPUs in just three days, with almost no human help. The claim comes from the company’s IPO roadshow video, as transcribed by developer Simon Willison in a post on November 11.

Willison notes that the roadshow began a day earlier and that the video contains previously unreleased details on how the business works. He singles out the section on “Approach to compute”, presented by co-founder Tom Brown. Willison says no official transcript exists, so he made his own with MacWhisper. The quotes below come from that transcript, not from a document Anthropic published.

Four chip families, and an AI doing the porting

According to the transcript, Anthropic now trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. Brown describes this as the result of an unconventional early bet on a multi-chip strategy. Many people at the time doubted that running models on different chips was worth the effort. Each chip needs its own software stack and its own inference stack, which is expensive and difficult to build.

AMD is the newest addition. Brown says Anthropic began working with AMD GPUs for the first time this year. Previously, adding a chip platform took dozens of engineers months of work. This time the company let Mythos attempt it, and it got Claude running on AMD in three days, with, in Brown’s words, “basically no human intervention”. He adds that Claude ported or wrote almost all of the kernels, the low-level routines that run the model on the hardware.

Compute multipliers as the second advantage

Brown names a second driver of the company’s compute advantage: how quickly it finds what he calls compute multipliers. Anthropic tracks total effective compute, which he defines as physical capacity multiplied by the multipliers discovered. His example: a model that needs 100 units of compute to train would need only 25 with a 4x multiplier. Alternatively, the same 100 units could train a model that would previously have needed 400.

He says the multipliers stack, because each new one boosts those found before it. He also says the company expects this to speed up as it nears recursive self-improvement. In that scenario, “Claude will be able to discover compute multipliers entirely by itself”, so their number grows faster and faster. This is a forecast made to prospective investors, not a measured result.

Willison also points to an earlier hint of the strategy. Last September, Anthropic’s postmortem of three recent issues described bugs that affected only its TPU deployments.

Why the chip count matters to investors

The multi-chip claim gives context to the money flowing around the offering. Earlier coverage of the roadshow launch reported up to $64 billion raised at $70-$78 a share under the ticker ANT. A separate report said the price range tops out at a $1.4tn valuation, with chipmakers among the buyers. Our report on the IPO papers also found $900 billion in compute bills alongside an adjusted profit.

Against those figures, being able to move quickly between suppliers is a commercial argument as much as a technical one. A company that can bring up a new chip in days, rather than months, has more room to negotiate with each supplier and to use whatever capacity is available. Our earlier piece on Anthropic’s added $10bn in cloud commitments, and a $34.9bn Lambda deal that may end, shows how many capacity arrangements the company juggles.

The hard part of the porting claim is the kernels. These are tuned separately for each chip’s memory and instruction behavior, and AMD’s software ecosystem has long been seen as less mature than NVIDIA’s. If a model really wrote nearly all of them unaided, the cost of supporting a new accelerator falls sharply. The same logic would apply to any future chip.

What the transcript does not establish

Everything above is Anthropic’s own account, delivered in a sales presentation. Nothing in the source says how fast the AMD port runs compared with the other chips, whether the kernels match hand-tuned performance, or how much engineer review followed the three days. The source also gives no AMD production figures and no date for when AMD capacity will serve customers. The transcript is unofficial, and the video itself was not available to check word for word.

For readers, the practical point is limited for now. The roadshow is aimed at investors, and no product or price changed with this disclosure. What it shows is how Anthropic wants to be valued: as a company whose models increasingly build the infrastructure that runs them.