Anthropic claims its Mythos model ported Claude to AMD chips in three days — but the fine print narrows the claim
In Anthropic's IPO roadshow video, co-founder Tom Brown said its frontier model Claude Mythos got Claude running on AMD Instinct MI355X GPUs 'in just three days, with basically no human intervention,' versus roughly 6–12 months under traditional engineering. But a footnote in Anthropic's own roadshow material limits the claim to bringing up prompt processing for one model 'at a useful level of performance' — not training or full inference — and AMD has not commented.
The News
In Anthropic's IPO roadshow video, co-founder and Chief Compute Officer Tom Brown made a claim that, if true, would rattle the chip industry. He said Anthropic's frontier model Claude Mythos got Claude running on AMD Instinct MI355X GPUs 'in just three days, with basically no human intervention,' porting or writing 'almost 100% of the kernels' — the low-level software that tells a chip how to run a model's maths. Traditional engineering, Brown said, would take roughly 6 to 12 months. Anthropic is a private company preparing to go public, and the roadshow is its sales pitch to investors — which is the first reason to read the claim carefully.
The second reason is the footnote. In Anthropic's own roadshow material, a footnote limits the achievement to bringing up prompt processing — the stage where a model reads its input — for one Anthropic model on MI355X GPUs 'at a useful level of performance.' That is not training a model, and it is not running full inference, the stage where a model generates its answers. Getting one stage of one model working usefully is a real engineering feat, but it is a long way from 'Claude runs on AMD now.' The figures come from Anthropic's own IPO marketing, have not been independently verified, and AMD has not commented.
The backdrop is real enough. Anthropic says it now trains and serves Claude across four chip families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. And in July, Anthropic and AMD announced a partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs, with AMD agreeing to invest up to $5 billion in Anthropic as milestones are hit. Anthropic has a genuine commercial interest in making AMD hardware work — it is not a neutral observer of its own claim.
A three-day port of one stage of one model is a demonstration, not a dead moat.
Why It Matters
The claim matters because Nvidia's grip on AI rests on more than its chips. It rests on CUDA, the software platform that developers use to run AI on Nvidia hardware, and on the thousands of kernels — small, hand-tuned programs — that make models run fast on it. Porting a model to a rival chip has historically meant months of rewriting those kernels by hand. If an AI model can do that work itself in three days, the switching cost that protects Nvidia's position gets much smaller, and AMD's hardware becomes a credible alternative far faster.
But the footnote is the story. Prompt processing is one stage of running a model, and 'a useful level of performance' is a deliberately soft phrase — it does not say the model runs as fast on AMD as on Nvidia, or that training works at all. Until Anthropic or AMD shows Claude training on MI355X chips, or running full inference at competitive speed, the claim is a demonstration of a capability, not a change in the market. And because it comes from Anthropic's IPO roadshow — a document designed to sell shares — it deserves the scepticism any sales pitch deserves.
For investors, the right frame is optionality, not revolution. Anthropic now has a commercial reason to make AMD work, a $5 billion AMD investment tied to milestones, and a 2-gigawatt deployment plan. That alone makes AMD a more credible second source for AI compute, which is good for AMD and mildly uncomfortable for Nvidia. But the 'CUDA moat is dead' framing is not supported by the evidence: a three-day prompt-processing port is not a dead moat, and nobody has shown otherwise.
Analyst's Take
Our call is Mixed. If Anthropic's claim holds up — an AI model porting another model to new hardware in days rather than months — it is a genuine crack in the switching costs that protect Nvidia's position, and a real boost for AMD. But the claim is Anthropic's own, unverified, narrowed by its own footnote to one stage of one model, and AMD has not commented. A roadshow claim is not a benchmark.
What to watch: whether AMD confirms the port, and whether Anthropic shows Claude training — not just prompt processing — on MI355X or MI450 hardware. Training is where the money is, and it is the stage the footnote does not cover. Until then, this is a fascinating claim with a long way to go before it changes anyone's chip orders.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is a kernel?
A kernel is a small, specialised program that tells a chip exactly how to carry out one piece of a model's maths. Models need thousands of them, tuned to each chip — and writing them by hand for a new chip is slow, skilled work.
What is prompt processing?
It is the stage where a model reads and digests its input — the prompt — before it starts generating an answer. Anthropic's footnote limits its claim to this stage, not to training the model or to generating full answers.
What is CUDA, and why does it matter?
CUDA is Nvidia's software platform for running AI on its chips. Developers build on it, and the kernels written for it do not run on rival chips — so switching to AMD has historically meant months of rewriting. That switching cost is often called Nvidia's 'moat.'
Why should investors be sceptical?
The claim comes from Anthropic's IPO roadshow — a sales pitch to investors — and has not been independently verified. Anthropic's own footnote narrows it to one stage of one model, and AMD has not commented.
What did Anthropic and AMD actually agree in July?
A partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs for Anthropic, with AMD agreeing to invest up to $5 billion in Anthropic as milestones are hit. That deal is confirmed; the three-day porting claim is not.
👁
Source: Wccftech; Anthropic IPO roadshow video/transcript (Simon Willison)