GPU accelerator boards and cooling pipes inside a data center server rack.

Claude Mythos Ran on AMD GPUs in Three Days: Why That Matters

Key Takeaways

  • Anthropic told prospective investors that its next-generation model, Claude Mythos, completed in about three days the software work needed to run its AI on AMD graphics processing units, a job the company says previously took dozens of engineers six to twelve months.
  • The claim comes from Anthropic's IPO roadshow video of 10 November 2026, delivered by co-founder and chief compute officer Tom Brown, who said Mythos got Claude running on AMD hardware "with basically no human intervention" and ported or wrote almost 100% of the GPU kernels itself.
  • A footnote on Anthropic's own slide narrows the scope considerably: the three days covered bringing up prompt processing, known as prefill, for a single Anthropic model on AMD's MI355X GPU at a "useful level of performance".
  • Anthropic has not said whether the response-generation stage, known as decode, was included, and it has published no performance comparison against comparable Nvidia hardware, so the claim has no independent verification yet.
  • The story matters because Nvidia's CUDA software platform, built over nearly two decades, has been the main reason AI developers stay locked to Nvidia chips; if an AI model can do the porting work in days, that lock-in weakens.
  • Anthropic's first AMD MI450 chips, deployed in Helios racks, are due in the first half of 2027, which will be the first real test of whether the three-day result can be repeated at production scale.

What Anthropic Actually Said

The claim surfaced in materials Anthropic prepared for investors ahead of its initial public offering. In a roadshow video released on 10 November 2026, Tom Brown, Anthropic's co-founder and chief compute officer, said that Claude Mythos "got Claude running on AMD in just three days with basically no human intervention". According to reporting of the materials, the model ported or wrote nearly 100% of the kernels, the low-level pieces of code that tell a GPU how to carry out each mathematical operation in a model.

The comparison Anthropic offered was stark. The same work, bringing a new accelerator into use, previously consumed dozens of skilled engineers for six months to a year, according to the company's own account of its engineering history. A slide in the roadshow deck summarised the change in a small table: time to bring up AMD's MI355X, six to twelve months under the old approach, about three days with Mythos.

It is worth pausing on the setting. This was not a peer-reviewed paper or a technical blog post. It was a pitch to prospective investors, made days before a planned public listing. That does not make the claim false, but it does mean it was presented by a company with every incentive to show its technology in the strongest possible light, and it has not been subjected to the kind of independent testing that would normally settle a claim of this size.

The Footnote That Changes the Story

The most important detail in the whole episode is a footnote. On the slide Anthropic released, a footnote limits the three-day figure to bringing up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance. Prompt processing, usually called prefill, is the stage in which a model reads and processes the input it has been given before it starts replying. It is a real and substantial part of running a model, but it is not the whole job.

Anthropic has not said whether the second stage, decode, the step in which the model generates its reply token by token, was included in the three days. Nor has the company said what "useful level of performance" means in measurable terms. There is no published figure for speed, no throughput number, and no comparison with the same model running on Nvidia hardware. The company's securities filing, the prospectus filed as S-1/A No. 2 on 10 November, states in general terms that new chip platforms can now reach production with far less engineering effort, but it does not contain the three-day figure itself.

None of this is a reason to dismiss the claim. Getting prefill working on a new accelerator in three days, if that is what happened, would still be a remarkable engineering result. But the gap between "prefill for one model at a useful level of performance" and "Claude running on AMD" is exactly the gap that matters for anyone deciding whether to buy AMD hardware, and readers should keep that gap in mind.

What Is Claude Mythos?

Claude Mythos is Anthropic's next-generation model, presented to investors as the system that did the porting work. The version named in the roadshow materials is Claude Mythos Preview. Beyond the AMD claim, little has been disclosed publicly about the model's size, training, or capabilities, and Anthropic has not published a model card or technical report for it in the materials reported so far.

What can be said with confidence is what the model reportedly did: it wrote or ported the GPU kernels needed to run an Anthropic model on AMD hardware. A kernel is a small, highly optimised programme that runs on the GPU and performs one specific operation, such as a matrix multiplication or an attention calculation. A modern AI model needs hundreds or thousands of them, each tuned to the exact hardware it runs on. Writing them by hand is slow, fiddly work that requires deep knowledge of the chip's architecture. That is why porting a model to a new chip has historically been a project measured in months.

The idea that a model can write its own kernels is not new in principle. AI-assisted code generation has been improving for years, and kernel generation is a natural target because kernels are small, well-specified, and testable: either the code produces the right numbers at the right speed, or it does not. What would be new is the scale and speed Anthropic claims: nearly all of the kernels, for a frontier model, on unfamiliar hardware, in three days, with almost no human help.

What Porting an AI Model to a New GPU Actually Involves

To understand why three days would be striking, it helps to understand what "bringing up" a model on a new chip actually means. An AI model is, at its core, a very large set of numbers and a long sequence of mathematical operations. Running it requires far more than copying files onto a new machine.

First, every operation in the model needs a kernel written for the new chip's instruction set and memory layout. Nvidia's GPUs and AMD's GPUs organise their compute units and memory differently, so code written for one does not simply run on the other. Second, those kernels need to be tuned: the same operation can run ten times faster or slower depending on how memory is moved, how work is split across compute units, and how operations are fused together. Third, the surrounding software stack, the libraries that handle communication between chips, the scheduling of work, and the serving infrastructure, must all be adapted. Finally, the result has to be validated: the model must produce the same answers, within tolerance, as it does on the original hardware, and it must do so reliably under load.

Historically, this is why a new accelerator takes six to twelve months to adopt, in Anthropic's telling. It is not a matter of one engineer spending a week. It is a team of specialists working through hundreds of kernels, chasing performance regressions, and fixing the long tail of edge cases that only appear at scale. The claim that Mythos compressed that into three days is, in effect, a claim that the most labour-intensive part of AI infrastructure work can now be automated.

Why CUDA Is the Moat

Nvidia's dominance in AI chips is usually described in terms of hardware, but the hardware is only half the story. The other half is CUDA, Nvidia's programming platform, which has been built up over nearly two decades into a mature ecosystem of compilers, libraries, debugging tools, and a large community of developers who know how to use them. When an AI lab builds its models on CUDA, every tool, every optimisation, and every engineer's hard-won experience is tied to Nvidia's chips.

AMD's answer is ROCm, its own software stack, with HIP as the programming layer. Code written for CUDA has to be transformed to work on ROCm, and AMD even built a tool called HIPify to automate part of that conversion. But HIPify and similar tools have never fully closed the gap, because the gap is not just about translating code. It is about the thousands of small optimisations and workarounds that accumulate around a platform over twenty years, and about the fact that most AI software is written and tested on Nvidia hardware first.

That is why Anthropic's claim, if it holds up, is read as a challenge to Nvidia rather than simply as good news for AMD. If an AI model can do the porting work itself, in days rather than months, then the cost of switching away from CUDA falls sharply. The moat does not disappear overnight, because hardware performance, supply, and the existing ecosystem still matter, but the moat gets narrower. Anthropic's own stated goal, according to reporting of the roadshow, is to create a "fungible compute pool", in which its models can run on whichever hardware is available and cheapest, rather than being tied to a single supplier.

Anthropic, AMD, and the Bigger Picture

Anthropic's relationship with AMD is not new, but it has deepened quickly. The company began working with AMD GPUs this year and has signed a supply agreement for up to 2 gigawatts of AMD capacity, according to reporting of its investor materials. That is a very large commitment, and it follows a pattern of huge chip deals across the industry, including Anthropic's own $63 billion chip financing arrangement with Broadcom, which shows how seriously the company is taking its supply of compute. Anthropic's interest in AMD is strategic rather than experimental: the company wants a second source of compute at a scale that matters.

The timing also matters. Anthropic is heading into its public listing, and the Mythos claim arrived as part of that story. Anthropic's IPO filing, with its sales figures and losses laid out for investors, shows a company whose biggest cost is compute, and investors in an AI lab are, in effect, buying exposure to that cost. A credible claim that Anthropic can cut its dependence on one supplier, and cut the engineering cost of adopting new hardware by an order of magnitude, is a powerful part of that pitch. It is also a claim that will be tested in public, because the company's first AMD MI450 chips, deployed in Helios racks, are due in the first half of 2027. If the three-day porting result repeats there, at production scale, the scepticism will fade quickly. If it does not, the footnote will look more important than the headline.

The chip market context is worth noting too. Nvidia's position has been under pressure from several directions at once: customers want second sources, regulators are watching, and AMD has been investing heavily in both hardware and software. The scale of the money now moving around AI chips was underlined recently when Nvidia and AMD were tied to a $4 billion Meta investment and a $95 billion raise, a reminder that the contest for compute supply is being fought in tens of billions of dollars. Anthropic's claim landed in that context, which is why it was read by market commentators as good news for AMD investors and bad news for Nvidia, as the Motley Fool put it on 15 November 2026. But share-price commentary is not evidence about engineering, and readers should separate the market reaction from the technical claim.

What This Means for AI Developers

For most developers, the immediate practical effect of the Mythos claim is small. Nobody outside Anthropic has access to Mythos, and nobody has reproduced the three-day port. But the direction of travel matters, because it points to a world in which the choice of AI chip is less sticky than it is today.

If model porting becomes cheap and fast, several things follow. First, AI labs and large users gain bargaining power with chip suppliers, because they can credibly threaten to move workloads. Second, smaller chipmakers, including AMD and newer entrants, get a fairer hearing, because the software barrier to trying their hardware falls. Third, the pace of hardware competition could quicken, since a new chip would no longer need a year of software work before anyone could use it seriously. And fourth, the engineers who currently spend their careers writing kernels may find their work changing, with more time spent on verification, architecture, and the hardest performance problems, and less on routine porting.

There is also a caution. Faster porting does not prove that AMD hardware has matched Nvidia's performance. As far as public reporting shows, Anthropic has not released performance comparisons with similar Nvidia hardware, and independent benchmarks of the Mythos work are not available. A model can run on a chip without running well, and "useful level of performance" is a phrase that covers a wide range of outcomes. Until Anthropic or an independent party publishes numbers, the performance question is open.

The Questions That Still Need Answers

Several questions remain, and they are the ones to watch over the coming months.

First, does the three-day figure include decode, the response-generation stage, or only prefill? Anthropic has not said, and the answer determines whether the claim covers a full working model or one stage of it.

Second, what does "useful level of performance" mean in numbers? Without a throughput or latency figure, and without a comparison to the same model on Nvidia hardware, it is impossible to judge whether the ported model is production-ready or merely a proof of concept.

Third, can the result be repeated? A single port, for a single model, on a single chip, under the eye of the team that built the model, is a demonstration. Repeating it across models, chips, and teams is what would make it a capability. The MI450 deployment in the first half of 2027 is the natural test.

Fourth, how much human intervention was there really? "Basically no human intervention" and "almost 100% of the kernels" are the company's words, and they leave room for meaningful human involvement in setting up the task, choosing the approach, or fixing the failures. The details matter, and they have not been published.

Fifth, what does this mean for the CUDA ecosystem itself? Even if porting becomes cheap, CUDA's libraries, tools, and developer community do not vanish. Nvidia's moat may narrow without disappearing, and the company has its own AI-assisted tooling. The competitive question is not whether CUDA dies, but whether it stops being a reason to pay Nvidia's prices.

What We Could Not Verify

  • Whether the decode stage was included in the three-day figure. Anthropic has not disclosed this, and reporting of the roadshow materials notes the omission.
  • The performance of the ported model on AMD MI355X hardware. No throughput, latency, or benchmark figures have been published, and no comparison with Nvidia hardware exists in the public record.
  • The exact extent of human involvement in the port. The phrases "basically no human intervention" and "almost 100% of the kernels" are Anthropic's own characterisations, and no detailed account has been published.
  • Whether the result has been independently reproduced. As of 18 November 2026, no independent benchmark or third-party verification of the Mythos porting work has been reported.
  • The financial terms of Anthropic's AMD supply agreement. The reported figure is up to 2 gigawatts of capacity; pricing and timing beyond the MI450 deployment window have not been disclosed.

Sources

Similar Posts