Anthropic Says Its SI Did Months of Chip Work in Three Days. The Footnote Says Less.

Anthropic is selling stock next week, and one line from its sales pitch has the whole chip world talking. The company says its most powerful model did a job in about three days that used to take a team of engineers the better part of a year.

The job was getting Claude to run on a new kind of chip. The chips were AMD’s. And the target everyone heard was Nvidia, whose grip on SI computing rests partly on how hard it is to leave.

The claim comes from Anthropic’s IPO roadshow video, which the company posted on its investor site after it launched the roadshow on Nov. 10. The final share price is expected Nov. 18, with trading on Nasdaq under the ticker ANT the next day, according to Mad Robot’s write-up of the roadshow.

Tom Brown’s three days

The line belongs to Tom Brown, an Anthropic co-founder who now runs its compute as Chief Compute Officer. In the video’s “Approach to compute” chapter, he explains that Anthropic trains and serves Claude across four chip families: Amazon’s Trainium, Google’s TPUs, Nvidia’s GPUs and AMD’s GPUs.

“We train and serve Claude now across all four major chip families,” Brown says, according to a transcript Simon Willison published, since Anthropic didn’t post one. “That’s only possible because early on we made an unconventional bet on a multi-chip strategy.”

Then comes the part that went everywhere. “This year, we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work. This time, we had Mythos try it out. It got Claude running on AMD in just three days, with basically no human intervention. Claude ported or wrote almost 100% of the kernels.”

Kernels are the small, heavily tuned programs that actually do a model’s math on a particular chip. Every chip family needs its own set, and writing them well is slow, specialized work. That’s a big reason companies stick with Nvidia. Its CUDA software has had nearly two decades to pile up kernels, tools and engineers who know them by heart.

A chart on screen puts it plainly: about three days with Mythos, against “~6-12 months with traditional approach.”

What the small print actually covers

Look under that chart and there’s a footnote, and it’s a lot narrower than the clip that spread online. It says the three days reflects “the approximate time for Claude Mythos Preview to bring up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance.”

That’s one model, not all of Claude. It’s one chip, the MI355X. And it’s prompt processing, the stage where a model reads what you typed before it starts writing back. The footnote says nothing about training on AMD, and nothing about the half of the work where the model generates its answer.

It also doesn’t give a speed. “A useful level of performance” could mean nearly as fast as Nvidia or a long way behind. The 6-to-12-month comparison, the footnote adds, “is based on the Company’s historical experience bringing up new accelerators using traditional engineering approaches.” In other words, both numbers are Anthropic’s own estimates.

None of this has been independently checked. Anthropic hasn’t released technical details of the port, and the figures come from marketing for a share sale, where companies show their best side. AMD hadn’t commented, according to Mad Robot.

Still a big deal, even shrunk to size

Here’s the thing. Even the narrow version is impressive if it holds up. Bringing up the first working piece of a model on brand-new hardware is the hardest, most tedious part of a port. If an SI model can write nearly all of those kernels on its own in a long weekend, the next chip, and the one after that, gets cheaper to adopt.

That matters to anyone buying computing power, which today mostly means buying from Nvidia. The less it costs to switch, the more room there is for AMD, Google, Amazon and whoever comes next to compete on price. A post from Marcelo P. Lima on 𝕏 claiming “the Nvidia CUDA moat is gone” drew close to a million views. That’s the hype version. The footnote version is that a crack showed up, on one stage of one model.

It also fits a deal the two companies already have. In July, Anthropic and AMD announced a partnership under which Anthropic would deploy up to two gigawatts of AMD’s next-generation MI450 chips starting in the first half of 2027, with AMD investing up to $5 billion, as Mad Robot recounts. That deal already included using Claude to speed up ROCm, AMD’s answer to CUDA. AMD is also buying $2.5 billion of shares alongside the IPO.

Mythos isn’t something you can use

One more thing to keep in mind. Mythos is not a public product the way regular Claude is. Earlier in the same video, Anthropic says it has deployed models like Mythos only to cyber defenders, according to CellCog’s breakdown of the roadshow. So nobody outside the company can rerun the test.

Brown went further in the same chapter. He talked about “compute multipliers,” the tricks that let a lab train the same model on less hardware, and said Anthropic believes this “can accelerate even further as we approach recursive self-improvement,” with Claude eventually finding those tricks “entirely by itself.”

That’s a big promise from a company asking investors for money. The three-day port is the first concrete example it has put on the table, and the footnote tells you exactly how far that example goes. When the shares start trading next week, buyers will be paying for the promise. The fine print is worth reading first.