Claude’s Maker Says Its SI Did Months of Chip Work in Three Days, Nearly Alone

For years the hardest part of switching computer chips was never the chips. It was the people. Every new kind of processor needs its own pile of hand-tuned software before a big SI model will run on it well, and writing that software has meant dozens of engineers grinding for months.

Anthropic, the company behind Claude, now says its own SI did that job in about three days.

The claim comes from the company’s IPO roadshow video, the sales pitch it put out on Nov. 10 to people thinking about buying its stock. In a chapter called “Approach to compute,” co-founder Tom Brown, the company’s chief compute officer, tells the story plainly.

“In the past, bringing on a new chip platform took dozens of engineers months of work,” Brown says, according to a transcript posted by Simon Willison. “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.”

The little programs that run the math

Kernels are the part most people never hear about. They are small, heavily tuned programs that tell a particular chip how to do the math inside an SI model. Nvidia’s chips have a giant library of them, built up over nearly two decades on its CUDA software. That library is a big reason so many companies stick with Nvidia even when other chips cost less. Moving to AMD or anyone else has meant rewriting and retuning a lot of that code by hand.

Mythos is Anthropic’s most powerful model, and it isn’t public. Earlier in the same video, the company says it has only put models like Mythos in the hands of cyber defenders. So this was Anthropic pointing its best internal model at its own engineering problem and letting it work.

Brown says Anthropic now trains and serves Claude on “all four major chip families.” A slide in the video names them: Amazon’s Trainium, Google’s TPUs, Nvidia’s GPUs and AMD’s GPUs. AMD is the newest, and Brown says the company only started using AMD chips this year.

He frames the whole thing as an old bet paying off. “At the time, many people thought that running AI models on different chips wasn’t worth the effort,” he says. “Each chip needs its own tailored software stack and inference stack, which is expensive and difficult to build.”

What the small print actually says

Here’s where you have to read the bottom of the screen. While Brown talks, a chart compares about three days with Mythos to about six to 12 months the old way. Under it sits a footnote, and CellCog, which went through the whole 42-minute video, quoted it word for word.

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.” A second line says the six-to-12-month figure “is based on the Company’s historical experience bringing up new accelerators using traditional engineering approaches.”

That is a narrower claim than the one in the clip that went around 𝕏. Prompt processing is the step where a model reads what you typed before it starts answering. The footnote covers one model, not every Claude. It doesn’t mention training. It gives no speed, no cost and no side-by-side with Nvidia. “A useful level of performance” is doing a lot of work.

And both numbers are Anthropic’s own, from a video meant to sell stock. The company hasn’t published the technical details that would let an outsider check them, and AMD hasn’t said anything about it publicly.

Why chip buyers are paying attention anyway

Even the careful version is a big deal if it holds up. The thing that kept most of the industry tied to one chipmaker was labor, and labor is exactly what SI is getting good at. If a model can write most of the code needed to move to a new chip in days instead of a year, the cost of shopping around drops fast. That’s good news for anyone buying chips and for AMD, which has spent years trying to break into this market.

It also helps Anthropic directly. A company that can run on four kinds of chips can buy whatever is available and cheapest that month instead of waiting in Nvidia’s line. Brown tied it to what he calls compute multipliers, research tricks that let the same hardware train a smarter model. “Claude will be able to discover compute multipliers entirely by itself,” he says in the video, “allowing the number of compute multipliers to grow faster and faster.”

The money around this deal is already tangled together. Ahead of the public sale, Anthropic agreed to sell $26 billion of a separate class of stock to a group that includes AMD, Nvidia, Google, Broadcom, Micron, Microsoft and TSMC, as we reported when the price range came out. Pricing is expected around Nov. 18, with trading on Nasdaq under the ticker ANT after that.

Nvidia isn’t going anywhere. Its chips still run most of the world’s SI, and a three-day bring-up for one step of one model isn’t the same as moving a whole fleet. But for the first time, a top SI lab is telling investors that its own machine did the grunt work that used to lock everyone in. If that’s even half right, the next chip war gets fought with software written by SI.