Anthropic has told investors in its initial public offering that its Claude Mythos Preview model did the low-level software work to get an Anthropic model running on AMD's MI355X accelerators in about three days, with “basically no human intervention,” a task the company says used to take six to 12 months and a team of dozens of engineers. Tom's Hardware read the claim out of Anthropic's investor materials this week.

The three days cover prompt processing, the prefill stage where a model reads and organises its input before generating anything, for one Anthropic model at what the company calls a useful level of performance. Mythos ported or wrote almost 100 percent of the kernels involved, the small compiled routines that map a model's operations onto a specific chip. Tom Brown, Anthropic's co-founder and chief compute officer, made the comparison, which rests on Anthropic's own history of bringing up new accelerators the traditional way.

Why kernel work was the moat

Nvidia's CUDA toolkit is nearly two decades old, and the reason it holds customers is not the API but the years of hand-tuned code that make each model run well on each generation of silicon. AMD's equivalent, ROCm, uses HIP as its programming layer, and code written for CUDA has to be converted, which is what AMD's HIPify tool does. Conversion is only half the work; the performance tuning afterwards is where engineering teams spend their quarters.

That is the part Anthropic says it automated. If a model can write its own kernels in days, the cost of adding a second supplier stops being a year of scarce engineers and starts looking like a purchasing decision. Anthropic has run models across Nvidia and AMD GPUs, Google TPUs and AWS Trainium since early on, manages the whole fleet as one pool of compute, and has an agreement with AMD for up to 2 gigawatts of capacity.

The claims that came from AMD first

Most of this argument was made by AMD before Anthropic put a number on it. In September 2025, chief executive Lisa Su said on a podcast that much of the AI software written for CUDA needs translating to AMD and that a large share of it can now be done in software. By July this year, senior vice president of AI Vamsi Boppana said his engineers were generating kernels with AI models, and Su recounted asking whether her team was helping the Anthropic engineer working on the MI355X and being told the help was not needed.

AMD's vice president of AI software, Anush Elangovan, told EE Times in April that Claude beats HIPify because it has web search built in. OpenAI's Philippe Tillet said at AMD's Advancing AI event in July that a single engineer confirmed its Helios racks worked within days once they arrived, and that agents were taking over kernel engineering work. Andrew Dieckman, who runs AMD's data center GPU business, told reporters he has almost no conversations with customers about CUDA any more.

The reason Anthropic's version carries more weight is that Anthropic buys the chips rather than selling them, and is about to be a public company. Its prospectus says it can onboard new chip platforms in a fraction of the time it once took, without the number. Anthropic has already been arguing in public that its models compete on capability rather than on which accelerator runs them, and it has been signing power deals to feed that fleet on the same logic.

What the claim leaves out

Anthropic has not said whether the three days included decode, the stage where the model produces its answer, which is the part that matters most for serving cost. It has not published throughput on MI355X against an Nvidia part, so nobody outside the company can say whether AI-written kernels are as fast as hand-tuned ones. The footnote on the slide limits the comparison to prefill on one model.

The counterargument is older than the claim. Nvidia's Jensen Huang said in an April interview that Nvidia uses AI on its own kernels and knows its architecture better than anyone, and dismissed labs that use other accelerators as a unique instance rather than a trend. SemiAnalysis made a related point in July: an agent writing kernels needs GPUs to test them against, and AMD has had a persistent shortage of spare hardware for exactly that work, which turns compute into the new bottleneck.

There is also the money. Nvidia and AMD are among seven companies buying $26.0 billion of Anthropic stock at the IPO price, which means Nvidia is investing in the customer that just published a case for not needing Nvidia. The next test is scheduled: Anthropic's first MI450 chips in Helios racks are due in the first half of 2027, a new generation with new kernels to write. If the three-day bring-up repeats on silicon nobody has tuned yet, the moat argument gets much harder to make.