Anthropic's Model Taught Itself to Run on AMD in "Three Days."

Anthropic told IPO investors that its Mythos model got Claude running on AMD chips in about three days, work the company says used to take six to 12 months, according to tikr.com's report. "It got Claude running on AMD in just three days with basically no human intervention. Claude ported or wrote almost 100% of the kernels," Anthropic co-founder Tom Brown said in the company's roadshow video.

This piece covers what Anthropic actually claimed, the narrow footnote behind the three-day figure, what NVIDIA's CUDA software lock-in is worth, and how the claim lines up with AMD's growth and valuation. It is written for engineers following GPU portability, and for readers tracking the AI chip market around Anthropic's IPO.

01Three days versus six to 12 months

TIKR's analysis frames the gap as at least 60x faster than Anthropic's old approach. Brown said Anthropic only started working with AMD GPUs "this year," and that Mythos did most of the porting work.

The claim quickly spread beyond the roadshow. After an investor on X argued that every AI IPO gets priced off NVIDIA's multiple, Replit CEO Amjad Masad pushed back, writing that Anthropic "made it clear that they're not dependent on Nvidia and that they ported their software to run on amd in 3 days."

The timing added weight. The claim surfaced during a brutal week for NVIDIA, whose stock sank as CEO Jensen Huang resigned after The Wall Street Journal reported the board was preparing to oust him, while AMD and Broadcom rose on bets they would pick up business.

02CUDA porting and AMD's software gap

AMD's hard problem has been software rather than silicon. Code written for NVIDIA's CUDA has to be reworked before it runs well on anyone else's chips, which is the step Mythos is said to have largely automated for Claude.

AMD has been pointing at faster bring-up for months. TBPN co-host John Coogan said that when the show spoke with AMD CEO Lisa Su over the summer, "the big story of the day was that the bring up time had been significantly reduced from these AI speed ups."

03AMD revenue growth and analyst forecasts

AMD's business is already growing fast. Third-quarter revenue rose 45% to $13.4 billion, data center revenue jumped 97% to $8.5 billion, and management guided to about $17 billion for the fourth quarter.

Analysts expect AMD revenue to climb from $34.6 billion in 2025 to $95.9 billion in 2027, then to $135 billion in 2028, nearly tripling over the first two years. TIKR, which tracks 53 analysts on AMD, argues that a faster path off CUDA makes that share-gain forecast easier to hit.

04NVIDIA's 70% gross margin and the CUDA moat

NVIDIA can charge what it does partly because its customers' code is hard to move. Its gross margin has stayed above 70% for three straight fiscal years, through the year ended January 2026.

That pricing power is already under some strain. On its Aug. 26 call, NVIDIA management said it expects gross margin to bottom at 71% to 72% in the January 2027 quarter as memory prices climb, and TIKR's author expects buyers to use cheaper ports as leverage while the company's leadership is in flux.

05The MI355X footnote and 7.5 million CUDA developers

Anthropic's own slide is narrower than "running on AMD" sounds. Its footnote says the three days covered prompt processing for one Anthropic model on AMD's MI355X GPUs, at a useful level of performance, so the claim covers one task for one model.

NVIDIA's annual report counts more than 7.5 million CUDA developers. Cerity Partners' Jim Lebenthal, who owns NVIDIA, argued on the Halftime Report that a new CEO would inherit "one of the best jobs in America," because "while you can replace the CEO, you can't replace the CUDA architecture. Every developer builds on it."

TIKR's author counters that the tool doing the porting is the product Anthropic sells, and expects other big labs and cloud customers to point coding models at the same job. The piece also notes that Anthropic has every reason to talk up its independence from NVIDIA on a roadshow.

06AMD's 37x multiple and NVIDIA's Nov. 17 report

AMD trades at 37x forward earnings, below its three-year average of 39x and well under the peak of 67x it reached at the end of June. TIKR's author calls AMD "the better side of this trade" at that below-average multiple, while noting the CUDA developer base won't leave overnight.

The next checkpoint is NVIDIA's earnings report on Nov. 17. Its gross margin guidance for the January quarter will show whether customers are already pushing back on price. The TIKR article is analysis, not investment advice, and TIKR says it holds no position in the stocks mentioned.