What was actually claimed
The sentence that traveled is real. Brown is describing a deliberate bet: instead of standardizing on one chip supplier, Anthropic built the software work to run Claude on several. Bringing up a new platform used to be the expensive part of that bet. This time, he says, the company let Mythos try the job, and Mythos had Claude running on AMD in about three days, porting or writing almost all of the kernels, the small tuned programs that carry out a model's math on a particular chip.
Posts sharing the clip on X went further and declared Nvidia's software advantage finished. The video does not say that. Its narration never mentions Nvidia or CUDA at all, and the slide behind Brown lists Nvidia GPUs as one of the four families Anthropic still uses.
Where the footnote cuts it down
The chart on screen compares about 3 days with Mythos against about 6 to 12 months the traditional way. Underneath, in small type, sits the definition: the three days is the approximate time to bring up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance.
Each qualifier removes something. One model, not the Claude family. Prompt processing — the input-reading stage of inference — with the footnote establishing nothing about output generation or training. Useful performance, a phrase the video never turns into a number: no tokens per second, no comparison against the same model on Nvidia hardware, no cost. And the 6-to-12-month side of the comparison is Anthropic's estimate of its own past projects, not a measured baseline anyone can rerun.
The prospectus is quieter still, on CellCog's reading of the filing: Amendment No. 2 says new chip platforms historically required intense engineering effort over many months, and that Claude now helps Anthropic's teams bring them into production more quickly, but it does not contain the words "three days," "MI355X" or "kernels." ALFW's desk has not independently checked that absence. It does mean the specific, quotable version of this claim exists, so far, only in marketing material published while the company is selling shares.
Why the narrow version still matters
Strip the claim to what the footnote supports and it remains a substantial result, if it holds. The practical barrier around Nvidia is not only the chips rivals can build. It is also the months of kernel work needed before a model runs well on different silicon, work that has to be repeated for each new platform. A tool that compresses the first stage of that work from months to days, even for one model and one stage, changes the arithmetic of trying a second supplier. Anthropic has a commercial reason to prove that point now: under its July partnership with AMD, reported by MadRobot from AMD's announcement, it plans to deploy up to 2 gigawatts of AMD's MI450-series chips from the first half of 2027.
What would let a reader evaluate the reported result is specific: the tested model and workload, the performance criteria behind "useful," where the three-day clock started and stopped, how much human involvement "basically none" contained, and a comparable historical baseline. Demonstrating the broader claim — a complete port — would additionally require evidence about output generation. None of that has been published. Until some of it is, the honest summary is the footnote's, not the clip's: one stage of one model crossed to a new chip in about three days, by its maker's account, in a sales video. That is a lead worth following. It is not a moat draining.
Sources
- [1] Anthropic, IPO investor site and roadshow video, "Approach to compute" chapter (claim at 20:42; site confirms roadshow from November 10, 2026, pricing expected November 18, trading expected November 19 under ANT, subject to regulatory clearance)Read source
- [2] CellCog, "Anthropic Roadshow Video: The AMD and Disease Claims," November 11, 2026 (full read of the video's 6,468-word captions file and slides; verbatim footnote; prospectus comparison)Read source
- [3] MadRobot, "Anthropic's Mythos got Claude running on AMD chips in three days," November 11, 2026 (read first-hand in full by the desk; independent transcription of the same passage and footnote; AMD partnership context from AMD's July announcement; AMD has not commented)Read source