Anthropic's IPO Roadshow Video Says Mythos Got Claude Running on AMD MI355X in About 3 Days, With Claude Porting or Writing Almost 100% of the Kernels, vs 6-12 Months Traditionally

Background and industry pain point

For AI labs, the hard part of switching chips is software, not hardware. Every accelerator needs its own kernels, inference stack and tuning, which has traditionally taken dozens of engineers months of work, and that is why NVIDIA's CUDA ecosystem is called a moat. Anthropic launched its IPO roadshow on November 10 (825 million Class A shares at an expected $70-$78, applying to list on Nasdaq as "ANT") and published a 42-minute roadshow video on its official IPO site. The video's compute chapter uses AMD as an example of AI helping the company bring up new silicon.

Architecture and mechanism

In Anthropic's words: it started using AMD GPUs for the first time this year, and "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." An on-screen footnote is more precise: Claude Mythos Preview brought up prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance. So this is kernel porting and tuning for the prefill stage, not full training and not every model in full production. Anthropic says it trains and serves Claude on Trainium, TPUs, NVIDIA and AMD, and treats them as one fungible pool, matching each chip to the job it does best.

Benchmarks and measured results

Anthropic did not publish throughput, latency or any AMD-versus-NVIDIA comparison, only "a useful level of performance". The 3 days versus 6-12 months comparison is a vendor figure based on the company's own history, with no third-party check. Financial figures from the same video (vendor numbers): revenue grew more than 45x over six quarters while compute costs grew 10x, and year-to-date gross margin is 85%, up from 29% a year earlier. Posts on X declaring the CUDA moat dead overstate it, and infrastructure analysts have pushed back: getting a model to run is different from running it reliably at scale with competitive cost.

Developer guide

For teams building GPU kernels and inference engines, the signal is that frontier coding agents can now take on most of a cross-hardware port. Mythos is still limited to a small set of cyber defenders, though, so you cannot reproduce this today. You can start with Claude Opus or Fable plus ROCm and Triton, porting one operator at a time with unit tests, and make performance benchmarks and numerical checks the agent's acceptance gate. The full video and captions are in the Roadshow video section of ipo.anthropic.com.