Anthropic's IPO roadshow video

Anthropic, the company behind Claude, began its IPO roadshow on November 10. A roadshow is the round of presentations a company gives prospective investors before its shares go on sale. Simon Willison pointed to the section of Anthropic's roadshow video on compute, narrated by co-founder Tom Brown, as containing information about the business that had not been released before. Because Anthropic posted no official transcript, he made his own with the transcription app MacWhisper (simonwillison.net, transcript). Brown's central claim is that Anthropic now trains and serves Claude on all four major families of AI chips: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.

Running one model on several kinds of chips is costly because each chip needs its own tailored software, including the low-level routines, called kernels, that carry out the model's arithmetic on that hardware. Brown says Anthropic accepted that cost early, and that it now lets the company buy whichever chips offer the best price for the performance in a market short of supply, assign each chip to the training or serving work it does best, and treat everything it runs as one interchangeable pool. Anthropic's own postmortem of three infrastructure bugs had already described serving Claude on Trainium, NVIDIA GPUs and TPUs. AMD is the new addition. Brown says adding a chip platform used to take dozens of engineers months, but this year a Claude model called Mythos got Claude running on AMD GPUs in three days with basically no human intervention, porting or writing almost all of the kernels itself.

The second half of the section is about efficiency. Brown describes "total effective compute" as physical computing capacity multiplied by compute multipliers, meaning discoveries that get the same result from less hardware. In his example, a 4x multiplier lets a model that needed 100 units of compute be trained with 25, or lets those 100 units train a model that would previously have needed 400. He says the multipliers stack, and that Anthropic expects Claude eventually to discover them by itself as it approaches recursive self-improvement, an AI improving the process that produces AI.

All of this is Anthropic's account of its own strategy, given in a presentation meant to sell shares. The transcript names no measured multiplier and offers no way to check the three-day AMD port, so read these as company claims rather than demonstrated results. What the video does establish is the company's stated position: it presents running on all four major chip families, and letting its own models do the porting, as a cost advantage it expects investors to value.