Anthropic IPO roadshow video details four-chip compute strategy
Simon Willison wrote on 11 November 2026 that Anthropic kicked off its IPO roadshow the day before, and that the roadshow video contains previously unreleased information about how the business works. He singles out the "Approach to compute" section, which features co-founder Tom Brown. It describes how Anthropic trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.
There is no official transcript, and Willison says he does not know why. He made his own with MacWhisper and posts highlights from it.
The first excerpt says Claude is now trained and served across all four major chip families. According to the transcript, that is possible only because Anthropic made an unconventional early bet on a multi-chip strategy, at a time when many people thought running AI models on different chips was not worth the effort. Each chip needs its own tailored software stack and inference stack, which is expensive and difficult to build. The speaker says Anthropic has now been doing it for years and has become very good at it.
The second excerpt gives an example: this year Anthropic started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work. This time, the speaker says, "we had Mythos try it out". It got Claude running on AMD in three days, with basically no human intervention, and Claude ported or wrote almost 100% of the kernels. The excerpt does not explain what Mythos is.
The third excerpt covers what the speaker calls the second driver of Anthropic's compute advantage: its rate of discovery of compute multipliers. Anthropic focuses on total effective compute, the product of physical capacity and the multipliers it discovers. The speaker's illustration is hypothetical: if a model of a certain intelligence needs 100 units of compute, a 4x multiplier lets you train the same model with 25 units. Or you can spend the same hundred physical units on a more intelligent model that would previously have needed 400 units.
The speaker adds that multipliers are a competitive advantage that is hard to replicate, because they build on each other: each new one has a positive effect on the ones before it, so they stack. Anthropic believes this can accelerate further as it approaches recursive self-improvement, and the speaker says Claude will be able to discover compute multipliers entirely by itself, so their number can grow faster and faster.
Willison ends with a pointer to earlier evidence. Last September, Anthropic's postmortem of three recent issues described bugs that affected just its TPU deployments, which he says gave a hint of the multi-chip strategy.
Key facts
- Anthropic says it now trains and serves Claude across four major chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.
- AMD GPUs are new this year. The transcript says Mythos got Claude running on AMD in three days with basically no human intervention, and Claude ported or wrote almost 100% of the kernels.
- Past chip ports took "dozens of engineers months of work", according to the transcript.
- Anthropic says it focuses on total effective compute: physical capacity multiplied by the compute multipliers it discovers. A hypothetical 4x multiplier turns 100 units of compute into 25 for the same model.
- The speaker says Claude will be able to discover compute multipliers entirely by itself. The text comes from Willison's own MacWhisper transcript, since no official one exists.
Why it matters
The excerpts put Anthropic's compute approach in its own words, in a video made for the IPO roadshow. The claim is that supporting several chip families, long seen by some as not worth the effort, is a deliberate advantage that Anthropic has built over years. The AMD example is offered as evidence: a port that once took dozens of engineers months of work was done in three days, with Claude doing nearly all of the kernel work. Willison calls the video a source of previously unreleased information about how the business works.
Who it affects
The four chip families named are AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs, so the vendors behind them are part of the story. Anyone following Anthropic's business or its IPO has a first-hand description of how the company says it handles compute. The excerpts say nothing about customers or pricing.
How to use it
This is a reading item, not a product. The most reusable idea is the arithmetic of compute multipliers: with a 4x multiplier, the same model needs a quarter of the compute, or the same compute buys a model that would once have needed four times as much. Willison's post is a convenient place to read the excerpts, since there is no official transcript.
How solid is it
The material is first-hand from Anthropic's own roadshow video, but the text here is Willison's MacWhisper extraction rather than an official transcript, so wording may differ from the video. The excerpts are not individually attributed; Willison says only that the "Approach to compute" section features co-founder Tom Brown. The 100, 25 and 400 unit example is explicitly hypothetical, not a measurement. The three-day AMD result and the 'almost 100%' kernel figure are statements in the video, with no supporting data in the post.
Risks and caveats
This is promotional material for an IPO roadshow, and the claims about a hard-to-replicate advantage and about Claude discovering multipliers by itself are forward-looking beliefs and predictions. The source gives no detail on the performance or production use of the AMD deployment, does not explain what Mythos is, and gives no financial figures, valuation or listing details from the video.
“Compute multipliers create a competitive advantage for us that's hard to replicate. They build on each other.”
— Anthropic IPO roadshow video, per Simon Willison's MacWhisper transcript