Anthropic's IPO roadshow video
Anthropic began its IPO roadshow yesterday, and the accompanying roadshow video contains material that had not previously been public about how the company runs its business. The most detailed section, presented by co-founder Tom Brown, covers Anthropic's approach to compute and explains how Claude is trained and served across four different chip families.
Brown states that Anthropic now trains and serves Claude across all four major chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. He frames this as the product of an unconventional early bet on a multi-chip strategy, noting that many people at the time doubted running AI models on different chips was worth the effort. The stated reason for that skepticism is the software burden: each chip family needs its own tailored software and inference stack, which is expensive and difficult to build. Anthropic has been doing this for years and says it has become very good at it.
The concrete example Brown gives is AMD. Anthropic started working with AMD GPUs for the first time this year. Historically, bringing up a new chip platform took dozens of engineers several months. This time, a system referred to as Mythos did the work: Claude was running on AMD in three days with basically no human intervention, and Claude ported or wrote almost 100% of the kernels. For developers, kernels are the low-level compute routines that determine how efficiently a model runs on a given piece of hardware, so having a model write them for a new accelerator is a meaningful sign of maturity in automated hardware bring-up.
The second driver Brown describes is the rate at which Anthropic discovers compute multipliers. He defines the company's focus as total effective compute, the product of physical capacity and the multipliers it discovers, so each new multiplier enlarges the effective pool. His illustration: if a model of a given capability needs 100 units of compute, a 4x multiplier means the same model can be trained on 25 units, or the same 100 physical units can train a more capable model that would previously have required 400. Anthropic argues these multipliers compound and build on one another, creating an advantage that is hard to replicate, and Brown suggests the process could accelerate as the company approaches recursive self-improvement, with Claude discovering multipliers on its own.
A September postmortem of three recent issues offered an earlier hint of this multi-chip setup, describing bugs that affected only the TPU deployments. That kind of platform-specific failure is a natural cost of the strategy Brown describes: a multi-chip fleet buys supply flexibility and negotiating leverage but multiplies the surface area for bugs, since each stack behaves differently under load and failure.
This suggests the story is really about supply and efficiency rather than raw chip counts. Spreading workloads across Trainium, TPUs, NVIDIA and AMD reduces dependence on any single vendor, and Anthropic's framing of effective compute indicates it sees algorithmic and software gains as roughly as valuable as additional hardware. The Mythos AMD result, if it holds generally, points to a future where porting to new silicon is a matter of days rather than quarters, which would make hardware choices more flexible over time.
Why it matters: teams building on Claude are indirectly affected by where and how cheaply Anthropic can run inference, and anyone operating their own models should note the implication that automated kernel porting is becoming practical (an inference from the reported three-day AMD result). If Anthropic's multi-chip and multiplier strategy works as described, expect continued pressure on both model capability and serving economics, with vendor lock-in on the infrastructure side looking less necessary than it did a few years ago.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Anthropic's IPO roadshow video Anthropic kicked off their IPO roadshow yesterday, and their roadshow video has some previously unreleased information about how their business works. The section on "Approach to compute" with co-founder Tom Brown is particularly interesting, describing how Anthropic train and serve Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. I'm not sure why there's no official video transcript, but here's one that I extracted using MacWhisper. Highlights below: We train and s