Anthropic’s IPO roadshow video

Anthropic began its initial public offering roadshow yesterday with a video that reveals previously unreleased details about their computing infrastructure. The segment…

By Vane November 11, 2026 2 min read

Anthropic began its initial public offering roadshow yesterday with a video that reveals previously unreleased details about their computing infrastructure.

The segment featuring co-founder Tom Brown outlines how the company trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.

No official transcript was released, but Simon Williston extracted one using MacWhisper. The key points follow.

Multi-chip strategy

The team now runs Claude on all four major chip families. This outcome stems from an early decision to support multiple hardware types.

Many observers previously dismissed the effort as unnecessary. Each chip requires its own tailored software and inference stacks, a costly and difficult engineering task.

Anthropic has spent years building this capability and claims to have mastered it. As one example, the company started working with AMD GPUs this year.

Historically, adding a new chip platform required dozens of engineers and months of work. This time, the system called Mythos handled the integration.

Mythos got Claude running on AMD in just three days with basically no human intervention. Claude ported or wrote almost 100% of the required kernels.

Compute multipliers

The second driver of their compute advantage is the rate of discovery for compute multipliers. The company focuses on total effective compute, calculated as the product of physical capacity and discovered multipliers.

Each new multiplier expands the pool of total effective compute. For instance, if a model requires 100 units of compute to train, a 4x multiplier allows the same result using only 25 units.

Alternatively, the same 100 physical units could train a more intelligent model that would have previously needed 400 units.

Compute multipliers create a competitive advantage that is hard to replicate. They build on each other, where each new discovery improves the efficiency of previous ones.

The team believes this process can accelerate further as they approach recursive self-improvement. Eventually, Claude will be able to discover compute multipliers entirely by itself.

This autonomy would allow the number of compute multipliers to grow faster and faster.

Readers got a hint at this multi-chip strategy last September. Anthropic described bugs affecting just their TPU deployments in a postmortem of three recent issues.

Scroll to Top