Anthropic shows how it runs Claude on four types of chips at once — and did it faster this year

What you could do with it

• AI companies could switch between chip makers like AWS, Google, NVIDIA, or AMD if one runs out of supply or raises prices, which could keep their computing costs down for users ¹.
• Researchers could build custom tools like Anthropic's Mythos to get AI models running on new chip hardware in just a few days instead of requiring dozens of engineers working for months ¹.
• Companies pursuing AI development could stack up "compute multipliers" — optimizations that let the same physical chips do more work, either cutting compute usage or enabling them to train more capable models from the same hardware ¹.

Running on every major chip

Anthropic kicked off its IPO roadshow on November 10, 2026, releasing a video that included information not previously shared publicly ¹.

In a section about how the company handles computing, co-founder Tom Brown explained that Anthropic now trains and serves Claude across all four major chip families: Amazon's AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs ¹.

This is unusual. When Anthropic first started doing this, many people in the industry thought running AI models on different chips was not worth the effort. Each type of chip needs its own custom software stack for both training and running the model, which is expensive and difficult to build. Anthropic stuck with the strategy anyway, and after years of doing it, says it has become very good at it ¹.

The company had already hinted at this multi-chip approach last September, when it published a postmortem about three infrastructure bugs — some of them affecting only their TPU deployments ². That document noted Anthropic serves Claude through its own API, Amazon Bedrock, and Google Cloud's Vertex AI, across AWS Trainium, NVIDIA GPUs, and Google TPUs .

A new chip took three days instead of months

The most striking example came this year, when Anthropic began working with AMD GPUs for the first time ¹.

In the past, bringing a new chip platform up to speed required dozens of engineers working for months. This time, Anthropic used a tool built into its own systems called Mythos ¹. Mythos got Claude running on AMD hardware in just three days, with almost no human help . According to Brown, Claude "ported or wrote almost 100% of the kernels" — the small pieces of code that tell the chip exactly what to do .

Multiplying compute without buying more chips

Brown also described Anthropic's second source of advantage: discovering what he calls "compute multipliers" ¹. These are improvements that let the same physical chips do more work.

He offered a straightforward illustration: when a firm typically requires 100 units of compute power to build a model at a specific intelligence tier, a 4x multiplier opens up two paths—it could cut its usage down to just 25 units while achieving the same outcome, or it could deploy the full 100 units to produce a more capable model that would have demanded 400 units before ¹.

These multipliers build on each other, creating a kind of snowball effect. Brown believes the process could accelerate further as Anthropic approaches recursive self-improvement — a stage where the AI improves itself — because at that point Claude would be able to discover compute multipliers entirely on its own ¹.

Words to know

Compute multiplier: a way to make chips do more useful work without adding physical hardware · Kernels: small programs that run directly on a chip to carry out calculations · IPO roadshow: a series of presentations a company gives to investors before going public

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

Key facts

• Anthropic trains and serves Claude on four chip types: AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs
• Adding AMD support took three days using Mythos, compared to months and dozens of engineers in the past
• Claude wrote nearly 100% of the kernels for the AMD port with almost no human intervention
• Anthropic co-founder Tom Brown said the company is approaching recursive self-improvement, where Claude could discover compute multipliers on its own

Sources

1. Anthropic’s IPO roadshow video — simonwillison.net
2. A postmortem of three recent issues