Anthropic details multi-chip strategy in IPO roadshow video

Training and serving across four chip families

Anthropic shared previously unreleased details about its business in its IPO roadshow video. The "approach to compute" section, presented by co-founder Tom Brown, reveals that Claude is trained and served across four major chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs.

Brown says the multi-chip strategy was an unconventional bet early on. Each chip requires its own tailored software and inference stack, which is expensive and difficult to build; Anthropic says it has refined this capability over years.

Claude ported itself to new chips

According to the company's example, it began working with AMD GPUs for the first time this year. While bringing on a new chip platform used to take dozens of engineers months of work, this time Claude was running on AMD in just three days with basically no human intervention. Claude wrote or ported almost all of the kernels.

Compute multipliers

Brown describes a second advantage: discovered compute multipliers. Anthropic views total effective compute as physical capacity multiplied by these multipliers. A 4x multiplier, for example, lets the same model be trained with a quarter of the compute, or the same hardware to train a more capable model. The company argues these multipliers stack and are hard to replicate.

Anthropic suggests that as recursive self-improvement approaches, Claude could discover these multipliers entirely on its own, accelerating the process further.