Anthropic IPO video: Claude runs on all four chip families

The roadshow video is notable less for the IPO mechanics than for what Anthropic chose to disclose about its compute stack. Brown frames the early multi-chip bet as unconventional—many people thought running AI models on different chips wasn't worth the effort, since each chip needs its own tailored software and inference stack—and says the company has been doing it for years. The AMD example is the concrete proof point: where bringing on a new chip platform had taken dozens of engineers months, Claude got it running in three days with basically no human intervention.

Brown pairs that with a second driver: the rate at which Anthropic discovers compute multipliers, which he describes as multiplying physical capacity to raise total effective compute. He argues these multipliers stack—each new one has a positive effect on the earlier ones—and that the process could accelerate toward recursive self-improvement, with Claude discovering multipliers entirely by itself. That is a forward-looking claim from Anthropic's co-founder, not an established result.

A prior hint: last September's postmortem of three recent issues described bugs that affected just Anthropic's TPU deployments, an early sign of the multi-chip approach in practice.