Lambert at Harvard: Chinese labs lead open-weight AI race

Nathan Lambert told a Harvard class that Chinese labs now anchor the open-weight model ecosystem while US frontier labs still hold a wide compute lead, in a guest lecture delivered October 22 for CS 2881R: AI Safety.

By his Artificial Analysis Intelligence Index numbers, MiMo-V2.6-Pro scores 46, DeepSeek V4.1 Flash and GLM-5.3 both at 45, and Kimi K3 at 44, putting a bench of Chinese releases within striking distance of top closed US systems. Open models as a whole, he said, 'trail the frontier by a few months (~3 months via Artificial Analysis).'

The compute picture pushes the other way. Lambert put leading US labs (OpenAI, Google DeepMind, Anthropic) at 1.2M to 1.7M H100-equivalents each, versus roughly 71k to 167k for the leading Chinese labs, a gap he framed as a '16× average total-compute advantage.'

On distillation Lambert was deliberately cool. He flagged 'very little public information/research on the impact of distillation,' and said Chinese labs run 'dedicated teams to circumvent the intended use of the API, extracting reasoning tokens' from closed systems, while calling the measurable effect on frontier capability unclear. Two of the researchers in our Who's Who directory had already passed around the slide deck.

He tied the arc back to history, framing openness through the FSF (1985) and OSI (1998) traditions and arguing that 'openness lead us to so much of this AI revolution.' In his OpenRouter data, open-model inference grew from roughly 1.5 trillion tokens per day in January 2025 to about 2.8 trillion by December 2025.