A week in Beijing and Shanghai with the people building AI in China
8 hours ago
- Traffic in Beijing is light due to economic slowdown, with many young graduates driving Didi as the only available job; youth unemployment is 17.9%.
- Chinese AI is the only working sector of the struggling economy, heavily state-subsidized, but characterized by 'involution'—intense competition with no profit.
- Compute scarcity forces Chinese labs to be efficient, but they lack high-performance chips; open-source models are a necessity for distribution, not a choice.
- Benchmark parity is distrusted; the real gap is in long-horizon agentic work due to a lack of an RL environment caused by China's closed tech ecosystem.
- Chinese labs use US inference providers as compliance airlocks; they excel at on-prem and on-device deployment, which is now valuable for sovereign AI.
- The idea of a unified 'China's AI strategy' is a myth; domestic competition is more brutal than competition with US labs, with hundreds of labs undercutting each other.
- Chinese capital markets are broken with RMB and USD funds not mixing; valuations are wild, and the Manus acquisition unwinding highlights regulatory risks.
- Embodied AI sees US intelligence running on Chinese hardware; China has advantages in energy, shipping cadence, institutionalized criticism, and studying US closely.
- The bear case: Chinese gains rely on distilling US models, the chip gap is insurmountable, and capital markets demand early monetization, limiting long-term R&D.
- Implications for US builders: own deployment, expect cheap Chinese tokens but lag in agentic work; integration surfaces and data loops are key moats.