A week in Beijing and Shanghai with the people building AI in China

A week spent with model labs, robotics teams, academics, VCs and founders in Beijing and Shanghai yields a single, stark argument: Chinese AI is shaped by “involution” - intense internal competition and state propping that yields engineering excellence but little domestic economic payoff unless exported. The domestic economy - weak consumer demand, a real estate hangover and 17.9% youth unemployment - frames AI as the one prioritized sector, buoyed by subsidized power, fast-tracked listings and price wars. Persistent compute scarcity (US effective compute estimated >10x China) forces trade-offs: slower training, constrained experimentation, gated models like Kimi K3 and infrastructure workarounds such as modular data centers in Southeast Asia. Alibaba’s in-house PPU and Moonshot/other gating exemplify capability under constraint.

That scarcity explains distribution and openness strategies: open-sourcing weaker weights is a survival wedge when you lack frontier compute, while labs with distribution (ByteDance’s Doubao, ~300M users) keep flagships closed; others (Alibaba, Zhipu, MiniMax) hybridize openness with API limits or revenue thresholds. Benchmark parity is treated skeptically; a clear gap remains on long-horizon, agentic tasks (only ~10-20% complete) because China lacks an interoperable RL environment - B2B SaaS and open APIs are rare, and companies fear sharing data. North American usage commonly routes through third-party providers (Fireworks, Bedrock) as compliance airlocks, often with little revenue back to the labs. The upshot: Chinese labs are technically formidable but structurally constrained by compute, ecosystem closure, and dependence on export to realize value.