The Frontier-Lab Landscape — US vs China
A full map of the US–China frontier-AI race — who can train at the frontier at all, why compute is still the master chokepoint, and where the gap is closing.
The frontier is no longer “many AI companies.” It’s a small club with different bottlenecks on each side of the Pacific. This report maps that club — the US and China (美中) rosters, the value chain where the stack narrows to chokepoints, and the single dependency that decides everything: compute.
The core thesis is that frontier AI inherits the semiconductor chokepoint. The model race looks like software, but its decisive chain is still physical: lithography → foundry → packaging → memory → accelerator → power → labs. The US still leads at the absolute frontier; China compresses the gap through implementation speed, open weights, and aggressive price. And the next chokepoint isn’t chips at all — it’s permitted deployment, as model access itself becomes export-controlled.
The best single sentence: China is close enough to disrupt pricing and adoption, but not yet structurally independent enough to own the absolute frontier repeatedly. Inside: Team USA and Team China rosters, a price ladder, the open-vs-closed gating timeline, the capital and power walls, and what all of it means for decentralized compute.