Fund consult notes: 9 judgments on US-China embodied AI landscape and data flywheels

aigclink · x · 2026-08-29

Nine industry observations from an embodied-intelligence consulting session with a fund:

US-China split: US leads in "brain" companies, China leads in "body" hardware
China's edge: manufacturing scenarios give unique physical-world data at scale, better in volume and quality than the US
Scenarios: industry favors semi-structured settings (warehouse logistics, convenience stores, pharmacies); fully structured industrial scenes are oversaturated
Data flywheel players: Zhiyuan, Galaxea, Xinghetu in China; Tesla Optimus, Waymo abroad — all with real robots collecting data in the field
Standard stack: WAM+VLA model fusion is now table stakes
Brain players: Ant's Lingbot series and Skild pursue one-brain-many-robots; Ant's healthcare/consumer scenes are a unique moat
Evaluation: count human interventions over long runs, and chained-action accuracy — ten 95%-accurate actions yield under 60% end-to-end
Data mix: 70% of pretraining data is internet video (no actions) plus egocentric data (ego, umi); only a small share is real-robot or sim data with actions
Unitree's gap: shipments but no data flywheel — scenes limited to performances and labs; a partnership with a brain company could change that

More from Companies & People