VCs on Physical AI: brain architectures haven't converged, data needs are 6 orders of magnitude beyond LLMs

创业邦 · wechat · 2026-10-15

At a panel during the 5th Global Digital Trade Expo, investors Cao Guoxiong, Shen Jin and Wang Xiang discussed Physical AI's path to real deployment.

Stage: Cao says it's still in a validation phase—demos work but real-scenario loops are unproven; Shen applies the S-curve, with the 15% early-adopter tipping point years away; Wang stresses household robots hinge on the "brain" model, not the body.
Scenarios: First wins are where humans can't go—hotel delivery is solved, cleaning and hazardous-site inspection are刚需-like demand; embodiment will eventually restructure production workflows. The investment anchor is real deployments generating proprietary real-machine data flywheels, not demo products.
Data gap: Wang notes Physical AI data should be 6 orders of magnitude beyond trillion-token LLM corpora, yet physical-world data is near zero; in-deployment collection is the most efficient path, and EEG/muscle-signal labeling could be revolutionary.
Infrastructure: Shen lists six pillars—compute, data engineering, models, simulation/synthetic data, runtime environment (safety), and hardware; safety is seen as the final blocker and few companies work on it.
Patience: Embodiment's brain loop likely needs 5+ years; hard tech cycles run 10-15 years.

More from AGI Musings