The Carve-Out Play

The frontier-model lab is learning to deploy itself in vertical silence. 36Kr's report on ByteDance, asked about autonomous driving, surfaced a denial that closed no doors. The carve-out separates physical AI from the consumer AV product, and the gap is where the business lives.

The wording 36Kr carried splits the question: plenty of "early research and exploration" in physical AI, no plan for an "intelligent driving business." One sentence accepts a domain; the other rejects a product. That is the design choice, on the record.

Call it the Carve-Out Play: a foundation-model lab reaches a physical-AI surface without ever shipping a car, a robot, or a consumer brand. The on-ramp is an existing B2B cloud unit — here Volcano Engine's auto-industry line. The model work rides in from Seed's world-model team under Zhou Chang. The first deployed surface is unmanned logistics: fleets, not passengers; forklifts, not robotaxis; one B2B customer absorbing the cost of a frontier model so the regulator never has to ask whether the lab is a car company.

Every frontier lab sits on a stack whose marginal compute is too cheap to monetize in chat and too expensive to idle. Vertical B2B is the cleanest lane where a world model gets priced like infrastructure, not a subscription. The denial survives because the legal exposure sits in the consumer product, not the API.

The 36Kr report, in the same breath that disclaims driving, opens physical AI. The footprint is the design doc. The lab is not entering mobility; it is escaping the model-compression grind by climbing down into a stack that still has a margin.