China's EV makers pivot to humanoids as car margins thin

CNBC reported on September 9, 2026 that China's electric-vehicle makers — XPeng, BYD, Nio, Xiaomi, Li Auto and Geely among them — are pushing humanoid robot programmes forward as domestic car sales growth cools. The South China Morning Post has been reporting the same rotation, framing it as Tesla's Chinese imitators picking up Tesla's second act as well as its first.

The reason is arithmetic, not vision. Per CNBC, average profit margins across China's vehicle manufacturing sector ran at 1.5% in the first half of 2026. XPeng shares are down more than 45% this year; BYD is off more than 13%. When the core business earns a point and a half and the market has stopped paying for volume, a second growth vector is worth building — and robotics is the one that can be built out of parts already on the shelf.

What actually transfers

Jefferies analyst Xiaoyi Lei, quoted by CNBC, estimates Chinese automakers can reuse roughly 85% of their motors, chips and smart-driving software in a humanoid. That number is high but the direction is right, and the transferable list is specific. Battery cells and pack engineering: a humanoid runs on the same chemistry, at lower current. Actuator supply chains: harmonic drives, planetary gearboxes, rare-earth magnets and precision bearings are sourced from the same tier-two Chinese vendors that feed EV motor lines. Automotive-grade manufacturing: the discipline of building at unit cost, with tolerance control and a functional-safety quality system, is genuinely rare in robotics and genuinely present in a car plant. In-house silicon: XPeng puts three of its own Turing chips into IRON for a combined 2,250 TOPS, compute it designed for driving.

XPeng's IRON scale-up is the clearest test of that thesis, because the company built a dedicated line rather than bolting robots onto a car plant. The $900M Dogotix carve-out is the financial version of the same claim: the robotics unit is now marked at roughly what public markets pay for the cars.

What does not

Three things. Manipulation data: a carmaker's fleet telemetry is driving data, and driving data does not teach a hand to seat a connector. Nobody has solved this by having sold a lot of cars. Task generality: a car does one task extremely well, and the whole automotive design method — freeze the requirement, then optimise cost against it — is the opposite of what a general-purpose robot needs. Customers: XPeng's first deployments are its own showrooms, and BYD's Xiao Di greets buyers at Di Space venues. Both are internal placements. They are useful as pilots and they are not evidence of demand.

Unitree founder Wang Xingxing, cited in the same CNBC report, put commercialisation years out and the sector's pivotal moment possibly a decade away. That is the counterweight to every production-line photograph published this month. The most defensible read is narrower than the headlines: China's EV industry has a real cost advantage in humanoid hardware, sharpened further as domestic chips move into the joints, and no particular advantage in the software or the market. Whether hardware advantage is enough depends on which half of the problem turns out to be scarce, and that is not settled.