In the summer of 2026, China’s auto industry is moving its gaze from the steering wheel to two legs. On 24 August, XPeng said its humanoid business closed a first round above USD 900 million at over USD 6.3 billion post-money, with Tencent and Alibaba rarely joining together. Around the same time BYD’s first commercial service humanoid “Xiaodi” made its live debut. Chery said its MoJia Robotics passed 3,000 cumulative deliveries, 2,000 overseas, across 60-plus countries, and started IPO prep.
From Chery, XPeng and BYD to Li Auto, Xiaomi and Nio, in a short span over a dozen major carmakers entered humanoids via self-development, incubation or investment. The speed and resolve beat expectations. But few ask: can car logic build robots?
“Selling cars no longer feeds the carmakers”
To read the migration, see the numbers. On 27 August, Cui Dongshu disclosed that January to July 2026 auto-industry revenue was RMB 6.078 trillion but profit only RMB 216.2 billion, a 3.6 per cent margin, versus 6.5 per cent for downstream industry. Early-year was worse: January to February margin fell to 2.9 per cent, a decade low.
A RMB 200,000 car nets the plant only RMB 3,000 to 4,000. One blogger’s data: Geely at 5.2 per cent was a top performer at RMB 6,389 per car; Changan 0.6 to 0.8 per cent, RMB 619 to 811; Leapmotor 0.55 per cent, RMB 589; many were negative. An executive said plainly: “We do not chase fat profits, but we should at least have profits.”
With NEV penetration past 50 per cent, the market shifted from growth to stock competition. Homogenisation and price wars crushed margin. Over 16 carmakers and nearly 70 models cut prices at the year’s start; the industry ministry had to step in. Policy stops the bleeding but not the cause, as lithium carbonate jumped from RMB 75,000 to RMB 200,000 a tonne and auto chips rose 180 per cent.
So they hunt a new story. Humanoids are the sexiest one. Morgan Stanley in June raised its 2026 China humanoid shipment forecast from 28,000 to 50,000, seeing 446,000 and USD 15 billion by 2030, and globally 1 billion units and USD 7.5 trillion by 2050.
On one side a 3.6 per cent margin mire, on the other a trillion-dollar ocean. He Xiaopeng’s line on becoming robot-business CEO sounded like the whole industry: “We made this call because we stand at a historic turning point.”
The trigger was Tesla. In July Tesla showed a video of the Fremont plant tearing down the Model S/X line in 46 days, freeing space for Optimus Gen 3 with a 1 million-unit annual design capacity.
“Cards others do not hold”
But anxiety alone is not enough; carmakers hold real cards. A sharp analogy: stand a smart car up and it is embodied AI and humanoids. Lidar, cameras and millimetre-wave radar at perception; end-to-end AI at decision; motors, electronics and motion control at execution, architectures are highly convergent.
“Seventy per cent of carmakers’ tech reserves translate directly to robots,” He said. XPeng’s IRON uses the same Turing AI chip as its autonomous driving, 2,250 TOPS, with XNGP perception and planning lightly adapted. BYD’s Li Ke said software complexity moves to robots “very easily”. In August “Xiaodi”, 1.61 metres with six-fingered hands, debuted in Zhengzhou.
Beyond tech transfer, carmakers hold two cards pure robot startups lack. First, supply chain and mass production: GAC’s Huilun chief said labs “over-value prototypes, under-value reliability”, while carmakers bring automotive reliability, testing and supply-chain discipline, and can pull humanoid cost from millions of yuan rapidly down. Chery’s MoJia already ships at RMB 285,800 and globally, targeting 10,000 deliveries in 2027 with IPO prep.
Second, ready-made scenario closure. Humanoid commercialisation’s hardest part is cold start: no scene, no orders; no orders, no data; no data, no iteration. Carmakers bring dual buffers, factory lines for sorting, handling, inspection and patrol, and thousands of stores for greeting and guiding. Xiaomi’s “Iron” trained in an auto plant; SAIC put “Nengzai 1” on a Buick battery line; XPeng’s IRON first serves stores and plants; Chery’s first served its own lines and overseas dealerships. Internal scenes give first orders and continuous real-world data, a flywheel labs cannot match.
“20 times the difficulty, fuzzy commercial path”
Yet the flip side is grim. “Advanced general humanoids are over 20 times harder than smart cars,” He said on 24 August. Cars are standardised, on structured roads; robots face unstructured reality, a spilled cup, a shifted part, a running child. Cars move in two dimensions; robots coordinate dozens of joints in three, exponentially harder.
So most carmaker robots stay “demo-grade”, walking and talking in showrooms, failing in real scenes. Earlier this year XPeng’s IRON lost balance and fell in its first public walk, legs twitching, lifted off by staff.
Editor’s note: This is an adapted translation of the original Sohu IT report. It has been trimmed and restructured for readability for an international business audience.