A car pulls into the kerb outside a metro station on the eastern edge of Beijing and there is nobody in the front seat. The passenger opens the back door, taps confirm on a screen, and the vehicle pulls into traffic without a word.
That is the part everyone has already seen on video. The part that decides whether the business works happens elsewhere: a depot where cars are cleaned and charged; a desk where one person watches several vehicles at once; an insurance policy; a municipal office that decides how many cars may carry paying passengers, and where.
China's robotaxi industry spent a decade proving it could drive. What it now sells abroad — to Zagreb, to London and Manchester, and this week to Astana — is the machinery around the driving.
Key takeaways
- Beijing's autonomous-driving demonstration zone ran more than 1,500 vehicles and had logged over 70 million test kilometres by October 2026 (municipal coordination office, via Xinhua, 22 October 2026).
- Baidu's Apollo Go (萝卜快跑) reached 27 cities and more than 22 million rides by May 2026 — a cumulative count, not an annual one (Xinhua, 22 October 2026).
- Pony.ai (小马智行) reported US$12.1m of robotaxi revenue in the second quarter of 2026, up 691% year on year, on a fleet of 1,975 vehicles at the end of June (company results, 18 August 2026).
- A Ministry of Transport guideline from 5 December 2023 allows fully driverless taxis with remote operators at no fewer than one per three vehicles.
- Pony.ai's 23 October 2026 memorandum with inDrive and Nvidia puts its seventh-generation robotaxi in Kazakhstan, with the local platform owning operations (China News Service).
The driving is the easy part
A driverless taxi in a Chinese city needs more than a good model: a mapped operating domain, a transport licence, a depot with chargers, a remote-operations desk, an insurer willing to write the policy, and a city willing to let it charge fares.
The Ministry of Transport set the shape of this in a guideline issued on 5 December 2023: a fully driverless taxi may run without anyone aboard only with municipal approval, inside designated areas, with remote safety operators (安全员) at no fewer than one per three vehicles, and with at least 90 seconds of pre-failure operating state recorded.
Autonomy does not remove labour; it relocates it, from the front seat to a room with screens and to depots that still need people at three in the morning. That is why the operating company, not the model, is the unit that has to make money.
Three numbers, three different questions
Robotaxi scale claims mix three measurements that answer different questions.
- 22 million rides is what Baidu had accumulated by May 2026 — endurance since the service began, not current demand, and nothing about how full the cars are.
- 1,500 vehicles and 70 million kilometres in Beijing's demonstration zone, now about 600 square kilometres, measure capacity and test exposure. They are not revenue.
- 27 cities measures regulatory footprint: each city means a separate licence, map, depot and remote team, so footprint and density pull against each other.
WeRide (文远知行) said it had nearly 2,000 test and operating vehicles in China, and Shenzhen has opened more than 2,000 kilometres of test roads (both via Xinhua, 22 October 2026).
A unit that works is not a company that works
Pony.ai's second-quarter filing is the clearest window available. Robotaxi revenue reached US$12.1m, a third of total revenue of US$36.2m, and the company said per-vehicle economics turned positive in Guangzhou and Shenzhen. Its net loss that quarter was US$45.4m, and it ended June with 1,975 robotaxis against a target of more than 3,500 by December.
Chief financial officer Wang Haojun put the break-even threshold at 40,000 to 50,000 vehicles (National Business Daily, 3 September 2026). That is about twenty times its fleet.
The gap is not a contradiction; it is a definition. Per-vehicle break-even is a contribution margin, struck before research spending, corporate overhead and the cost of opening the next city. It means a car can earn its keep on a given route, not that the company earns anything. Closing the gap requires fleet scale, and fleet scale in China is granted city by city, not ordered from a factory.
The export is the same playbook with local owners
The overseas deals now arriving share one structure. In Zagreb, Pony.ai supplies the vehicle and the driving system, Verne runs the fleet, and Uber brings demand; the commercial service opened in April 2026 (company announcement), and the partners are extending it.
In August 2026 Pony.ai and Uber said they would deploy more than 2,000 robotaxis across five European cities; London followed on 8 October, Manchester on 14 October. On 23 October Pony.ai signed a memorandum with inDrive, which operates in 47 countries and more than 1,200 cities, and with Nvidia, naming Kazakhstan and its capital Astana. That project is still being evaluated; no timetable or fleet size has been published.
The shape is not modesty. Three inputs cannot cross a border: a local transport licence, local capital, and aggregated demand. European data rules add cost to admission, Pony.ai vice-president Zhang Ning told Xinhua on 22 October 2026. What travels is the vehicle plus the operating manual: depot layout, remote-desk ratios, the 90-second incident record, the way a city contract is written.
Honest limitations
Fleet, revenue and per-vehicle figures above are company-disclosed; no operator we examined publishes an audited city-level profit-and-loss statement. Cumulative ride counts and test kilometres are not rates and cannot be compared across companies. China has no national, standardised per-kilometre disengagement table comparable to California's, so safety claims cannot be independently audited. The 40,000–50,000 break-even is a management estimate given to a newspaper, not a filed forecast, and "per-vehicle break-even" is not publicly defined.
What readers can do now
- When a fleet operator quotes city counts, ask instead for licensed vehicles and operating-area size per city, plus rides per vehicle per day.
- Before accepting a unit-economics claim, ask what the unit contains — depot, charging, insurance, remote operators — and over what period it is measured.
- If you are buying or partnering, establish who holds the transport licence, the insurance policy and the incident-reporting duty; in this model that is the operator, not the supplier.