Per vehicle profit on the 7th generation fleet in Guangzhou is RMB 299 a day at 23 rides, with similar favorable economics in Shenzhen, Pony AI says.
Pony AI's seventh-generation robotaxi fleet in Guangzhou is now generating per-vehicle profit, roughly RMB 299 in net revenue per car per day at 23 daily orders per vehicle, the company said on its Q3 2025 earnings call and in a 36kr interview with CEO James Peng. The numbers are city-specific and net of discounts and refunds, management said, but they remain the first per-vehicle profit claim from a Chinese robotaxi operator that has been running paid public service at scale.
The deeper claim Peng made to 36kr is that removing the driver was never the hard part. The hard part is everything that comes after: the unglamorous operational layer of charging, cleaning, maintenance, passenger handling, and remote assistance that has to be built from scratch the moment the human behind the wheel disappears. That infrastructure, he argued, is the actual moat. And it is the part that traditional automakers (the OEMs that build cars) and ride-hailing platforms like Didi systematically underestimate, because they confuse fleet management with robotaxi operations.
Pony AI has staked its strategy on five bets that run against the prevailing consensus in Chinese autonomous driving. In the 36kr interview, Peng laid them out as deliberate anti-consensus choices.
First, Pony AI declined to pivot into the L2 supplier business, selling Level-2 advanced driver-assistance systems to automakers, where many former robotaxi-first peers have moved to chase volume. Peng described L2 as a low-margin price war driven by OEMs that use the technology as a marketing feature rather than a product. Second, the company kept a portfolio of smaller, specialized models instead of consolidating into one large end-to-end foundation model of the type that has dominated the LLM, or large language model, era. Third, it held onto a lightweight-map stack, keeping high-definition maps as a reference layer even as much of the Chinese ADAS industry moved toward map-light or map-free architectures. Fourth, the company declined to chase embodied AI (the field of giving AI systems physical bodies, such as humanoid robots) as the next land grab. Peng called it another ten-year cycle. Fifth, it rejected the assumption that ride-hailing platforms or automakers would naturally run robotaxi services, on the grounds that post-driver operations are not an extension of fleet management but a from-scratch build.
Peng saved his sharpest lines for Tesla and the legacy automaker approach. In the 36kr Q&A, he was quoted as saying that announcing robotaxi plans is the easy part ("宣布总是一件容易的事"), and that "Tesla has been shouting about it for ten years. Has it actually done it?" He also pointed out that even Tesla still uses maps for navigation in unfamiliar territory. ("You know the roads you drive often, but unfamiliar places tire you out. … Even Tesla uses maps.") General Motors' Cruise, by contrast, has effectively collapsed as a robotaxi operator after a high-profile accident and regulatory shutdown.
The operational argument is what Peng framed as Pony AI's capability-blind-spot insight. A robotaxi that drives itself still needs to be charged, cleaned, maintained, recovered when stranded, supported remotely when it cannot resolve an edge case, and integrated with passenger pickup and drop-off logistics. Per-vehicle cost of remote assistance is targeted at one operator per 30 vehicles by year-end 2025, the company said on its Q3 2025 call. The seventh-generation vehicle's bill of materials is 70% cheaper than the prior generation, with another 20% reduction planned for 2026 production. Over 600 Gen-7 units had been produced by November 2025. None of this is the kind of headline-grabbing technology story that drives autonomous-driving coverage, and that is exactly the point.
The financial picture behind the per-vehicle claim is mixed but real. Q3 2025 robotaxi revenue reached $6.7 million, up 89.5% year-over-year and 338.7% quarter-over-quarter. Fare-charging revenue, excluding the Gen-7 commercial fleet, was up 233.3% year-over-year. Robotruck revenue hit $10.2 million. Licensing and application revenue was $8.6 million, up 354.6%. Total revenue of $25.4 million was up 72% year-over-year on a gross margin of 18.4%, compared with 9.2% a year earlier. Cash on hand stood at $587.7 million as of September 30, 2025, after a Hong Kong dual primary listing that raised more than $800 million, the largest autonomous-driving IPO globally that year, per the company's release and SEC filing.
The corporate net loss remains substantial at $61.6 million in Q3 2025, up from $42.1 million a year earlier. The breakeven claim is therefore per-vehicle and per-city, not company-wide. It also reflects city-specific economics in Guangzhou and Shenzhen, where fleet density, regulatory access, and per-ride pricing are favorable. Pony AI does not yet have similar data from other Chinese cities or from its overseas deployments, which by Q3 2025 had reached eight countries with Qatar newly added.
There is also a small on-record discrepancy worth flagging. On the Q3 2025 earnings call, management said the 2026 fleet target was "over 3,000" vehicles. In the 36kr interview, Peng gave the number as 3,500. The gap is small but worth tracking. Readers should attribute each number to the source it came from rather than treat either as a settled figure.
From roughly 900 vehicles on the road in November 2025, with a year-end target of more than 1,000, the company is betting that its ten-year head start on the operational layer compounds faster than the technology lead of any new entrant. The argument is that an automaker can buy or license an autonomous-driving stack from a supplier, but it cannot buy ten years of learning how to clean a stranded robotaxi at 2 a.m. or hand off a confused passenger to a remote operator in three seconds.
The competitive watch items are concrete. Tesla's Robotaxi service in Austin has begun paid public rides but remains small and tightly geofenced. XPeng and Geely have publicly announced robotaxi programs but have not put paid fleets into commercial operation at the same density. Cruise is out of the running. China's automated-logistics namesakes (White Rhino, Neolix, and Jiushi) have raised capital but operate in adjacent verticals, not passenger service. Pony AI's contention is that the real competition is not the autonomous-driving leaderboard but the operational one: who can run a robotaxi fleet with a single human in the loop per thirty vehicles, every day, in every city, at per-vehicle profit.
What to watch next: whether the RMB 299 per-day Guangzhou economics hold as the fleet scales toward 3,000-plus vehicles in 2026, whether the Gen-7 BOM target of another 20% reduction lands on schedule, and whether Pony AI can replicate the per-vehicle profit number outside its two strongest Chinese cities.