Third-quarter 2026 earnings call, November 18, 2026
Insurance premiums emerge as a hard safety and cost signal
The most concrete new data point from Pony AI's third-quarter call was a third-party read on its safety record. Chairman and CEO James Peng said the company recently renewed its vehicle insurance, and the annual premium per Gen 7 Robotaxi came in at 40% of the average for human-driven taxis, down from 50% last year. "This is a clear indicator of our superior safety track record," Peng said. CFO Leo Wang added that higher vehicle costs would normally push premiums up through higher repair costs, but the safety record "more than offset this upward pressure." He described the result as structural savings in fleet operating costs. In the Q&A, Wang said the per-vehicle premium for the Gen 7 fleet has fallen a further 19% this year.
PonyWorld: post-training in simulation as the core technical argument
CTO Tiancheng Lou used the call to lay out the company's technical thesis in detail. Physical AI, he said, follows the same two stages as frontier language models, pretraining and post-training. Pretraining on multimodal data builds basic driving skills but is "bounded by human driving data, leaving long-tail hallucination unresolved." Breaking through that ceiling requires reinforcement learning in post-training, where, in his words, "over 90% to 95% of core work happens."
That is the role of PonyWorld, which Lou called a "closed-loop virtual driving school." He said it now generates 10 billion kilometers of training every week. The central challenge is the sim-to-real gap: the simulated distribution of events must mirror reality, because edge cases that appear too rarely are not learned, and edge cases that are exaggerated make the driving policy overly conservative. Lou said Pony built a self-evolving pipeline in which edge cases from public-road fleet operations automatically sharpen the simulator's fidelity. He referred to the latest iteration as PonyWorld 2.0, and said it cuts the time and compute needed to train the world model by an order of magnitude.
Lou tied the system to four commercial outcomes. First, safety, using what he called intention enumeration to generate all possible intentions and trajectories of surrounding road users and stress-test the driving policy before deployment. Second, area expansion without retraining: where traffic rules and driving styles are consistent, the model generalizes with zero retraining, so growing from more than 1,000 square kilometers in Guangzhou to "tens of thousands of square kilometers" is "purely a function of deploying more vehicles." Third, international replication through localized interaction training, which Lou credited for the move to driverless operations in Zagreb within months. Fourth, cross-platform generalization, with Robotaxis and Robotrucks sharing one core architecture that adapts to heavy-duty and light trucks through post-training.
Moat and compute: not bottlenecked, and not distillable
Asked by Citi about chip restrictions, Lou said Pony is "not bottlenecked by the AI compute." On the vehicle, Gen 7 Robotaxis have run on automotive-grade chips since launch last April, which he said lowers BOM cost and energy consumption and extends range per charge. On training, he said the on-vehicle models are much smaller than large language models, and most post-training compute goes to running simulation inside the world model, which needs far less compute than pretraining. He said Pony is the first company to achieve L4 driverless operation on automotive-grade chips, and contrasted this with other L4 players that rely heavily on on-vehicle compute, implying larger models and more training compute. Lou added that Pony has evaluated more than one China-made chip for model training and has deployed its training pipeline on them: "If needed, we can train our models on those chips."
Responding to Bloomberg Intelligence on whether GPT-6-class models plugged into embodied AI could commoditize autonomous driving, Lou said embodied AI has not reached large-scale commercial deployment and that a few minutes of demo "proves nothing about operating safety at scale." He argued that language is "a slow and lossy intermediate representation for the physical control," and that foundation-model pretraining is "only the starting line." Because Pony's models run inside a closed-loop system rather than behind a public API, he said, they cannot be distilled, and a competitor would need to build the post-training infrastructure and a real-world fleet data flywheel from scratch.
Financials: revenue up 75%, losses narrowing slowly
Total revenue rose 75% year over year to $44.6 million. Robotaxi and Robotruck together made up 80% of revenue versus 66% a year ago. Robotaxi services revenue reached $17.8 million, up 165% year over year and 47% sequentially, and accounted for 40% of the total versus 26% a year earlier. Fare-charging revenue grew about 400% year over year. Robotruck revenue was $18 million, up 76% year over year and 35% sequentially, driven by freight services with Sinotrans. Intelligent Solutions revenue was $8.9 million, broadly flat.
Gross profit rose 69% to $7.9 million, a gross margin of 17.8% versus 17.5% in the second quarter. GAAP operating expenses rose 3% to $76.6 million, while non-GAAP operating expenses fell 2% to $66.2 million. Operating margin improved from negative 274% to negative 154%. The GAAP loss from operations narrowed slightly to $68.7 million from $69.7 million, and the non-GAAP loss from operations fell 8% to $58.3 million. GAAP net loss was $136.3 million, including a non-cash $77 million loss from fair value changes in trading securities. Non-GAAP net loss narrowed 14% to $48.9 million....