338 yuan per vehicle, per day, net. Twenty-three orders. That is the number Pony.ai published for Shenzhen in February 2026 — and it is the only number in this week's flood of autonomous-vehicle announcements that tells you whether the ride actually makes money.
Everything else is geography.
Last week the story was robotaxi scaling in China and the U.S. This week the map got wider in every direction at once: London took its first paying self-driving rides, Zoox started running to and from the Las Vegas airport, WeRide showed off a Robotaxi in Riyadh and ran passengers through Punggol, and Kodiak announced its long-haul trucks are 93% "ready" to drive with nobody in the cab.
Line the announcements up and the pattern is unmistakable. Each one trumpets a milestone — a destination, a percentage, a fleet count — and each one omits the same thing: cost-per-mile or net revenue per vehicle.
Start with London, because it was sold as the West's proudest launch of the week. Wayve and Uber put 15 electric Ford Mustangs on the capital's streets picking up paying passengers, beating Waymo to commercial service in the city.
Read the fine print Uber itself supplied: "The initial launch phase features supervised autonomous rides, meaning a trained and TfL licensed private hire driver is onboard to oversee the trip."
There is a human being in the driver's seat, licensed and paid, on every single ride. Bloomberg's own reporter described the experience as "autonomous rides with hands still hovering over the steering wheel."
That is not a rounding error in the economics. That is the economics. A robotaxi service whose entire premise is removing the driver cannot claim a path to profit while it is still paying one.
A robotaxi service whose entire premise is removing the driver cannot claim a path to profit while it is still paying one.
Wayve's answer is the future, not the present. At the launch event it showed a Nissan LEAF prototype — sensors integrated into the roofline and bumpers, running the Wayve AI Driver on NVIDIA DRIVE Hyperion — that a company representative said would eventually be capable of driverless operation.
Eventually means "next year," and only "depending on several factors, including technology and regulations." The Mustang you can ride today has a person in it; the car that might not is a car you cannot ride.
So London's milestone is real motion with negative unit economics baked in by design. The stunt beats Waymo to a headline. It does not beat anyone to a profit.
Now Zoox. The Las Vegas expansion is genuinely clever positioning — its toaster-shaped robotaxi will pick up and drop off at both airport terminals near baggage claim, while Uber and Lyft riders trudge to a parking garage a five-to-ten-minute walk away.
Zoox is the only robotaxi at the Vegas airport, and it earned that spot the hard way, spending more than a decade building a vehicle with no steering wheel or pedals. Since a NHTSA exemption cleared commercial operation, it began charging for rides on August 10, opened testing in San Diego and Houston, and now claims the airport.
Here is what Zoox still has not said, a year into giving Vegas rides and a month into charging for them: what a ride costs, or what one earns.
No fare. No cost-per-mile. No per-vehicle revenue. The company that removed the human — the hard part, the expensive part, the part everyone else is still faking with a safety driver — has published none of the numbers that would prove the removal pays off.
That silence is the tell. If your airport runs were throwing off Pony.ai-style net revenue, you would lead with it, not with baggage-claim convenience.
WeRide is the interesting middle case, because it is half-transparent. At LEAP 2026 in Saudi Arabia it displayed Robotaxi GXR, Robobus, and Robosweeper, and its CFO Jennifer Li framed the moment precisely: "The next chapter is about proving scalability."
Proving scalability, note — not proving it has been proven. In Punggol the same GXR is doing first-and-last-mile runs between homes and heartland malls. The geography is spreading fast.
But WeRide, unlike Zoox, actually put an operational number on the table. In its H1 2026 report it disclosed that Q2 daily orders per vehicle exceeded 21, up 24% quarter-on-quarter, with China ride-hailing revenue up roughly 140% in the quarter.
It did not disclose per-vehicle revenue or cost. But 21 orders a day is a demand signal that lives in the same universe as Pony.ai's 23. It is the difference between a company hiding the number and a company that doesn't yet have the clean, city-level figure to publish.
Which brings us back to the only operator that showed its work.
Pony.ai's Shenzhen disclosure is granular in a way nothing else this week comes close to. 338 yuan net revenue per vehicle per day, from 23 orders — net meaning after depreciation, charging, maintenance, insurance, and remote assistance are stripped out.
The cost stack is itemized: a 270,000-yuan seventh-generation vehicle depreciated over six years works out to 105 yuan a day, plus 60 yuan energy, 40-50 yuan maintenance, and about 30 yuan for insurance and support. Set that against 338 in revenue and the city-level unit economics are positive.
Pony.ai says it crossed into per-vehicle profitability in Guangzhou in November 2025 and Shenzhen in February 2026. That is two cities, and the caveat matters: R&D — $104 million in H1 2026 alone — is not allocated into that figure, and neither is corporate overhead.
This is not company-wide profit. It is proof that the marginal robotaxi, in the right city, at the right order density, earns more than it costs to run. That is the thing the entire industry needs to demonstrate, and only Pony.ai has demonstrated it with published numbers.
Now score the field against that yardstick.
London: negative by construction — a paid driver on every ride. Zoox: unknown, and the year-long refusal to name a fare reads as a choice, not an oversight. WeRide: 21 orders a day, revenue growing 140% a quarter, no per-vehicle profit disclosed but a demand base that rhymes with Pony.ai's. Pony.ai: positive, itemized, and confined to two Chinese cities.
Which leaves Kodiak, and the most seductive number of the week.
Kodiak's long-haul Autonomy Readiness Measure hit 93% at the end of August, up from 91% in July and 84% in February. It tracks how ready the Kodiak Driver is to run long-haul with no one in the cab; what remains is "final engineering verification and validation."
That climb from 84 to 93 in six months is a real engineering signal. But read the metric for what it is: a validation percentage, not a P&L line. Readiness is not revenue. Ninety-three percent ready still has a person in the truck.
And yet the trucking path is the one to watch, precisely because of what it standardizes. A driverless long-haul truck strips out the human — the expensive, negative-economics human sitting in every London Mustang — and runs a repeatable route on a repeatable vehicle. That is the same recipe that makes Pony.ai's math close: kill the driver cost, standardize the platform, run enough volume to amortize the machine.
So here is the answer to the framing question. The two credible near-term paths to sustainable AV unit economics are driverless long-haul trucking and single-platform Chinese robotaxi fleets — because both remove the human and standardize the vehicle, which is exactly where the cost stack collapses in Pony.ai's disclosure.
The commercialization is unmistakably wider this week. London, Vegas, Riyadh, Punggol, and the American interstate all got a little more autonomous.
But wider is not the same as profitable. Outside two Chinese cities with published numbers, nobody has shown the ride makes money — and the West's proudest launch this week still has a TfL-licensed driver sitting in the seat, hands hovering over the wheel.