The Cheapest Robotaxi Fleet Is One Nobody Buys

Based on the research ofWang, Zhang, and Ma, "Sharing economy in the era of full automation: Evidence from autonomous vehicle on-demand mobility services," arXiv preprint, 2026

Every Parked Robotaxi Is a Balance-Sheet Liability

Waymo now operates more than 4,000 vehicles across 14 U.S. cities, up from roughly 3,067 as of a December 2025 filing with federal safety regulators. Over the same stretch, its weekly paid rides climbed tenfold, from 50,000 in May 2024 to 500,000 by March 2026, with the company targeting 1 million a week by the end of this year. Read those two numbers together and the story is not really about robots driving themselves. It is about a company squeezing far more utilization out of a fleet that grew much slower than its ridership, because every additional vehicle is sensors, compute, a depot slot, insurance, and maintenance staff, whether or not anyone is riding in it at 3 a.m.

Tesla's Austin pilot shows what happens when that capital constraint bites harder. Elon Musk told investors the service would have 500 vehicles running in Austin by the end of 2025. The actual count sat closer to 40, and of those, only four to eight vehicles were verified operating without a safety monitor as of this spring, still under Tesla's remote supervision. The company's geofence has grown to roughly 245 square miles, about twelve times its original footprint, but expanding the map has proven far cheaper than expanding the fleet running inside it. Even Waymo, with its commanding lead, opens each new city cautiously: its newest markets, San Diego, Tampa, and Denver, launched this month with what the company describes as "dozens" of vehicles, with "hundreds over time" promised but not delivered on day one. Every operator in this race is discovering the same constraint at a different scale: a platform-owned fleet is a balance sheet, and balance sheets do not scale as fast as demand.

What the Paper Actually Models

This is the gap Wang, Zhang, and Ma step into. Their paper does not describe an existing product; it builds a time-expanded network flow model of a hypothetical arrangement they call AV crowdsourcing, in which private owners of autonomous vehicles lease their cars to a mobility-on-demand platform during the hours they are not using them personally, much as an Airbnb host lists a spare room only for the nights it sits empty. The authors solve for exactly when this arrangement beats a platform simply owning its own fleet, and they find it depends on three things: how well owner travel patterns complement passenger demand patterns, how much slack time an owner is willing to reserve for the platform, and how far a crowdsourced vehicle has to reposition itself between an owner's drop-off and a rider's pickup. Get those three right, and crowdsourcing genuinely lowers the platform's operating cost. Get them wrong, and the promised savings are eaten by empty repositioning miles and vehicles that are elsewhere exactly when demand shows up.

It is worth noting how close this already sits to a real, stated corporate plan rather than a purely academic exercise. On Tesla's July 2025 earnings call, Musk told investors that owners would be able to add their personal cars to the robotaxi network "confidently next year," explicitly comparing it to renting out a home when you are not there. That promise is now well over a year old. Tesla's own directly owned Austin fleet still has not reached the count Musk projected for 2025, let alone opened the network to vehicles it does not control. The paper's model describes the arrangement Tesla proposed and has not yet built; the real world so far shows only the capital-intensive, fully owned version straining to scale even by itself.

Chicago's Verdict: There Was Never a Flat Rate on the Table

The paper's case study, built on Chicago travel data, finds substantial heterogeneity in optimal prices and service quality across both geography and time of day. That alone would sound like bad news for anyone outside downtown. What the model actually shows is more interesting: centralized dispatch can serve heavy downtown demand and maintain relatively high service quality in peripheral areas simultaneously, but only by letting price and quality float, never by holding either one flat across the city. A single citywide fare or a single promised wait time is not the safer, more equitable choice in this model. It is the choice that would force the platform to shortchange one side or the other, because a flat number cannot represent two neighborhoods with opposite supply-and-demand pictures at the same hour.

A flat, one-price-fits-the-city robotaxi service is not the cautious option. In a model where supply and demand vary this much by block and hour, it is the option that guarantees somebody gets shortchanged.

Real deployments are already living this reality without any owner-crowdsourcing at all. Zoox's first paid fares, launching in Las Vegas this August after nearly a year of free rides, are priced from a base fare plus distance and time, with extra destination fees for trips like airport runs, while its San Francisco service remains free and limited to selected riders with no announced date for paid expansion. Waymo runs over 100 vehicles across roughly 90 square miles in Austin with a reported 4.9-star average, a maturity level its brand-new San Diego, Tampa, and Denver markets will not match for months. Tesla's unsupervised service exists in only a sliver of its own expanding geofence. None of these companies has built a uniform product; they have built a sequence of unevenly resourced local products under one brand, and riders in a newly opened city are, in effect, getting a lower tier of the same service that a Phoenix or San Francisco rider takes for granted. The paper's finding is that this unevenness is not a rollout hiccup to be smoothed away. It is close to the efficient outcome, and pretending otherwise with a flat national price or a flat promised wait time would likely make service in the weaker markets worse, not better.

That unevenness carries real political weight already, without a single crowdsourced vehicle on the road. New York's governor shelved a proposal that would have allowed commercial robotaxi service in parts of the state this February, and New York City separately let a Waymo testing permit lapse this spring. A crowdsourcing model built on top of that unevenness adds a new wrinkle: if owner participation clusters in wealthier neighborhoods with more private AV ownership, then a platform's cheapest supply, and by extension its best-served rider pockets, would track owner geography as much as rider demand. That is a data problem operators can measure before regulators make them explain it.

The Ownership Question Nobody Is Pricing Yet

There is a second, quieter shift underway that points at the same tension. Uber and Waymo are unwinding their exclusivity arrangement in Atlanta and Austin, effective 2028, after Waymo built enough direct rider trust to expand into nine other markets without needing Uber's app at all; Uber, meanwhile, has been signing its own deals with other AV developers, including Waabi, Wayve, Nuro, and Rivian, so it can put non-Waymo vehicles on its own platform. Waymo separately struck a non-exclusive deal with Lyft for robotaxi service in Nashville last year. None of this is owner-crowdsourcing in the sense the paper models; it is still fleets owned by AV developers, just decoupled from any single ride-hailing app. But it is the industry practicing the underlying skill the paper's model depends on: treating vehicle supply and demand-side distribution as separable businesses that can be matched flexibly rather than bundled into one vertically integrated company. The piece still missing is the one Musk promised and has not delivered, letting supply come from vehicles the platform never bought at all. Whoever solves that matching problem for real, at the complementary zones and hours the model identifies, gets a structurally cheaper cost base than anyone still racing to simply buy more cars.

The uncomfortable part of this paper's argument is that it undercuts two comfortable stories at once. It undercuts the fleet-race story, in which the winner is whoever manufactures and owns the most robotaxis fastest, by showing a leaner, borrowed-supply model can beat pure ownership on cost. And it undercuts the equity story, in which uniform citywide pricing and service would be the fair outcome, by showing that uniformity is precisely what would make the periphery worse off. What is left is a harder, more honest claim: the mobility platform that wins this decade will not be the one with the biggest fleet or the flattest price sheet. It will be the one that works out, block by block and hour by hour, exactly where somebody else's idle car is worth more to them than to its owner, and prices that difference openly instead of pretending it does not exist.