Robotaxis Are Scaling by Splitting the Stack
Executive summary
Europe's first autonomous rides bookable through Uber went live in Zagreb on August 20. The visible product is simple: a rider requests UberX or Comfort and may be matched with a Pony.ai-powered vehicle operated by Verne. The operating structure underneath is not simple at all. Pony.ai supplies the Level 4 driving system, Verne owns and runs the local fleet, and Uber supplies demand, booking, payment and customer service.
The mainstream interpretation is that robotaxis are approaching a platform-scale breakthrough. Zagreb supports a narrower and more useful conclusion. Autonomous mobility is being unbundled into technology, fleet operations and distribution before any one participant has proved it can own the whole stack profitably across markets.
That structure can accelerate deployment. A technology developer does not need to recreate local licensing, depots and customer acquisition in every city. Uber does not need to choose one autonomous driver or finance every vehicle. A local operator can supply regulatory standing and physical operations. But modularity does not make the cost base disappear. It divides revenue, control and liability among partners while the launch still carries an on-board licensed operator and limited hours.
The contrarian thesis is that the near-term robotaxi winner may not be the company with the most vertically integrated vehicle. It may be the network that orchestrates several autonomous systems and a less glamorous layer of fleet operators. This thesis would fail if vertically integrated fleets establish superior utilization, safety and contribution margins that partners cannot match, or if contractual complexity keeps the modular model from resolving exceptions quickly.
Market context
The Zagreb launch is Europe's first instance of autonomous rides being available through the Uber app. Service operates in selected central areas from 07:00 to 21:00. A licensed operator remains behind the wheel during the phased introduction. Verne says it had already completed thousands of autonomous rides since launching its own Zagreb service in April.
One week earlier, Uber and Pony.ai expanded their agreement to target more than 2,000 vehicles across five European cities. Their description of the model is explicit: autonomous-driving technology, a mobility platform and day-to-day fleet operations can be provided by different partners, and vehicle funding can sit with different parties by market.
This is not unique to Zagreb. In Tokyo, Uber plans to combine Wayve's driver, Nissan vehicles and Hinomaru Kotsu's authorized taxi operations. Hinomaru is responsible for depots, charging, cleaning, inspection, maintenance and uptime. In the United States, Hertz affiliate Oro is preparing to manage a fleet of Lucid vehicles using Nuro technology, including repairs, cleaning, charging and depot staffing. Zurich plans pair WeRide and Uber with fleet operator Rydera.
Regulation reinforces the local layer. The European Union has a type-approval framework for fully automated vehicles, but commercial deployment still depends on practical permitting and operating rules. In June, 18 EU member states signed a declaration to coordinate cross-border testbeds and common approval principles. The European Commission described harmonization as work in progress, not a finished single-market operating license.
Findings
Finding 1
The robotaxi market is separating the autonomous driver from commercialization.
Uber's second-quarter prepared remarks argue that different autonomous developers will suit different geographies, vehicles and operating conditions. Uber positions itself as the commercialization layer: demand aggregation, dispatch, vehicle integration, fleet operations, charging, insurance, financing and local regulatory relationships.
That is more than investor framing. The company's disclosed partnerships now repeat the same architecture with different suppliers. Pony.ai and Verne in Zagreb, Wayve and Hinomaru in Tokyo, and Nuro and Oro in the United States all preserve Uber's place at the customer interface while swapping the autonomy and fleet layers.
The architecture gives Uber option value. A marketplace can route demand across human drivers and autonomous fleets, use different AV systems where each performs best and reduce the risk of betting the network on one technology supplier. Uber said AVs were live on its platform in seven cities at the end of the second quarter, potentially reaching 15 by year-end, with partners committing about 120,000 vehicles over coming years. These are company-reported commitments, not deployed fleet counts, but they show the intended breadth.
The same structure creates new dependencies. A failed ride can involve the developer that made the driving decision, the owner responsible for vehicle condition, the platform that set expectations and took payment, and the licensed operator accountable to local authorities. The customer will still expect one answer. Commercial scale therefore depends on shared event records, response rights and liability rules as much as miles driven.
| Deployment | Autonomous technology | Vehicle / fleet operator | Customer platform | Launch-stage human role |
|---|---|---|---|---|
| Zagreb | Pony.ai Gen-7 | Verne | Uber and Verne | Licensed operator on board |
| Tokyo pilot | Wayve AI Driver | Nissan vehicle / Hinomaru Kotsu | Uber | Hinomaru safety operator planned |
| U.S. Lucid program | Nuro | Lucid vehicle / Oro Mobility | Uber | Not specified in cited fleet announcement |
| Zurich plan | WeRide | Rydera | Uber | Subject to launch approval |
Source: company announcements from April-August 2026. Unit: functional role by planned deployment. Method: Blackrock Research mapping of named responsibilities. Limitations: partnership scopes may change; announced vehicles and programs are not equivalent to deployed driverless service; human-monitoring requirements vary by market and phase.
Finding 2
Rapid robotaxi revenue growth has not yet proved company-level economics.
Pony.ai's second-quarter results give the freshest financial view of the stack. Robotaxi-services revenue rose 691.2% year over year to $12.1 million, about one-third of the company's $36.2 million total revenue. Fare-charging revenue rose 849.3%, the fleet reached 1,975 vehicles at June 30, and management said revenue from joint deployments increased sequentially.
Those figures establish commercialization, not mature profitability. Pony.ai reported $6.4 million of gross profit at a 17.5% gross margin, $72.1 million of operating expenses and a $65.7 million operating loss. Capital expenditure increased to $32.2 million from $9.6 million a year earlier as the company invested in Gen-7 production, data centers and servers. Management says it has reached city-level unit-economic breakeven in several Chinese markets, but it does not publish a comparable Zagreb contribution statement.
The distinction matters because joint deployment can improve a technology provider's capital efficiency by moving vehicle ownership or fleet work to partners. It can also move costs outside the reporting boundary used to make the claim. A city can produce positive per-vehicle economics for the driver system while a fleet partner bears financing, downtime, cleaning or local staffing. A platform can gain incremental trips while pricing incentives or partner guarantees sit elsewhere.
| Q2 metric | 2025 | 2026 | Change / relationship |
|---|---|---|---|
| Robotaxi-services revenue | $1.5m | $12.1m | +691.2% |
| Total revenue | $21.5m | $36.2m | +68.8% |
| Robotaxi share of total revenue | 7.1% | 33.3% | +26.2 percentage points |
| Gross profit | $3.5m | $6.4m | +83.4% |
| Operating loss | $61.3m | $65.7m | Loss widened 7.3% |
| Capital expenditure | $9.6m | $32.2m | +235.4% |
Source: Pony AI Inc. unaudited Q2 2026 results released August 18, 2026. Period: three months ended June 30, 2025 and 2026. Units: U.S. dollars, rounded millions; percentages as reported except robotaxi revenue share and percentage-point change, calculated from disclosed values. Limitations: total company figures include robotruck and intelligent-solutions businesses; joint-deployment revenue is not separately disclosed; rounded inputs create minor calculation differences.
Finding 3
Fleet operations are becoming a distinct market rather than a temporary chore.
Removing the paid driver does not remove the work required to keep a commercial vehicle earning. Robotaxis still need financing, charging, cleaning, inspection, maintenance, tire and body repair, software updates, incident handling, depot capacity and repositioning. Uptime has to match demand by hour and neighborhood. A vehicle idle during the evening peak can erase savings generated elsewhere.
The emerging operators are not interchangeable contractors. They bring local licenses, facilities, labor systems, regulator relationships and experience managing vehicle availability. Hinomaru's role is required in part because Japanese passenger transport must be operated by authorized taxi companies. Verne handles ownership and operations in Zagreb. Hertz created Oro specifically to sell fleet orchestration across autonomous and driver-led mobility.
This layer may capture more value than current robotaxi narratives assume. If autonomy software becomes available from several credible suppliers and marketplaces aggregate demand, reliable fleet uptime becomes the scarce local capability. Conversely, the layer can be commoditized if vehicles become highly self-diagnosing, charging is abundant and marketplaces can tender operations among many suppliers.
The open question is who holds residual risk. Vehicle owners bear utilization and depreciation unless contracts shift it. Technology providers may face safety and performance warranties. Platforms may offer minimum volumes or make capital commitments to secure supply. Uber says it expects to commit more than $10 billion across investments, infrastructure and vehicle offtake over coming years while also seeking third-party financing. The strategy is asset-light only in relative terms.
Implications for operators
Mobility platforms should measure autonomous supply by fulfilled demand, not fleet announcements. The core dashboard needs paid trips per available vehicle hour, peak coverage, deadhead miles, cancellation and intervention rates, downtime by cause, customer-support cost and contribution after partner guarantees. Human and AV supply should be compared within the same operating zone and time band.
Fleet owners need contracts that distinguish technology failure from operations failure. Uptime guarantees, repair responsibility, software-update windows, battery degradation, incident data access and insurance deductibles determine whether apparent vehicle economics survive. Minimum-volume protection may be valuable, but only if it does not leave the operator holding uncapped maintenance or residual-value risk.
Autonomy developers should resist using partner capital as a substitute for end-to-end cost evidence. Report economics at three levels: autonomous-driver gross margin, vehicle contribution after direct fleet costs, and city contribution after supervision, maps, regulatory work and support. A claim of breakeven without the cost boundary is not portable across markets.
Regulators and cities should require a single accountable operating interface even when the stack is split. Customers need one incident channel. Authorities need access to a consistent record that connects the trip, vehicle state, software version, remote assistance, maintenance history and platform dispatch decision.
Risks & open questions
Zagreb is a meaningful commercial launch, but it is not yet a driverless scale proof. Operating hours are limited, the service zone is partial and a licensed operator remains in the vehicle. Thousands of rides since April do not reveal paid utilization, intervention frequency, subsidy, fleet size or contribution margin.
Company announcements provide most of the available evidence. Partnership vehicle counts describe plans, not funded and licensed deployments. Pony.ai's city-level breakeven claims are not accompanied by market-level financial statements. Uber's claim that AV presence grows category share in mature U.S. zones does not isolate causality or disclose the size of the effect.
The thesis would weaken under three conditions. First, vertically integrated providers could sustain higher utilization and lower exception cost because they control vehicle, software, fleet and customer service. Second, fragmented liability could slow approvals or make insurance uneconomic. Third, local fleet work could prove too small and standardized to support attractive returns once autonomous vehicles mature.
The evidence that would strengthen the thesis is equally specific: repeated launches using different autonomy suppliers, third-party fleet owners financing vehicles without broad guarantees, published city economics that include all direct costs, and stable customer experience across mixed human and autonomous supply.
Appendix / methodology notes
This report reviewed primary announcements and financial materials published from April through August 20, 2026 by Pony.ai, Uber, Verne, Hertz/Oro and the European Commission. Functional roles were mapped from responsibilities explicitly named in the announcements. Financial values were transcribed from Pony.ai's unaudited Q2 results; calculated shares use rounded disclosed figures.
No vehicle commitments were treated as deployments, and no city-level profitability claim was treated as company profitability. The report distinguishes a supervised commercial launch from fully driverless service. A stronger future chart should compare paid trips, vehicle hours, interventions, direct fleet cost and city contribution across deployments, but those inputs are not yet disclosed on a comparable basis.