Uber's Robotaxi Network Puts Two Workforces on One App
Uber put two labor systems into its second-quarter earnings remarks on August 5.
One was already at enormous scale. A record 10.2 million drivers and couriers earned more than $25 billion through the platform during the quarter. Uber also added more than 100 cities, while active drivers reached another record.
The other was still being assembled. Autonomous vehicles were live on Uber in seven cities. Management expected that number to reach as many as 15 by year-end. Partners had committed roughly 120,000 vehicles over the coming years, and Uber expected to deploy more than $10 billion of capital across equity investments, infrastructure, and vehicle-offtake commitments.
These figures appeared in the same prepared remarks, yet they describe different states. The 10.2 million people earned money during a completed quarter. Roughly 120,000 is a partner commitment tied to future deployments. The capital is an expectation across several categories, not an expense already paid or a count of jobs already created.
Uber calls the intended result a hybrid network. Human drivers and autonomous vehicles would receive trips through the same marketplace. Each could cover demand the other handles poorly. That operating claim sounds less dramatic than a driverless takeover, but it creates a more difficult workforce question.
This is a dispatch change, not a clean move from people to machines. Trips will be allocated among human drivers and autonomous fleets. Another group will charge and clean the vehicles, fix them, watch for exceptions, support riders, and recover stranded cars. Financial reporting rarely identifies how many people do that work, who employs them, or how they are paid.
Those omissions matter before 120,000 vehicles arrive. Investors cannot understand margins or capital intensity from a vehicle count alone, and cities need both street operations and local employment data. Drivers need to know whether a growing platform is shrinking their accessible work. Fleet partners have to staff promises that look driverless only from the passenger seat.
Two workforces appeared in one quarter
Uber’s current network already operates at a scale that makes “bridge” the wrong word for it. In the second quarter, customers completed 3.9 billion trips and generated $58 billion in gross bookings. Uber reported $14.2 billion in revenue and $2.8 billion in free cash flow in its earnings release. Trailing 12-month free cash flow exceeded $10 billion for the first time.
Drivers and couriers produced the physical service behind most of that scale. Dividing their collective $25 billion of quarterly earnings by 10.2 million people would produce a tidy average, but not a useful income measure. Some drove nearly every week. Others delivered a few evenings. The total combines rides and delivery, markets with different prices, tips, incentives, vehicle costs, and tax treatment. It says that the labor base is huge. It does not say what a representative worker took home.
Uber’s own autonomy white paper provides an hours distribution. It says 61% of drivers and couriers spend fewer than 20 hours a week on the platform and 85% spend fewer than 40. The figures support Uber’s case that driving is often supplemental or transitional work.
Part-time status does not make a lost hour trivial. A worker may use six evening hours to cover a utility bill or use weekend driving between jobs. A decline that looks small beside full-time payroll can be large beside the purpose of the work. Any transition account therefore needs hours, earnings, and worker dependence, rather than a binary count of jobs.
A Nature Cities study published in July makes that distinction sharper. Adam Koling and his co-authors examined Uber and Lyft entry across 167 U.S. service regions from 2010 through 2019. They estimated a 3.7% increase in unstable or intermittent employment per working-age resident and a 2.9% increase in GDP per capita, with no statistically significant effect on total employment or earnings.
This is evidence about human-driven ride-hailing entry, not autonomous vehicles. It describes an earlier transition in which the platforms expanded intermittent work without clearly raising overall employment. Robotaxis now expose some of the work that ride-hailing helped create.
Uber argues that human supply remains structurally useful. Its U.S. network reaches roughly 8,000 cities, many with too little density for autonomous fleets in the foreseeable future. Human drivers can enter a market without a depot, a mapped operating domain, a dedicated charging plan, or a fixed vehicle purchase. They can appear for a holiday surge and leave when demand falls.
Suburban and less dense markets were growing about 1.5 times faster than dense markets in the quarter, according to Uber, while active drivers reached another record. This is the part of the network where flexible human supply may retain an economic advantage.
Autonomous vehicles offer another advantage in dense areas. A fleet can stay on the road across long operating windows, deliver a consistent cabin, and produce data on every mile. When demand is predictable and the operating domain is constrained, higher vehicle utilization can spread a large fixed investment across more trips.
Competition between the two workforces varies by map square and hour. They meet inside dispatch, where a rider’s location, route, accessibility needs, wait time, price, and available vehicle determine which form of labor receives the trip. Aggregate platform growth can therefore coexist with a local driver income shock.
Seven cities are not 120,000 vehicles
Uber’s autonomy disclosures contain at least four measures of scale. Seven cities described where autonomous vehicles were live on the platform in early August. Up to 15 described a year-end target. Roughly 120,000 vehicles described partner commitments over several years, while more than $10 billion described expected Uber capital across investments, infrastructure, and purchase-like commitments.
Putting the four figures into one growth chart would erase their status. A live city may have a narrow geofence, limited operating hours, safety specialists in some vehicles, or only a small fleet. A city launch does not establish commercial volume. A vehicle commitment does not establish delivery, regulatory permission, passenger trips, or positive unit economics.
Closing that distance requires operating knowledge. Uber’s AV Labs program is deploying hundreds of sensor-equipped vehicles and expects to collect millions of miles of data each month. The work may improve mapping, simulation, and partner systems, but it first creates test operations, data review, vehicle care, and engineering support.
Each capital category buys something different. Equity purchases an ownership position. Infrastructure may cover depots, charging, data systems, or market preparation. A vehicle-offtake commitment can place demand or residual-value risk on Uber without making it the manufacturer. None translates directly into payroll.
For a workforce account, the most useful unit is a market-week, not a global press-release total. A market-week can show the operating domain, hours available, vehicles active, passenger trips completed, human-driver trips, wait time, and the people required to keep both supplies usable. It can also show what changed after a fleet expanded.
Consider two city launches that each receive 500 vehicles. In a compact district with steady demand, vehicles might accumulate paid miles quickly. In a dispersed market with airport queues, steep peaks, and long repositioning distances, the same 500 can spend more time empty or unavailable. The capital count is identical. The labor and margin effects are not.
An autonomous vehicle can accept only trips within its approved domain. A human driver can cross the boundary, reach a suburb, adapt to a closure, or help with an unusual pickup. The platform benefits from that flexibility without necessarily paying for the waiting capacity before a trip appears.
Dispatch is meant to use each supply where it performs best. It can also turn drivers into the shock absorber around a capital-intensive core. Human supply catches the irregular trips, remote pickups, peak overflow, and markets that do not justify fixed assets. The fleet takes dense, repeatable routes that once helped drivers maintain utilization.
Service can improve while the mix of work left to people worsens. A driver may see roughly the same number of offers but spend more unpaid time traveling to them. Trip distance, deadhead miles, pickup difficulty, cancellation risk, and earnings per online hour become as important as trip count.
Uber’s quarterly totals cannot reveal this movement. Neither can an AV ride counter. A useful disclosure would follow matched driver cohorts before and after deployment, within the same city and season, while reporting changes to geofence and price. Without that view, a growing hybrid network can produce two confident stories from two incomplete denominators.
Uber’s best economic answer is induced demand. If an autonomous fleet shortens waits, lowers prices, or makes late-night service reliable, people may take rides they would otherwise skip. A larger trip pool could leave drivers with more earnings even after autonomous vehicles take a share. Human supply could then cover peaks that a fixed fleet would be expensive to own for.
That mechanism is plausible, not automatic. A subsidized launch can move riders from transit or another ride-hailing service without creating enough incremental work for incumbent drivers. New trips may occur almost entirely inside the AV geofence. Platform bookings can grow because delivery or another country grew, while drivers near the launch lose their best hours.
A market-week account can separate these possibilities. It should identify trips that appear after the launch, where they originated, which supply completed them, and how the change affected driver earnings per online hour. The result may support Uber’s claim. The point is to test the claim where dispatch makes it true or false.
Hertz takes the work behind the empty seat
On April 30, Uber and Hertz announced a partnership that makes some of the labor behind an empty seat visible. Hertz’s Oro platform is supposed to handle charging, routine maintenance, repairs, cleaning, and depot staffing for Uber’s planned Lucid vehicles equipped with Nuro’s autonomous system. The first San Francisco Bay Area launch was expected later in 2026.
Oro is also running a separate driver-led fleet. Its employees would operate Hertz vehicles through Uber, beginning with an Atlanta pilot and planned expansion to Los Angeles, San Francisco, and New Jersey. The partner is preparing physical support for autonomous vehicles beside an employee-driver model and Uber’s independent supply.
No disclosed total shows how many jobs this creates. The release omits headcount, wages, schedules, benefits, retention, depot locations, and a completed Bay Area outcome. “Depot staffing” remains an operating commitment until positions are funded and filled.
The announcement does reveal how accountability can split. A rider books through Uber, but the vehicle may be financed through one party, equipped by another, operated by a developer’s system, and maintained by Oro under a city permit. A worker at the depot may carry a Hertz badge while supporting an Uber-branded trip and a Nuro technical stack.
Specialization makes job quality and incident ownership harder to see. A depot role may be direct employment, an agency shift, or a subcontracted service. Maintenance may require certified technicians while cleaners work around variable fleet returns. Remote support may sit in another state, leaving the host city with fewer jobs than the announcement suggests.
Fleet work also follows a different clock from gig driving. A driver chooses when to log in, subject to the economic pressure of demand and incentives. A depot needs coverage when vehicles return, when charging slots open, and before the next demand peak. Overnight and weekend shifts are not incidental. They are part of asset utilization.
Cleaning deserves more attention than it receives in autonomy forecasts. A person sitting in the driver’s seat can notice a spill, a lost phone, smoke, a damaged belt, or a passenger who needs help before the next ride begins. A driverless fleet must detect the condition, remove the vehicle from service, route it to a capable worker, and return it quickly enough to preserve utilization.
Maintenance has a similar feedback loop. More hours on the road can increase paid utilization, yet it compresses the window for inspections, tire work, sensor cleaning, calibration, software updates, and repairs. A cheap staffing model can become an expensive availability problem. The financial model needs vehicles purchased and labor hours per active vehicle.
Uber’s white paper is unusually direct about the transition limit. It says fleet management, maintenance, remote operations, and oversight jobs are unlikely to absorb a majority of drivers affected by autonomy. The new occupations may require different skills, locations, schedules, and employment arrangements. A displaced driver cannot be treated as a future technician through arithmetic alone.
Some of those jobs may still improve the bargain for the people who get them. A scheduled depot role can pay for waiting time, remove personal-vehicle depreciation, and come with benefits that independent driving does not. A driver who values control over six evening hours may see the fixed shift as a bad trade. Job quality has to be compared from the worker’s starting point, not assigned by employment label.
Operational job announcements should therefore identify funded and filled positions, the employer, work location, pay band, schedule, and minimum requirements. Those roles can then be compared with driver hours and earnings in the same market. A depot opening and a driver income decline belong in one account, but netting them into an unsupported “jobs created” claim would conceal both.
Waymo’s advisers expose the ratio
Waymo supplied one of the few visible staffing ratios in February. Approximately 70 remote-assistance agents were on duty worldwide at a time for a fleet of about 3,000 vehicles. The fleet was then completing roughly 400,000 rides and more than four million miles each week. Its operations explainer described agents as advisers, not remote drivers.
When a vehicle encounters an unusual scene, its autonomous system can request context. An agent might help interpret intent around a blocked road or an atypical object. The vehicle continues to control steering and braking and decides how to use the information. Waymo said the connection typically operates at about 150 milliseconds of latency in the United States and 250 milliseconds abroad.
Dividing the two figures produces one person on duty for about 43 vehicles. This concurrency ratio cannot reveal how many people rotate through a week of shifts, train, supervise the floor, review quality, or staff Waymo’s emergency-response team. Rider support, depot work, cleaning, charging, maintenance, mapping, software, and city coordination also sit outside it.
Waymo’s ratio does not describe Uber’s future partner network. A vertically integrated operator can build one remote-assistance practice, while Uber expects to connect many developers and fleet partners. Shared marketplace access may simplify rider demand while multiplying operating interfaces behind it.
MITRE’s review of scaled autonomous operations identifies seven related areas: remote assistance, rider support, fleet monitoring, incident response, operational resilience, staffing ratios, and organizational models. It is a framework drawn across transportation and other safety-sensitive domains, not an estimate of jobs per vehicle. Its value is the map of work that a narrow remote-operator count leaves out.
Rider support changes when no worker is in the cabin. A locked-out passenger, a lost item, a sick rider, a confused pickup, a child-seat dispute, or an accessibility need may require voice support, a field response, or another vehicle. First responders need a reliable way to identify the operator and make the vehicle safe. Cities need a contact who can resolve repeated obstruction rather than merely log it.
An empty front seat can also be a benefit. A rider who has faced discrimination, intrusive conversation, or unsafe behavior may prefer it. Some disabled passengers gain independence when they can travel without negotiating with a driver, while others need help with a walker, a door, or the path from curb to entrance.
Declaring either human or autonomous service universally better would flatten those experiences. Accessibility measures should cover completed trips, assistance requests, cancellations, wait time, and safe arrival for different rider needs.
Safety creates a harder counterweight to the labor argument. If autonomous operation materially reduces collisions or serious injuries within a domain, protecting today’s trip allocation would be a poor reason to withhold it. The comparison still needs actual exposure, severity, operating conditions, and independent review. A projected safety benefit cannot erase a driver-income loss, just as an income loss cannot erase a verified safety gain; each requires its own response.
One trained adviser can support many vehicles instead of one driver per car. A centralized team may build expertise and stable schedules, but it may also move paid work away from the neighborhoods where trips occur. Local officials counting employment need to distinguish an operating fleet from an operating workforce located inside the city.
Job design matters as much as the ratio. An agent who watches multiple streams and handles rare, high-consequence events needs sustained attention. Long quiet intervals do not eliminate fatigue. Rapid growth can turn a manageable queue into constant task switching. The safety case therefore depends on arrival rates, response times, workload distribution, escalation quality, and staffing during abnormal conditions.
Service data requires the same measures. A fleet can report high autonomous miles while riders wait for support or vehicles sit out of service. Assistance frequency, resolution time, field responses, and paid worker hours per thousand trips show how much work sustains the ride. Without them, “driverless” identifies an empty seat rather than the workforce operating the network.
Wuhan measured the first income shock
One early city estimate comes from Wuhan. Zhen Yu and four co-authors studied the rollout of Baidu’s Apollo Go robotaxis using 206,949 daily income observations from 1,040 taxis. Their paper in Humanities and Social Sciences Communications, accepted in April and published later that month, used a difference-in-differences design on data from January through August 2024.
The authors estimated a 10.9% short-run reduction in average daily income per taxi in the treatment district. A separate driver survey reported longer working hours, greater stress, lower job satisfaction, and stronger intent to search for another job in high-coverage areas.
This is not an Uber forecast. It covers taxis in one Chinese city over eight months. Platform rules, fares, licensing, labor supply, geography, and consumer behavior differ from U.S. ride-hailing markets. The operating data came from three taxi companies rather than a public replication file. Longer-run entry and exit could change the effect.
Still, this is closer to the worker question than a count of autonomous rides. It observes income while a competing supply entered service. It also shows how workers may respond: by staying online longer, searching elsewhere, or accepting lower utilization rather than disappearing instantly from a headcount.
Uber’s white paper says its experience in Austin, Atlanta, and Phoenix has shown trips rising while driver earnings remained consistent. It does not publish figures, cohort definitions, comparison periods, or a method for those claims. The same paper acknowledges that long-run driver work will probably be reduced or different.
When discussing autonomous-only networks in San Francisco and Los Angeles, Uber says driver utilization and hourly earnings declined as AVs took rides outside a hybrid marketplace. It estimates that one autonomous vehicle can do the work of roughly four drivers in California. Both claims support Uber’s preferred network design and lack a disclosed independent evaluation.
Together, the claims expose a testable proposition. A hybrid marketplace could reduce the harm if it expands demand, routes incremental trips to drivers, and prevents autonomous supply from stripping the densest work out of the human market. It could also obscure the harm if platform-wide earnings rise while incumbents in an AV geofence lose hours.
City analysis should separate drivers active before AV launch from people who joined later. Prior hours, ride versus delivery work, income reliance, vehicle cost, and home area define useful cohorts. Their record needs online and engaged hours, offers, paid and unpaid miles, gross earnings, incentives, tips, and estimated vehicle expense.
Publication should show the distribution. An unchanged market average can hide losses among full-time drivers if new occasional workers enter, prices change, or demand grows in another part of the city. Median and percentile changes are more informative than one total. A matched comparison area can help distinguish AV effects from seasonality, a convention, fuel prices, or a fare change.
Worker paths deserve their own record. One driver may shift from rides to delivery, while another moves beyond the AV operating window or leaves the platform. A few may enter fleet operations. Each path carries a different income, schedule, and skill consequence.
Growth cannot substitute for distribution. A city can receive more rides, lower waits, and safer service while a particular worker cohort loses income. Both outcomes can be true without either cancelling the other.
A hybrid-network labor account
Uber reports marketplace scale and financial results, its partners announce vehicle programs, and cities collect permits and incidents. A hybrid-network labor account can connect these records without pretending every measure has the same owner. It should operate at the market-week level and preserve the status of every number.
| Layer | Minimum fields | Decision supported |
|---|---|---|
| Market and operating domain | Geofence, hours, weather limits, airport access, accessibility service, launch stage, permit status | Defines the shared area for human and autonomous supply |
| Human-driver work | Active drivers, online and engaged hours, offers, trips, paid and unpaid miles, gross earnings, incentives, estimated vehicle cost | Shows distributional changes in accessible work and net earning capacity |
| Autonomous service | Vehicles available, service hours, passenger trips, paid and empty miles, utilization, disengagement or assistance events | Separates a fleet commitment from productive service |
| Depot and maintenance work | Filled roles by employer, location, pay band, schedule, contract type, vacancies, labor hours per active vehicle | Tests whether physical operations can sustain utilization and creates local work |
| Remote and rider operations | On-duty and total workers, queue volume, response time, escalation, field response, training and quality review | Connects staffing with service and safety rather than a vehicle ratio alone |
| Worker transition | Driver cohorts affected, redeployment, applications, training starts, completions, placements, pay change, attrition | Tests whether announced pathways reach people whose earnings changed |
| Public outcomes | Wait time, completion, accessibility, complaints, collisions and incidents, curb obstruction, emergency response | Keeps labor efficiency beside rider and city results |
| Evidence status | Actual, estimated, committed, planned, company claim, independently evaluated; source and review date | Stops unlike numbers from becoming one automation total |
Evidence status prevents commitments from turning into outputs. “120,000 committed” should never occupy the same cell as “vehicles active,” and “depot staffing announced” is not “workers employed.” An executive claim about stable earnings belongs beside the method and period needed to verify it.
The first window should cover eight or 12 weeks before a material deployment change. Use the same period afterward and select a comparison area if one is credible. Fare, incentive, geofence, demand, and regulatory changes stay in the record. A dashboard that begins on launch day has no worker baseline.
Avoid forcing every labor effect into job equivalents. Gig work varies by hour and worker. Fleet work may be scheduled employment. Report driver hours and earnings lost or gained, then report filled operations roles and paid hours separately. A conversion into full-time equivalents can be an additional view, provided the assumed hours are explicit.
Suppose a hypothetical city expands from 200 to 350 active autonomous vehicles. Over the next 12 weeks, passenger trips rise 9% and average wait time falls. Incumbent drivers inside the geofence complete 6% fewer engaged hours and earn 4% less per online hour after incentives. Drivers based outside it see no material change.
The operator also fills 48 depot roles, 12 maintenance roles, and 20 remote or rider-support roles across two locations. Only 30 roles sit in the launch city. Fifteen affected drivers apply, six begin training, and two accept positions at the end of the period. These figures are invented for illustration, but they show why a single “jobs impact” number would mislead.
Management can now test whether lower wait time came from utilization or a fare subsidy and whether drivers lost dense trips while accumulating unpaid miles. Fleet wages show whether a transition was plausible. Remote queue performance and accessibility outcomes show whether the operational gain reached riders.
Finance gets a better capital view as well. Vehicle utilization depends on depot throughput and downtime. Maintenance, cleaning, support, insurance, empty repositioning, and financing may offset some cabin labor savings. Those costs belong in the model rather than disproving it.
Cities can make expansion conditional on the same account. A larger geofence or higher vehicle cap can trigger a dated review of service, incidents, curb behavior, local workforce, and driver distribution. The review need not grant existing workers a veto over technology. It can require the operator to show what changed before receiving the next operating privilege.
Worker transition offers should be measured from invitation to durable placement. Report training invitations, applications, starts, completions, accepted jobs, 90-day retention, and weekly pay as separate stages. A role that cuts weekly income or requires relocation may be a valid choice, but it is not equivalent replacement.
Privacy sets a limit on this account. Cities and investors do not need individual trip histories or named workers. Uber can publish aggregated cohorts with minimum group sizes and independent evaluation. Workers should be able to inspect how their cohort was defined and challenge a material omission without exposing another person’s record.
Responsibility cannot end at the platform boundary. Uber controls dispatch and market design; fleet partners control vehicle availability; developers control the autonomous system; depot operators control shifts. Each party should supply its part of the account, while one named executive owns the combined review.
This account is less elegant than a vehicle count, but it describes the business being built: a transportation product that allocates paid work among people and fleets.
An airport pickup beyond the geofence
Picture the app at 11:40 p.m. after a delayed flight.
Rain has pushed arrivals toward the curb. The passenger is traveling with a folding walker and a large bag. Autonomous vehicles are operating downtown, but the airport sits outside their approved domain tonight. A human driver 18 minutes away accepts the trip.
The ride shows why Uber expects both supplies to remain. A flexible driver covers a place, hour, and assistance need that the fixed fleet cannot yet serve. It also shows the risk in the work left behind. The driver travels unpaid toward a congested pickup, waits through curb control, loads the bag, crosses the geofence, and returns without certainty of another fare.
Elsewhere, an autonomous vehicle finishes a dense downtown trip. A remote adviser has helped it interpret a temporary closure. It returns to a depot where an overnight worker cleans a wet floor, inspects a sensor, and connects it to a charger before the morning peak.
Between the two rides, paid work occurred at the network’s edge and inside the depot. A remote adviser contributed judgment during the exception. Only the first task happened behind a steering wheel, and none appears in a commitment for 120,000 vehicles.
Uber can build a useful service from that combination. More coverage, lower waits, fewer collisions, and new mobility for people who cannot drive would be real gains. Capital may turn those gains into a large business.
The workforce bargain will be decided trip by trip before it appears in an annual report. Dense work can move to a fleet while irregular work remains with drivers. Operations jobs may grow somewhere else, under another employer and schedule, as platform totals rise across every category.
The app already knows which supply received the airport request. A credible hybrid network should also know who did the work around it, what each worker earned, what changed after deployment, and whether the next expansion improved the service without hiding the transfer.
When that delayed passenger reaches home, the ride will look simple. The labor account should not.
This article examines Uber’s second-quarter autonomy plan alongside current evidence on fleet operations, remote assistance, platform work, and driver income. Published August 14, 2026.