September 2026 is the month the robotaxi industry became an operating business. On 1 September Waymo opened paid, fully driverless service in Denver, San Diego and Tampa, taking it to 15 US metros and past 20 million paid rides, while California cleared expansion across 18 counties. Uber answered within days by launching London’s first robotaxis with Wayve, safety drivers aboard at first, and folding Waymo, Wayve, WeRide and Apollo Go into one network. Spain issued its first national Level 4 permits for a Madrid fleet. Days later, Tesla’s no-controls Cybercab drew an NHTSA probe.
The incumbents have finished the capability race. What now decides who scales is everything around the car: operations, economics, permitting, liability and evidence. This page maps that terrain so you can evaluate and choose rather than follow a playbook.
In This Series
- What Changes When a Robotaxi Service Really Scales: The operational shift, and how to tell real paid rides from new permits.
- The Hidden Human Economy Inside Robotaxi Per-Mile Costs: The remote assistance, depots, cleaning and deadheading that set the real floor on per-mile cost.
- The Business Model Bets That Will Decide Robotaxi Scale: The vertical-versus-aggregator contest and the trust-and-permits bet against the cost-curve bet.
- How Permits, Exemptions and Liability Gate Robotaxi Expansion: Exemptions, self-certification versus pre-market approval, and liability.
- How to Read Robotaxi Safety Claims and Rider Data Risk: Reading self-reported safety metrics and rider-data risk.
What does it actually mean for a robotaxi service to scale beyond the pilot?
Scaling beyond the pilot means running the service under real demand: fully driverless, open to the public, and run as a business. What changes is everything around the car: depot throughput, servicing, charging, geographic extent and emergency behaviour, while the binding constraint moves from the self-driving stack to permits, servicing capacity and liability. Utilisation and paid rides are the real measure of scale.
September 2026 makes that shift visible. Waymo’s 15 metros and more than 20 million paid rides read as scale, but the operational design domain (ODD), the formal boundary of where and when a vehicle may operate, still leaves the edge cases outside: cars stalled in a San Francisco blackout and paused for a San Antonio flood. Keeping the fleet running inside that box is where the operating cost hides, which the per-mile economics below unpacks.
The full treatment covers what actually changes operationally.
What does a robotaxi ride actually cost per mile?
Any single per-mile figure compresses a stack of costs: vehicle amortisation, energy, servicing, remote assistance, insurance and deadheading. The number only means something if you know the utilisation and labour assumptions behind it. The driver leaves the front seat, and the work moves to an office: charging, cleaning, maintenance and remote supervision still need people. Treat a quoted per-mile cost as a claim to interrogate against those layers.
Headline figures such as Tesla’s projected $0.20 per mile hide that stack. Hardware-cost deflation lowers the capital layer; the operating layer, where deadheading and labour concentrate, holds its own price, and even in dense San Francisco Waymo vehicles deadhead more than 40% of the time. China’s one-supervisor-to-three-vehicles rule makes the labour ratio a policy choice with direct cost consequences. Whether those costs close is what the business model bets below are about.
Go deeper on the hidden human economy behind each ride.
Which business model bets will decide robotaxi scale?
The contest splits along two axes: whether to own the full stack and the rider relationship or aggregate third-party operators, and whether to buy the right to operate through demonstrated safety and permits or bet that manufacturing-cost leadership overwhelms friction. Each model assumes the moat forms in a different place. Weigh every bet against what it optimises for: control and trust versus breadth and cost. Hold off crowning a winner early.
Waymo and Tesla sit on the vertical side and Uber on the aggregator side. Uber’s multi-operator London network shows what breadth buys and what control it surrenders. Sensor strategy maps onto the same divide: lidar-and-HD-map richness against vision-only cost. Each bet also carries a regulatory and safety load, taken up in the next two sections.
Read the full vertical-versus-aggregator contest.
How do permits and liability gate robotaxi expansion?
Permits and liability now decide where a robotaxi service can grow. With no driver aboard, responsibility for harm shifts to the operator and manufacturer, which reshapes insurability and which deployments get authorised. The pathway differs by jurisdiction: self-certification in the US against pre-market approval in Europe. Controls-free designs like the Cybercab sit outside standards written for human controls and need a special approval path. Assess exposure before assuming expansion follows capability.
Spain’s first national Level 4 permits and London’s staged, safety-driver-first launch show Europe’s incremental path. California’s capped, staged expansion sits against Arizona’s permissive stance, and Australia’s state-led AV rules sit between the two.
See how permits and liability gate scale.
How should you read robotaxi safety claims and rider data risk?
Read an operator’s safety metrics as claims you interrogate. Check who reports each number, how it is defined, what it excludes and whether any independent party verifies it. Self-reported figures and external benchmarks like the IIHS per-mile study can tell very different stories. Rider data adds a second layer of exposure: know what the fleet records, where it flows and who can access it before you adopt a service.
The IIHS found Waymo’s rider-only vehicles were involved in 68% fewer police-reportable crashes than human drivers in the same cities, while Tesla’s own figures imply a crash roughly every 55,000 miles even with a safety monitor aboard. Disengagement rates have barely moved in a decade of testing, which says more about the metric than the technology and feeds the business model bets above.
Start reading safety and data claims critically.
Resource Hub: Robotaxi Scale Deep Dives
The Operational Shift
- What Changes When a Robotaxi Service Really Scales: The operating model behind a real service.
- The Hidden Human Economy Inside Robotaxi Per-Mile Costs: Where the per-mile cost floor sits.
The Strategic and Regulatory Contest
- The Business Model Bets That Will Decide Robotaxi Scale: Who owns the ride, and who owns the moat.
- How Permits, Exemptions and Liability Gate Robotaxi Expansion: What authorises or blocks growth.
The Evidence Layer
- How to Read Robotaxi Safety Claims and Rider Data Risk: How to verify the numbers.
Suggested reading order: Operational Shift first (what and why it changes), then the Strategic and Regulatory Contest (who wins and who decides), then the Evidence Layer (how to verify the claims).
Where to Start
The driving is settled for the incumbents; what happens around the car now decides who scales. Pick your doorway. If you are weighing what scale demands operationally, start with what actually changes when a service scales. If you are assessing an investment or a build-versus-partner decision for your team, start with the business model bets. If you are weighing risk, compliance or procurement, start with permits and liability. Each doorway leads into the full treatment.
Frequently Asked Questions
What is Level 4 autonomy, and how is it different from Level 5?
Level 4 means the vehicle can drive itself without a human in defined conditions, but it still has a boundary. That boundary is the operational design domain, so a Level 4 robotaxi may only work in mapped cities, in good weather or outside flood zones. Level 5 would carry no such limit. In practice, every commercial service running today is Level 4.
Why do robotaxis still have remote human operators if there is no driver?
Because removing the driver does not remove the human. Remote assistance staff monitor fleets, resolve stuck vehicles and confirm routing when the software cannot decide. One operator can watch many cars, so the cost is real but smaller than a driver’s wage. Uber’s London launch also used in-car safety drivers at first, a further human layer before full driverless approval.
How can I tell a real commercial robotaxi service from a pilot?
Judge it on paid, fully driverless rides, not permits or vehicle counts. A pilot may hold a permit and park cars at a depot without carrying the public, while a scaled service books real fares and runs without a safety driver. Check whether anyone can hail it, whether money changes hands and whether it operates beyond a single fixed route.
Is a robotaxi the same as a Tesla with Full Self-Driving?
No. A robotaxi is commercial, driverless and accountable to a regulator, while Full Self-Driving is an advanced driver-assistance feature that still expects a human to supervise. Tesla’s no-controls Cybercab is a separate product and drew an NHTSA probe days after deployment. Tesla’s claimed $0.20 per mile also hides the operating costs every operator faces.
How does a robotaxi actually get charged and cleaned between rides?
Off the car, at depots, and it is a real operating cost. Vehicles must return to charge, be inspected, cleaned and reset, and depot throughput becomes a bottleneck as fleets grow. This servicing layer is why removing the driver does not remove the labour around the car, and why per-mile costs do not fall as fast as hardware prices do.
What actually happens if a robotaxi is involved in a crash?
Because a robotaxi has no driver, legal responsibility for harm shifts to the operator and manufacturer, so claims run against the company rather than an individual. This makes insurance and authorisation gating factors, not afterthoughts. The vehicle’s sensors also record the event, and that footage typically becomes the central evidence in any dispute.
Why did Waymo’s cars stop in a blackout and a flood?
Because a robotaxi only works inside its operational design domain, the formal boundary of where and when it may operate, and blackouts and flooding sit outside that box. When conditions fall outside the mapped, validated envelope, the safe response is to pause rather than improvise. That is a limitation of the current system, not a sign the technology failed outright.
Will robotaxis come to Australia, and when?
Australia is likely to follow the state-led path rather than a single national rollout, with each state authorising and insuring fleets on its own terms. That patchwork sits between Europe’s pre-market approval and America’s permissive self-certification. Timing depends less on the technology than on how quickly each state writes and enforces its rules, so watch state regulators, not manufacturers’ launch dates.
What is the difference between a permit and an exemption for a robotaxi?
A permit is a regulator’s approval to operate under existing rules, while an exemption lets a vehicle run when it does not fit those rules at all. Controls-free designs such as Tesla’s Cybercab, which has no steering wheel or pedals, sit outside current standards and so need a specific exemption. Most deployments proceed by permit; the unusual designs need the harder exemption route.
Is it true robotaxis will put every driver out of work?
Not on current evidence. Robotaxi scale is creating new human roles even as it removes driving ones, including remote assistance, depot servicing, cleaning and maintenance. Lyft redeployed drivers to clean Waymo cars, a concrete example. The shift is a redistribution of labour around the car rather than its outright elimination, and it moves more slowly than capability headlines suggest.
How many robotaxis are actually operating on public roads today?
Reliable totals are hard to pin down because operators quote fleet sizes, permits and paid rides differently. Waymo’s 14 US metros and more than 20 million paid rides show real scale, but that is one operator. Globally most services remain small: Uber’s London launch began with about 15 vehicles, and Spain’s first national permits covered a 20-vehicle Madrid fleet.
