Who Needs a Self-Driving Car? Where Autonomy Pays, and Who Collects
No steed runs far that has no stable, nor any wain that lacks a wright to mend it; and a wain that drives itself must still be watched from some high window, by eyes that never sleep, along the roads it does not know.
Why We Asked
Most arguments about self-driving cars start from the wrong question. “Will robotaxis take over?” invites hype; “do we even need them?” invites a shrug. Neither tells you where the technology creates enough value to justify the sensors, software, rules and risks it drags in.
So we ran it the way we would run an investment case: map the evidence, build a model, and let the numbers argue back. The evidence map spans 15 domains, from safety and insurance to energy, labour and geopolitics, and cites 121 sources: regulator data, company filings, peer-reviewed studies and industry reports, current to 26 September 2026. We tested twelve working hypotheses against it and logged fourteen places where credible sources contradict each other.
A second pass went after the numbers that were still loose: Waymo’s California trip filings, state vehicle registrations, Alphabet’s segment results, federal crash reports, London’s taxi fare card and US job counts. The model is a Python script with every input tagged as sourced, derived or assumed. It prices a robotaxi mile, finds break-even utilisation, compares paying per ride with owning a car, and stress-tests traffic, energy and safety statistics. Its outputs are scenarios, not forecasts.
Throughout, the claims that carry the argument wear a label:
| Label | What it means |
|---|---|
| Known | Measured by a regulator or an independent study |
| Evidence suggests | Peer-reviewed or consistent across sources, but narrow |
| Company says | Self-reported and not independently checked |
| Our model | An output of our cost model: a scenario, not a forecast |
| Uncertain | Sources disagree, or the data do not exist yet |
The Short Answers
Eight questions shaped this piece. Here is where each one landed; the sections that follow show the evidence, and Where We Land pulls it together.
| # | Question | Short answer |
|---|---|---|
| 1 | Where does autonomy actually work today? | Where one operator controls the domain: mines, private roads, freight lanes and fenced-off city grids. No car on sale drives itself everywhere. |
| 2 | Is a robotaxi cheaper than a taxi? | Yes, before research costs: about $1.58 per paid mile against $2-plus for ride-hail. Waymo’s losses most likely sit in R&D and expansion. |
| 3 | Are they actually safer? | Where they run, yes: 68% fewer police-reportable crashes in independent data. But that evidence covers one operator. |
| 4 | What does a robotaxi need that your car doesn’t? | A whole operating system of remote assistants, depots, maps and permits. Its worst failures so far were fleet-wide stalls, not crashes. |
| 5 | Who pays when the machine crashes? | Increasingly the software’s maker, by four different legal routes. Crashes fall, but the bad-year tail gets fatter. |
| 6 | Will robotaxis ease traffic or add to it? | Add, unless pooled and priced: +58% vehicle-miles for solo robotaxis at scale, −7% for pooled, road-priced fleets. |
| 7 | Who’s winning, and is it even one race? | Four regional races, three business models, and a Gulf market where Chinese stacks scale. No single winner yet. |
| 8 | Will it replace the car in your driveway? | Not at today’s prices: only below about 3,800 miles a year at $2 a mile. The private case is time, not money. |
Where Does Autonomy Actually Work Today?
Wherever one operator controls the whole domain.
Mines, private roads, fixed freight lanes and fenced-off city grids already run without drivers. Nowhere can you buy a car that drives itself everywhere.
The bottom of the ladder is old. Komatsu has run driverless haul trucks since 2008 and commissioned its 1,000th ultra-class autonomous truck in April 2026 (Komatsu). One Caterpillar quarry hauled more than 2M tons in its first autonomous year (Caterpillar). On private oilfield roads in the Permian Basin, Atlas runs 28 driverless Kodiak trucks that have hauled about 7,000 loads of sand (FreightWaves).
Public roads come next, but only in pieces one company can manage. Aurora runs driverless trucks on Texas lanes, with no observer aboard since late July, and had logged nearly 440k driverless miles by 30 June (FreightWaves). Waymo has about 4,000 robotaxis, 80% of them in California and Texas, and carries about 500k paid rides a week (TechCrunch). Baidu’s Apollo Go runs in 28 cities (Baidu); Zoox, holder of the first commercial federal exemption for a purpose-built robotaxi (TechCrunch), has charged fares in Las Vegas since 10 August (The Charge Port); Tesla sells unsupervised rides in six metros (Electrek). Company says for most of those counts.
Everywhere else a human is still in charge or on call. London’s first “autonomous” ride-hailing, launched by Wayve and Uber on 3 September, carries a licensed driver (Electrek), and no operator has completed the first stage of a driverless permit in London (TNW). China’s first L3 approvals let two models drive themselves on set routes at up to 50 and 80 km/h (CarNewsChina). GM’s eyes-off Escalade IQ is a 2028 target (TechCrunch).
The same software can be safe and profitable in one place and unusable in another. That is why every driverless business so far is a fleet with a fence around it, and why the useful unit of analysis is the operating domain, not the vehicle.
A mapless, end-to-end system, Tesla’s or Wayve’s, running driverless in a new city within weeks and publishing crash data per mile. The ladder would flatten.
Is a Robotaxi Cheaper Than a Taxi?
Yes, before research costs.
Our model puts a US robotaxi at about $1.58 per paid mile: below the $2-plus that ride-hail charges, above the $0.88 a mile it costs to own and run a car driven 12,000 miles a year. Empty miles, not hardware, decide the margin, and Waymo’s losses most likely sit in research and expansion rather than in the ride.
Start with where the money goes Our model. With about $60k of vehicle and sensors, 69.5k miles a year and 45% of those miles empty, a robotaxi costs $0.87 per vehicle-mile: $0.21 of capital, $0.19 of fixed operations such as remote assistance and depots, and $0.47 of running costs, from electricity and maintenance to insurance and field crews. Because nearly half its miles carry nobody, each paid mile costs 1.8 times as much: $1.58.
The biggest lever is not the car. Moving the empty share between 25% and 50% swings cost from $1.16 to $1.74 per paid mile; moving the vehicle price between $45k and $90k swings it only from $1.48 to $1.77. A Monte Carlo across all 13 inputs gives $1.25–2.22.
How busy does a robotaxi need to be? That depends on the fare:
| Price per paid mile | Miles a year to break even | Rides per car a day |
|---|---|---|
| $6.00 (San Francisco today) | 9,100 | 4 |
| $3.00 | 21,900 | 10 |
| $2.00 (US ride-hail average) | 41,000 | 19 |
| $1.50 | 85,500 | 39 |
| $1.20 or less | never | — |
Two things follow. Below about $1.25 a mile the model never breaks even, because running costs and vehicle wear, which both grow with every mile, already exceed the fare. And at today’s prices the ride itself pays: Waymo’s average Bay Area fare of $19.69 in late 2025 (Obi) works out near $6 per paid mile on California trip lengths Our model.
Utilisation has started to slip, though. Waymo’s fleet grew to about 4,000 cars by September while rides held near 500k a week (TechCrunch). That is about 18 rides per car a day, down from about 23 in March. It nudges our base cost to $1.61, and utilisation is still well above break-even at current fares; Alphabet’s third-quarter results will show whether demand is catching up with the fleet Uncertain.
So where do the losses come from? Most likely research, mapping and expansion, though Alphabet does not break out Waymo’s costs. Its Other Bets segment, where Waymo sits alongside smaller units, lost $2.1B in the first quarter of 2026 and $1.8B in the second (Q1 release, Q2 call). Spread an assumed $4B a year of research over 3,000 cars and it adds $19 a mile; over 500,000 cars, $0.11 Our model. Carrying it at $0.10 a mile would take about 570k robotaxis, roughly 1.2% of all US vehicle-miles. Scale, not the ride, is the business problem.
Removing the driver is the business case, and it is also someone’s pay packet. The US has 455,700 taxi, shuttle and chauffeur jobs, and 86% of taxi drivers are self-employed (BLS); it also has 2.22M heavy-truck driving jobs, at a median $58,640 (BLS). China had 7.48M licensed ride-hail drivers in 2024 (Beijing Scroll).
In trucking, drivers are 44% of the $2.34 it costs to run a truck for a mile (ATRI). Aurora plans a driver subscription of at least $0.85 a mile (FreightWaves). Add trucks running about 200k miles a year, our estimate from Aurora’s revenue target, instead of ATRI’s 86k, and the cost falls 11–17% Our model: useful, not revolutionary. Goldman puts the all-in cost of an autonomous truck in 2025 at $8.56 a mile (Goldman Sachs), so today’s freight prices are entry prices.
Operators publishing cost per mile, which none does, or rides per car climbing back above 20 a day as fleets grow. Either would tell us whether the cost curve is real or only modelled.
Are They Actually Safer?
Where they run, yes, by a wide margin, for one operator.
Independent data put Waymo’s cars in far fewer police-reportable crashes than human drivers. The most striking claims rest on a handful of events, and no one else publishes comparable data.
Known The Insurance Institute for Highway Safety compared about 50M of Waymo’s driverless miles with human driving from 2021 to 2024 and found 68% fewer police-reportable crashes (IIHS). In Austin, on few miles, it found no advantage (+4%).
Company says Using peer-reviewed methods, Waymo reports 82% fewer injury crashes and 95% fewer serious-injury crashes over 271.3M rider-only miles to June 2026 (Waymo). A peer-reviewed study by Waymo and the reinsurer Swiss Re found 88% fewer property-damage claims and 92% fewer bodily-injury claims than for the overall driving population, over 25.3M miles (Waymo and Swiss Re).
The whiskers are the point. Waymo’s 95% cut in serious injuries rests on about three crashes against about 58 expected, so the true reduction could sit anywhere from 85% to 99% Our model. Its injury-crash figure, built on about 185 events, is far tighter at 79–84%. Swiss Re’s bodily-injury result rests on two claims.
Fatalities are harder still. To show with 95% confidence that a fleet beats the 2025 US rate of 1.10 deaths per 100M miles (NHTSA), you need about 272M miles with no fatal crash Our model. Waymo’s 271.3M miles are about that distance, but its data count every crash its cars were involved in. That includes at least one fatal crash it did not cause: a speeding driver’s six-car collision in San Francisco in January 2025, which hit an empty Waymo (NBC Bay Area). The data also cover only five metros, so “safer everywhere” needs the test repeated in every new domain.
Then there is what doesn’t compare. Tesla’s Austin robotaxis reported 14 crashes by mid-January, all with safety monitors aboard and narratives redacted, in what Electrek estimates at about 800k miles (Electrek); Tesla later released 19 unredacted narratives, which mostly put other drivers at fault (Not a Tesla App). In May a Tesla robotaxi crashed in Houston while a remote operator was driving, the third such report (NHTSA 13781-15395, Electrek). Yet in February Tesla’s vice-president of vehicle engineering had told the Senate Commerce Committee, answering questions on cybersecurity, that its driving controls “cannot be accessed from outside the vehicle” (Electrek).
Apollo Go reports one airbag deployment per 14.4M km, on its own definition (Baidu). Aurora says no collision has been attributed to its driver (Company says FreightWaves); with zero events in 440k driverless miles, that only bounds its rate below about 7 per million miles at 95% confidence, too few miles to tell it apart from human truckers Our model.
A second operator publishing rider-only miles and crashes in a form IIHS could compare, or a first at-fault fatality in a driverless fleet.
What Does a Robotaxi Need That Your Car Doesn’t?
An operating system around it.
Maps, remote assistants, depots, field crews, permits, and a plan for when the city’s power fails. The failures that have hurt most so far were not crashes but whole fleets stopping at once.
Two nights show why. On 20 December 2025 a blackout knocked out San Francisco’s traffic signals. Waymo says its cars crossed more than 7,000 dark signals, but a spike in requests for remote confirmation stalled vehicles and the service paused (Taipei Times); Waymo then updated its fleet’s software (CNBC). On 31 March 2026 about 100 Apollo Go cars stopped on Wuhan’s roads for around two hours (CarExpert), and China froze new robotaxi permits nationally until an end-June review (CnEVPost, electrive).
Both failed safe in the narrow sense: cars stopped rather than crashed. But a stopped fleet blocks junctions, buses and ambulances. Centralisation swaps collision risk for availability risk, and it makes the operator’s control room part of the city’s critical infrastructure.
The human layer is bigger than the name suggests. Waymo told a Senate hearing in February that about 70 remote-assistance agents are on duty worldwide at any time for some 3,000 cars, roughly one per 43, with about half of them in the Philippines (Axios). Goldman’s April report puts the ratio at six vehicles per remote operator (Goldman Sachs); the gap is most likely seats on duty versus all staff across shifts Uncertain. No operator discloses how often its cars ask for help (Markey report).
Should cities build smarter roads instead? China is trying. Its vehicle-road-cloud programme spans 20 pilot cities, and Shanghai alone has 400+ smart intersections (7ITS). But roadside kit costs an estimated CNY1–2M per km (Yicai). Fitting all 5.4M km of China’s roads would cost CNY5.4–10.8T, about $0.76–1.5T, against about $20B to put a $20k autonomy kit on a million robotaxis Our model. Pilots already report roadside units with few vehicles to talk to (36Kr).
Instrumented roads pay in bounded places: ports, mines, depots, dense junctions. Elsewhere the cheap public basics help human and machine drivers alike: clear markings, backup power for signals, and live roadworks data.
Incident data that attribute failures by layer (vehicle, cloud, operations, road) across operators. Until then, the system layer is where the risk hides.
Who Pays When the Machine Crashes?
Increasingly the company behind the software, though each region gets there by a different road.
Crashes get rarer, but the risk concentrates on a few balance sheets exposed to defects that can hit every car at once.
| Region | Who pays first | Legal basis | Where it stands |
|---|---|---|---|
| US | Whoever a court finds liable: driver, operator or maker | Tort and product liability, state by state; no federal AV statute | A $243M verdict against Tesla over a fatal Autopilot crash was upheld in February 2026 (CNBC); the SELF DRIVE Act cleared committee 12–11 (Fox News) |
| UK | The insurer | Automated and Electric Vehicles Act 2018; the Automated Vehicles Act 2024 adds an authorised self-driving entity and shields the user-in-charge | Passenger permits since 15 May 2026; full regime due in the second half of 2027 (Electrek) |
| EU | The producer, strictly, now including for software | Product Liability Directive, applying from 9 December 2026 (Gibson Dunn) | Traffic law and fault rules stay national |
| China | The maker, in principle, while L3 is engaged | Pilot rules for approved L3 models and routes (CarNewsChina) | National law still incomplete |
Insurers see the upside first. Swiss Re’s study with Waymo found 88% fewer property-damage claims and 92% fewer bodily-injury claims (Waymo and Swiss Re), and Lemonade cuts its per-mile rate by about half when Tesla’s FSD is engaged (Lemonade). Company says in both cases, and nobody publishes claim severity. A robotaxi crash may be rarer and dearer: as the Tesla verdict suggests, juries can aim at the deepest pocket, even over a privately owned car.
Business models move the risk around too. Aurora carries the insurance on trucks it runs under its full-service model; its planned subscription implies carriers take on more of that cover (FreightWaves).
The harder problem is correlation. Human errors are independent: one tired driver does not make another tired. A fleet shares one software stack, so one defect can reach every car at once. We simulated a 10,000-car fleet driving 700M miles a year Our model:
| 10,000 cars, 700M miles a year | Expected injury crashes | Bad year (1 in 200) | Bad year ÷ expected |
|---|---|---|---|
| Human drivers | 2,646 | 2,779 | 1.05 |
| Robotaxis, independent errors | 476 | 533 | 1.12 |
| Robotaxis plus a rare fleet-wide defect | 486 | 998 | 2.05 |
Expected crashes fall 82%, but the bad year relative to the average roughly doubles. Insurers hold capital against the bad year, so capital per unit of premium can rise even as claims fall. The defect settings are our assumptions, not estimates.
So far the fleet-wide failures have been stalls, not crashes, and fixes arrive as software pushed to every car at once, as after the San Francisco blackout (CBT News). The channel that fixes a whole fleet overnight can also break it overnight.
Insurers publishing claim severity for driverless fleets, or reinsurers pricing cover for fleet-wide defects. Either would show whether the tail is as fat as the structure suggests.
Will Robotaxis Ease Traffic or Add to It?
Add to it, unless they are pooled and priced.
Cheap, comfortable rides invite more trips, and today’s robotaxis drive nearly half their miles empty. Only shared, road-priced fleets cut vehicle-miles in our scenarios.
Known Start with the empty miles. Waymo’s California filings show 86.3M vehicle-miles over 13.8M trips, with passengers aboard for 53.6% of those miles (Abdelhalim 2026). Empty miles per trip fell from 5.1 to 2.8 as the service matured, yet cars still spend about 18 minutes empty between trips (Streetsblog). Trips average 1.4 passengers.
Two assumptions drive the result. The first is induced demand: when a Sacramento experiment gave 43 households a free chauffeur, their vehicle-miles rose 60%, over half of it empty, and their transit trips fell 70% (UC ITS). The second is transit substitution: ride-hailing pushed the growth in San Francisco’s traffic delay to 62%, against 22% without it (Erhardt 2019), and cut bus ridership by about 10% (Erhardt 2021). Our scenarios assume far smaller effects, and the direction holds regardless.
The occupancy arithmetic is unforgiving. At today’s 45% empty share, a fleet needs 2.65 passengers on every occupied trip just to match private cars; at 20% empty it still needs 1.82 Our model. Robotaxis average 1.4.
Energy follows the same logic. If the autonomy stack draws 0.3–1.5 kW, as we assume for current designs, it adds 3–42% to an electric car’s energy per mile, most in slow city traffic Our model. Industry estimates of 1.5–3 kW for 2021-era stacks would mean 17–83% (Design News). Per passenger-mile, a solo robotaxi uses about 2.1 times the energy of a private EV, and a pooled one about 1.2 times. Fleets also charge in bulk: 10,000 robotaxis draw about 28 MW on average, or 83 MW if charged in an eight-hour window, like a mid-sized data centre Our model.
A city where robotaxis carry more than 5% of trips publishing traffic and transit ridership before and after. No city has that data yet, and the choices that decide between +58% and −7% (pooling, road pricing, transit integration) are still open.
Who’s Winning, and Is It Even One Race?
It is four races run under different rules.
The US scales private operators under thin federal law. China scales state-supervised fleets and can halt them within weeks. The UK wrote the clearest liability law but has no driverless service yet, and the EU harmonises approval faster than it deploys. Meanwhile the Gulf has become neutral ground where Chinese stacks scale.
| US | China | UK | EU | |
|---|---|---|---|---|
| Driverless paid service | Yes: Waymo, Tesla, Zoox | Yes: Apollo Go, Pony, WeRide; new permits frozen Apr–Jun 2026 (CnEVPost) | No: supervised rides since Sep 2026 (TNW) | No: supervised pilots such as Hamburg’s |
| Rulebook | No AV statute; federal exemptions capped at 2,500 vehicles a year (TechCrunch) | National L2 standard from Jan 2027, automated-driving standard from Jul 2027 (China Daily) | AV Act 2024; passenger permits since May 2026 | Small-series approval, capped at 1,500 per type a year (InterRegs) |
| Value of removing the driver | High: ride-hail above $2 a mile (Goldman Sachs) | Low: about $0.47 a mile on a Shenzhen example (Wire China) | High: a six-mile black-cab ride costs £33–45 (TfL) | High, varies by country |
| Published safety evidence | Waymo’s data, an independent IIHS study, federal crash reports | Company claims only | None yet | None at scale |
| Edge | Capital, AI labs and chips | Cheap EVs, batteries and state backing | Liability law and AI talent | Carmakers and single-market approval |
Inside each race, three business models compete. Vertical operators run the software and the service and specify the vehicle: Waymo, Tesla, Zoox. Licensors sell the driver: Wayve to Nissan, Mobileye to Volkswagen, and Nvidia’s Hyperion platform to BYD, Geely, Stellantis and others (Nvidia). Aggregators sell the ride: Uber works with more than ten autonomy partners and has signed for 5,000+ Level 4 vehicles from Stellantis from 2028 (electrive).
The research arithmetic favours whoever spreads fixed cost widest. At $4B a year, research adds $19 a mile across 3,000 cars and $0.11 across 500,000 Our model. A vertical operator has to reach that scale alone; a licensor borrows other people’s fleets. That is the case for an “Android of autonomy”. The analogy skips one thing: when a phone crashes you reboot it. When a car crashes someone may be hurt, and whoever owns the whole safety case, liability included, may keep an edge a licensor cannot.
Geopolitics already splits the map. From model year 2027 the US bars driving software linked to China or Russia, and robotaxi services that run it (BIS). China now treats camera imagery of the outside world, radar data, trajectories, driving maps and training data as “important data” that needs a security approval before it leaves the country (DLA Piper). So Chinese operators grow where both sides are welcome: WeRide runs about 400 robotaxis in the Middle East (WeRide), Apollo Go carries paying riders in Dubai, and Pony operates in Doha and Singapore. Even Waymo’s new Ojai rides on a base vehicle from Geely’s Zeekr (TNW).
A licensed stack running driverless, at scale, for a fleet its developer does not own. That would settle whether autonomy becomes a component or stays a vertically integrated service.
Will It Replace the Car in Your Driveway?
Not at today’s prices.
A robotaxi beats owning a car only if you drive very little: under about 3,800 miles a year at $2 a mile. The private case for autonomy rests on time, not money, and nearly every car you can buy today still needs you watching.
The crossover is simple arithmetic Our model. Owning a new car costs about $6,240 a year before you drive it, plus about $0.36 a mile (AAA 2025, our split). Paying per ride wins only when your annual miles are low enough to beat that fixed cost. At $2 a mile the line sits near 3,800 miles a year, or about 5,000 if parking costs you $2,000. Only at $1 a mile does it reach 9,700–12,800 miles, close to a typical car’s use (13,100 against owning a self-driving car).
Time changes the sum. An hour not spent driving is worth something: we put it at $0.11–0.60 a mile, or $1,350–7,200 a year at 12,000 miles, depending on the value of your time and your average speed Our model. That value goes to anyone who can stop watching the road, owner or rider. It is the real private case for autonomy, and it is why carmakers want eyes-off driving: GM targets it on the 2028 Escalade IQ (TechCrunch), and Toyota is exploring personal vehicles with Waymo (Toyota).
Meanwhile the supervised middle is growing fast, and the evidence there is thin. In China, 70.5% of new cars had L2 features in the first half of 2026 (China Daily). Evidence suggests Partial automation brings no crash reduction (IIHS), while 53% of Super Cruise users and 42% of Autopilot users said they were comfortable treating their systems as self-driving (IIHS survey). A $243M verdict against Tesla over a fatal Autopilot crash was upheld in February (CNBC).
Yet the Dutch regulator approved Tesla’s FSD Supervised in April (Electrek), and Lemonade halves its per-mile rate when FSD is engaged (Lemonade). China will require driver monitoring for L2 systems from January 2027 (Gasgoo). Uncertain is the honest label: no regulator yet collects miles driven with L2 engaged.
Robotaxi prices near $1 a mile in a big market, or exposure data that separate L2-engaged miles from manual ones. The first would put private ownership in play; the second would tell us whether the supervised middle is safer or merely more comfortable.
Where We Land
So, do we need self-driving vehicles? We need autonomy where its value clearly exceeds the system it demands. That is a narrower set of places than the hype suggests, and a wider one than the backlash admits.
The thread through all seven: autonomy’s value is set less by how well the car drives than by how the network around it is run. The same technology can mean 58% more traffic or 7% less, a cheaper taxi or a pricier car, a safer city or a more fragile one. What decides it is who operates the fleet, how rides are priced and whether they are pooled.
Three Assumptions the Evidence Corrects
- “Fleets beat private cars.” True today, but because a fleet replaces a paid driver, not because a fleet mile is cheaper. It isn’t: $1.58 against $0.88.
- “Utilisation is the prize.” It depends on the baseline. Against a taxi, removing the driver dominates, because multi-shift taxis already run long hours. Against a private car there is no wage to save, and autonomy overhead plus 45% empty miles eat the utilisation gain.
- “Autonomy needs smarter roads.” The system that binds is the operator’s, not the road’s. The failures so far began in operators’ systems and the city’s power supply, not in missing roadside kit.
- Robotaxi fares near $1 a mile with positive unit margins. Giving up the car would come into play for typical drivers.
- A second operator, ideally an end-to-end one, publishing driverless crash data IIHS could compare. The safety case would generalise beyond one company.
- Waymo’s rides per car climbing back above 20 a day as its fleet grows. Demand, not supply, would be setting the pace.
- A city where robotaxis pass 5% of trips publishing traffic and transit data. The traffic question would finally get an answer.
- A fleet-wide defect that causes collisions rather than stalls. Insurance, permits and public trust would reprice overnight.
Autonomy already makes its case where one operator controls the domain and the driver is the costliest input: mines and yards since 2008, freight lanes in principle, and dense, high-fare cities where a robotaxi mile at about $1.58 undercuts a $2-plus fare. It does not yet beat the car in the driveway, its strongest safety evidence covers one company in five metros, and its worst failures so far have been fleets stopping at once rather than crashing.
The car can already drive. The open question is who runs the road around it. The same technology produces 58% more traffic or 7% less, a cheaper taxi or a pricier car, a safer city or a more fragile one, depending on who operates the fleet, how rides are priced and whether they are pooled. That is a policy choice as much as an engineering one, and it is being made now, city by city and permit by permit.
They asked whether the wain could find its own way, and it could. But none asked who would keep the roads, or who would answer when all the wains went wrong together; and that, in the end, was the only question worth the asking.
Leave a Reply