Rogue robotaxis: who pays for the driverless car that misbehaves

In the spring of 2026 Waymo robotaxis drove through the barriers of freeway construction zones in Phoenix and San Francisco at least thirteen times, past "road closed" signs a human would have heeded. In June the Alphabet-owned company recalled 3,871 of its cars, told them to stay off the highways of four American cities and went back to the software. These are software recalls, resolved over the air; their bite is the restriction, not the tow truck. Set them end to end and an escalation is visible: 444 cars recalled in 2024 for one sort of confusion, 1,212 in 2025 for a second, 3,067 for a third, and then two batches of almost four thousand this spring for a fourth and a fifth, flooded roads and closed lanes. In one of those incidents an unoccupied Waymo drove into a flooded San Antonio street and was swept into a creek. This is the age of the rogue robotaxi, and it has opened at the very moment the industry chose to declare the driving problem solved.

The declaration carries weight. Waymo now gives roughly half a million paid rides a week, ten times what it managed two years ago, and says it will pass a million before the year is out, in more than twenty American cities, with London and Tokyo to come. Tesla switched on unsupervised service across Austin in June and will unveil its purpose-built Cybercab there on September 3rd. Baidu, in China, moved 3.2m driverless rides in a single quarter. The technology question, as the industry likes to put it, is close to settled. The money question is not; and the safety record has become the ground on which it will be decided.

The economics are simple in outline and brutal in execution. Analysts at Morgan Stanley estimate that a Waymo mile costs about $1.36 to deliver, against $0.74 for a Tesla robotaxi, depreciation included. Both already sit below the $1.50–$2.50 a mile that a human-driven Uber costs, and the prize is the race downward to the point where the economics close on their own; Tesla's most enthusiastic projections imagine a Cybercab mile at a quarter or half a dollar. But a cheap mile is worth owning only if the vehicle is allowed to drive it. When Waymo pulled its cars off freeways it surrendered the industry's most valuable trips, the long airport hauls, until the fix was written. The recall was not merely embarrassing. It demonstrated that operational permission, not software, is now the scarce resource: a highway mile is worthless if a regulator has taken the highway away.

The regulator is already restive. In July the head of America's road-safety agency wrote of a "clear pattern" of driverless cars interfering with law enforcement and other first responders. Waymo's record is the best in the field — it claims thirteen times fewer serious injuries per mile than a human driver, across 170m autonomous miles — yet it has recalled cars for construction zones, flooded roads, stationary barriers and school-bus violations, and it is under investigation for striking a child near a Santa Monica school in January. Tesla faces a federal probe after a woman was killed by an automated driving system in June, and it has disclosed robotaxi crashes in which a remote operator, not the software, was at the wheel. Scale, argues Phil Koopman of Carnegie Mellon, makes rare failures impossible to ignore; on floods and construction zones, he concludes, the companies "haven't been able to fix them". Passengers may forgive the odd inexplicable detour. Insurers, juries and state regulators are less forgiving.

That is why the two public companies that own this future differ in kind rather than degree. Alphabet's Waymo has bought safety into every layer — lidar and dense mapping, streets mapped and approved in advance, a recall-and-restrict culture, and a co-chief executive, Tekedra Mawakana, who openly asks Washington for national safety standards that would lock caution in as regulation. The fee shows up in the cost per mile. Tesla has chosen the opposite: a cheaper vision-only system on cars it already builds, expansion at near-zero fleet cost, and humans sitting at consoles ready to seize the wheel from afar. Its cost advantage is partly genuine; it is also partly safety deferred. Deferred safety has a known invoice, because GM's Cruise wrote it: one 2023 incident in which a robotaxi struck and dragged a pedestrian brought federal investigations, a suspended operating permit and, in the end, the unit's shutdown. Even Uber, the incumbent with the most to lose, has committed $10bn to robotaxi partnerships rather than back a single winner.

The market has begun to price the difference. Tesla fell about 30% in 2026 — its worst stretch since 2022 — to a value of roughly $1.2 trillion in late July before recovering to about $349 a share, still about a third below a 52-week peak near $500. The retreat contains an irony visible in Tesla's own accounts. The company's headline proof of robotaxi momentum is a cumulative-mileage chart: about 2.4m paid miles since launch, with the second quarter adding roughly the same 0.9m as the first, even as the service map grew. The story is expanding; the odometer is flat. Profit, meanwhile, came in at $0.33 a share in the second quarter against analyst expectations of $0.53, a year after the company booked its first annual revenue decline and conceded the global EV crown to BYD. The multiple is paying for autonomy on schedule, which is why the chief executive's promise of a million robotaxis by the end of 2026 is doing so much work.

Alphabet's arithmetic runs the other way. The company entered the $4 trillion club in January and trades at about $347, so Waymo's $16bn fundraise in February — a valuation of $126bn is around 3% of the parent. That valuation is no longer fiction: Waymo's volume has grown tenfold in two years and still compounds, so Alphabet's shareholders own a real and expanding robotaxi business without staking the portfolio on a single failure mode; a Cruise-scale catastrophe inside Waymo would cost the parent a rounding error. But it also means Waymo barely moves the stock. Alphabet is valued for search, for cloud and for an AI-building programme expected to cost $180bn–$190bn this year alone — more than the entire sticker price of Waymo. Inside Google, the driverless option is a hedge. A spin-off, which investors periodically moot, would turn that hedge into a bet; doing so is, for now, Alphabet's call to make.

None of this resolves the argument in favour of caution over cheapness. Maturity in this industry, the past two years suggest, is measured in recalls — in how quickly and cheaply a company can be wrong in the field and put itself right. Waymo's higher cost per mile is partly the price of lessons already paid for in engineering and restriction; Tesla's lower cost is partly the price of lessons not yet billed. The investor's question, then, is where to watch. Not the next launch event but the safety ledger: the accident rate per mile of the unsupervised fleets, the premiums insurers charge them, the behaviour of a fully driverless car in its first serious incident. The company that can be wrong cheaply — careful enough to survive its mistakes, cheap enough to win the mile — is where the money question is likely to settle. The company whose entire price depends on never being wrong is holding the tail.