The driver is still in the system

Robotaxis are carrying paying passengers in 2026, but human supervision still sits inside the operating model. Depending on the company and city, people remain in vehicles, chase cars or remote-assistance centres, while permits, mapped operating areas, weather and liability constrain deployment. The market is commercially real, but it is not yet broadly autonomous.

On 22 January 2026, Elon Musk posted that Tesla had started running robotaxis in Austin with no safety monitor in the car. The post was one sentence long. The qualifications took months.

Tesla's AI lead, Ashok Elluswamy, said the company was starting with a few vehicles. Electrek reported that the monitors had not so much been removed as relocated, into a chase car trailing the driverless ones. By 3 March, Sherwood News found that the vast majority of Austin robotaxi rides still carried a supervisor in the passenger seat.

The seat was empty. The supervision was not.

This is the autonomous vehicle market in 2026. It works. It resists the sentence everyone wants to write about it.

The empty seat that wasn't empty

Tesla's Austin milestone carries weight because the company has spent a decade insisting that autonomy would fall out of the same camera-heavy consumer fleet it already sells. It has also spent that decade missing its own dates. In October 2019, Musk told investors that Tesla would have over a million robotaxis on the road the following year, woken up by an over-the-air update.

Six years later, the fleet running without an in-car monitor in Austin numbered a few cars, shadowed by a human in a second vehicle.

California is the sharper tell. Tesla holds a state DMV permit to test autonomous vehicles with a driver. The California Public Utilities Commission, which controls whether a company may carry public passengers in an autonomous vehicle, has confirmed that Tesla never applied for that authorisation. The most valuable carmaker on earth wants to run ride-hailing in its home state and has not filed the paperwork to do it. Tesla shares rose on the Austin news anyway.

What Tesla actually sells

The consumer product is called Full Self-Driving (Supervised). The parenthetical does the legal work. Tesla's own support pages tell owners to stay attentive and state plainly that the features do not make the vehicle autonomous — a sentence sitting directly beneath a brand name that says otherwise.

It sells. On the Q1 2026 earnings call in April, Tesla put paid FSD customers at roughly 1.3 million globally, with the growth arriving through subscriptions rather than one-off purchases. The car has become a delivery mechanism for the software licence.

China took the naming question less gently. After a fatal crash involving a Xiaomi SU7 in Anhui province, regulators in April 2025 banned carmakers from advertising driver-assistance features with the words "smart driving" or "autonomous driving" at all. Tesla's supervised system, cleared for Chinese roads after years of delay, is treated there as Level 2 automation. Local rivals — XPeng, Xiaomi, Baidu, Pony.ai, WeRide — never waited for it. The name is grand. The legal category is not.

Waymo's fleet got big enough to recall

Waymo is now the company most people mean when they say robotaxi. It runs about 500,000 paid rides a week across roughly ten U.S. metropolitan areas, with 3,871 vehicles in commercial service as of June 2026, and a stated target of a million weekly rides by year end. In parts of San Francisco and Phoenix, the cars have achieved the highest status available to a new technology: they are boring.

Its newest vehicle shows where the economics went. The Waymo Ojai is a pale-blue van built on a Zeekr platform from China's Geely, fitted out with 13 cameras, six radar units and four lidar sensors. It is cheaper than the Jaguars it supplements. It still has a steering wheel.

Then the arithmetic of maturity arrived. On 18 June, Waymo recalled nearly 3,900 robotaxis in the United States to patch software that had allowed cars to drive into closed freeway construction zones. The company restricted freeway operations while it shipped the remedy. A separate NHTSA investigation remains open.

The failures no longer look like science fiction. They look like municipal paperwork. When a power outage hit San Francisco at the end of 2025, Waymo vehicles stalled and snarled traffic, and the resulting Reuters coverage asked a question no demo video anticipates: what is a robotaxi during a blackout? A car that drives itself is also a car that must be recovered, rerouted and accounted for when a city breaks.

Redundancy versus scale

The engineering split between the two leaders is also a commercial one. Waymo has argued that a machine driver needs redundancy — lidar, radar, cameras, custom silicon, audio receivers tuned for sirens, prior maps, and a slow crawl outward from each validated operating area. Tesla has argued that vision plus fleet scale is cheaper and generalises faster.

Waymo is building a driver it can drop into any vehicle and any city. Tesla is trying to convert millions of privately owned cars into an autonomy network. Both bets route through the same bottleneck: the compute budget available on the vehicle, and the data centres behind it.

A hundred cars, frozen, for two hours

China supplies the volume. Baidu's Apollo Go delivered 3.2 million fully driverless rides in the first quarter of 2026, growth of more than 120% year on year, peaking above 350,000 rides a week in March. Cumulative public rides passed 22 million by April. By May the service reached 27 cities, with expansion into the Middle East.

It also supplies the cautionary tale.

Late on 31 March, police in Wuhan began receiving reports that Apollo Go cars had stopped in the middle of the road and could not move. At least 100 vehicles were affected, according to a traffic officer quoted by the Shanghai outlet The Paper. Doors could be opened, but some passengers were reluctant to step into moving traffic and called the police instead. Local media reported that some sat inside for close to two hours. Police attributed the outage to a "system failure" and recorded no injuries.

One software fault. A hundred cars. A city's arterial roads, at once.

The precedents were already on file. An Apollo Go car carrying a passenger drove into a construction pit in Chongqing in August 2025. A Pony.ai robotaxi caught fire on a Beijing road that May. Scale does not eliminate edge cases. It industrialises them, and it converts them into public records that regulators can count.

The middle level no one can sell in a slogan

The SAE taxonomy still describes the market more honestly than any marketing department does. Levels 0 through 2 are driver support: the human is driving, whatever the steering wheel is doing. Level 3 is conditional automation — the system drives inside a defined operational design domain, the specific roads and conditions it is approved for, and the person must be ready to retake control when it asks. Level 4 removes that handoff by fencing the geography. Level 5 goes anywhere a human can.

Level 5 is not late. It is absent.

Mercedes-Benz Drive Pilot remains the rare Level 3 system certified for American consumers, in Nevada and California. The conditions are exacting: approved freeway segments, clear lane markings, moderate to heavy traffic, speeds under 40 mph, daytime, clear weather, no construction zone, and a driver whose face stays visible to the cabin camera. The system lets the driver look away. It does not let the driver leave.

The chip behind the promise

Tesla's original argument for autonomy was data. The newer argument is silicon. On the Q1 2026 call, Musk confirmed the AI5 chip design had been finalised for manufacturing ('taped out') ahead of schedule. It is pitched as a chip for running AI on the device itself, serving Optimus, data-centre workloads and the autonomy programme alike, alongside talk of a research fab at Giga Texas.

The same call delivered the bill. Asked how Hardware 3 cars would ever reach unsupervised driving, Musk said Hardware 3 simply does not have the capability, pointing to memory bandwidth roughly one-eighth of Hardware 4's. Tesla offered upgrades or discounted trade-ins to the FSD buyers whose cars cannot get there. Customers who paid for a future feature were told the future needs different hardware.

Autonomy was sold for years as software waiting to arrive. In 2026 it invoices as cameras, memory bandwidth, chip yields, service-centre throughput, data-centre capacity, regional validation, insurance terms and a municipal permit. The car is the smallest line item on the list.

Where the market actually settles

Waymo is selling a driver it can install into vehicles and cities. Tesla is selling a subscription to 1.3 million people while it works on the fleet. Baidu is leaning on Chinese municipal scale and exporting the result. Zoox, Amazon's unit, builds a purpose-made pod with no steering wheel and no pedals — and since March has been bookable through Uber, launching in Las Vegas with Los Angeles slated for mid-2027. Uber, having built none of it, sits between the rider and whoever can supply the trip.

For anyone procuring this technology rather than reading about it, the useful questions have shifted. Not "can the car drive itself" — it can, somewhere. The questions are which streets, inside what operational design domain, on whose insurance, under which regulator, with what recall exposure, and who picks up the phone when a hundred of them stop at once.

Tesla's brand name remains the loudest phrase in the sector. Its own disclaimer is shorter, and it has not changed: the features do not make the vehicle autonomous.