Tesla's Cybercab is taking fares. What changed, and what it means for German SMEs

Tesla held a closed, invitation-only event in Austin, Texas, with no livestream and without Elon Musk in attendance. A handful of invited guests took the first rides. From 4 September at 5 p.m. local time, the Cybercab has been carrying paying customers through the city.

The vehicle is a two-seater built at Gigafactory Texas. It has no steering wheel, no pedals and no fallback for a human to take over. It drives on cameras and Tesla's own AI computer only, with no lidar and no radar; braking and steering are fully electronic. At launch, 45 Cybercabs were authorised for commercial robotaxi use in Texas. Ashok Elluswamy, Tesla's head of AI, announced one million miles of unsupervised robotaxi operation, and Musk said Tesla intends to build several times more Cybercabs per year than all its other vehicles combined, at a target price below 30,000 US dollars. Tesla named no fare, no delivery date and no next city.

Self-driving Teslas are not new. The fleet had covered more than 12 billion miles on FSD (Supervised) by Tesla's second-quarter 2026 report, and Tesla-focused trade coverage put the figure at around 13 billion in August. The difference sits in the word in brackets: with FSD, a human sits at the wheel, carries legal responsibility and intervenes when the system asks. The Cybercab removes that human. The car is built around the software, and whoever gets in is a passenger, nothing more.

Three figures were mixed together in this week's coverage and they measure different things. Tesla FSD (Supervised): more than 12 billion miles with a responsible driver, and Musk had named 10 billion in January 2026 as the data threshold for unsupervised driving. Tesla robotaxi, unsupervised: one million miles with nobody in the car. Waymo, driverless: more than 200 million miles with nobody at the wheel.

Setting one million against 200 million ignores that Tesla holds the largest driving dataset in the industry, roughly 60 times Waymo's driverless mileage. Setting 12 billion against 200 million ignores that Waymo has operated without a driver for years, while Tesla has been collecting that experience for barely one. Both numbers are correct. Tesla has the data; Waymo has the operating record without a fallback.

Waymo, Alphabet's robotaxi company, builds on redundancy: cameras, lidar and radar fitted to vehicles from other manufacturers, around 4,000 robotaxis in 14 US cities and 500,000 paid trips per week according to TechCrunch on 1 September 2026. Two days before Tesla's event, Waymo announced three further cities. Its vice president Srikanth Thirumalai put the position plainly: cameras are incredible, but they are not enough.

Tesla builds on cameras, AI and volume. Analyst estimates put the Cybercab's production cost at 23,000 to 25,000 US dollars, against 70,000 to 150,000 dollars for a Waymo vehicle. If Tesla's data is enough to drive safely without lidar, Tesla wins on price. If it is not, Waymo was right. The question is no longer settled by demos but by incidents per driverless mile and by permits.

Matthias Wessner's back-of-the-envelope figure is a purchase price of 18,000 to 25,000 US dollars against roughly 50,000 dollars of fare revenue per vehicle per year, which would return the purchase price in six months. The revenue side of that estimate is conservative. Waymo's 500,000 paid rides per week across around 4,000 vehicles, at 15 to 17 dollars per ride, work out to roughly 90,000 to 105,000 dollars per vehicle per year, and Tesla-focused analysts put a Cybercab at 200 to 400 dollars per day at high utilisation.

The cost side is what the six-month figure leaves out. Musk's 0.20 dollars per mile is a target for 2030 at scale, and ARK Invest models the same number. Current estimates for a Cybercab run at 0.30 to 0.50 dollars per mile before depreciation: about 2.6 cents for electricity, 7 to 9 cents for insurance, 5 to 10 cents for remote supervision, plus maintenance, cleaning and empty miles. Waymo, with far more sensors on bought-in vehicles, sits at 1.40 to 1.98 dollars per mile by third-party estimates and still reports losses despite its revenue per vehicle. On 50,000 dollars of fares and around 40,000 miles a year, a Cybercab at today's cost estimates clears 30,000 to 38,000 dollars. Against a 25,000-dollar vehicle that is an 8-to-12-month payback, longer in the first years of remote supervision and depots, and about seven months at Musk's target cost. A fleet owner on Tesla's network would also hand over a platform share that Tesla has not published. All figures in this section are analyst and trade estimates, not Tesla disclosures.

Even at twelve months, the payback is unusual. Investments rarely return their price in under a year, and the ones that do almost always replace labour: an AI phone assistant against a receptionist pays back in three to nine months, a collaborative robot in 12 to 18, a rented H100 GPU in 18 to 24, rooftop solar in Germany in 8 to 12 years. A ride-hailing car bought for 15,000 dollars earns 50,000 to 60,000 dollars a year, but the driver is the cost, 60 to 70 percent of every fare. The robotaxi is that calculation with the driver removed, and the same arithmetic applies to any process in which a person currently answers the phone, sorts the inbox or checks invoices.

Six months ago, a common estimate in Germany was that driving of this kind was five to ten years away. On 10 April 2026 the Dutch vehicle authority RDW became the first in Europe to approve FSD (Supervised). On 22 May 2026 Germany's Kraftfahrt-Bundesamt followed. Six weeks separated the two decisions.

German manufacturers moved in the opposite direction over the same months. Mercedes-Benz paused Drive Pilot in January 2026, the Level 3 system that had run on German motorways since 2021, and fitted the 2026 S-Class with a hands-on Level 2 assistant instead. BMW is removing Personal Pilot L3 with the 7 Series facelift in April 2026. On 2 July 2026 Volkswagen confirmed the end of its automated-driving alliance with Bosch, started in 2022 through its software subsidiary Cariad: around 1.5 billion euros invested, no continuation beyond Level 3, and a decision to buy the technology in from now on. Within six months, all three German groups have given up developing Level 3 and above for private cars themselves.

What still drives without a driver in Germany runs on bought-in technology. Volkswagen's subsidiary MOIA operates the ID. Buzz AD on Mobileye's system, with 27 sensors including nine lidar units, testing in Los Angeles and Hamburg, with commercial rides planned with a safety driver by the end of 2026 and without one in 2027. Mercedes is running a Level 4 prototype in Abu Dhabi with Nvidia. In private cars, the German brands have stepped back at the very moment Tesla is selling the level above as a taxi and getting its supervised system approved in Germany.

Germany's public transport operators are short of roughly 20,000 bus drivers and 3,000 train drivers today, according to the industry association VDV. Forty percent of bus and tram drivers are 55 or older, and about 6,000 retire every year until 2030. At the same time, autonomous Level 4 shuttles in the KIRA project have carried passengers in the Rhine-Main region since May 2025, Hamburg's ALIKE project puts up to 20 autonomous vehicles from HOCHBAHN and MOIA on the road in 2026 with a safety driver on board, and Uber plans robotaxi tests with Momenta in Munich in 2026. Germany passed its law on autonomous driving on 28 July 2021, and the implementing AFGBV regulation took effect on 1 July 2022: Level 4 vehicles may operate in an approved area with a technical supervisor outside the vehicle, which is exactly the basis KIRA has run on since May 2025. Regulation is not the blocker. An operator waiting for the legal framework is waiting for something that has existed for four years.

A contact at a large municipal transit operator told Matthias Wessner, managing director of RIT Services, this week that autonomous vehicles are not on the agenda there at all. Yet as early as next year, if operators and authorities move together, autonomous vehicles could fill the gaps that already exist: off-peak hours, outlying routes and feeder services for which no driver can be found. Not as a replacement for the tram, but as the part of the timetable that is currently being cancelled because nobody is available to drive it.

Optimus, Tesla's humanoid robot, runs on the same network architecture as FSD and is trained on the same compute cluster, according to Tesla's statements since its 2022 AI Day and to Tesla-focused trade coverage in 2026; independent verification does not exist yet. A system that finds its way through traffic on cameras alone is learning something general: how to move through a world nobody tidied up for machines. Road today, warehouse and living room later. The 12 billion miles are not only the dataset for a taxi; they are the foundation for machines that act in the physical world.

Lesson one: AI is moving from assistant to actor. Tesla went from supervised to unsupervised. The same threshold is arriving in bookkeeping, customer service and purchasing. The question is no longer whether AI may suggest, but in which process it may decide without sign-off.

Lesson two: let it assist first, then hand over responsibility. Tesla collected 12 billion supervised miles before the driver was allowed out of the car, and learned above all where humans had to intervene. A company that lets AI propose today and records every case in which an employee corrected it is building exactly the evidence it needs to let AI decide later. The RIT Services guide on which processes to automate first covers how to pick that starting point.

Lesson three: approval is the bottleneck, not the technology. The Netherlands and the KBA showed that an authority can decide in weeks when the data is on the table. For a company, the EU AI Act demands a risk classification and documentation for every AI deployment. Doing that early makes scaling faster, not slower; the free Compliance Check platform at compliance.rit.services walks a company through that classification.

Companies without a repeatable, measurable process have nothing to hand over to an AI system, and no corrections to collect. The lesson applies once a process runs often enough to produce data, not before. Businesses in sectors where a human decision is legally required, such as parts of medicine, law and lending, should read the robotaxi story as a signal about pace, not as a template. And a company that is still choosing between an off-the-shelf assistant and custom development is better served by the make-or-buy analysis on this site than by anything Tesla did this week.