In northern Munich, a group of invited people opened Uber and called a car.
A safety operator sat in the driver's seat. The ride was free.
That is the pilot Momenta and Uber announced on November 18: operator on board, geofenced, invitations in batches, no fare, under a German test permit. Five conditions, none resembling a product launch. It looks like an entrance exam.
A permit is harder to get than an algorithm
Self-driving companies are never short of demo video. They are short of one sentence: you may drive on my roads.
Software iterates. A permit negotiates something else: who carries liability in a crash, whether data stays in country, who certifies the operator, who pays. None of that is in the code. All of it decides whether the car moves.
Germany is the car industry's heartland and one of its most cautious regulators. A permit there is an asset that earns nothing and makes the next negotiation easier.
So the real result is not that a few cars drove. It is the clause: under a test permit from the German authorities.
The car is yours, the door is someone else's
Momenta has the algorithm. Uber has the entrance.
An entrance cannot be built quickly. It is an icon already on the phone, a payment method already linked, and the trust that a car will come when called in this city. Those three take a decade, tens of millions of trips and countless complaint tickets. Building that from zero in Munich would mean years, and an unfamiliar user habit. Plug into Uber and there is a passenger on day one.
The price: the passenger remembers Uber, not the car.
Aviation has run this way for decades: the aircraft belongs to a leasing company, the crew is outsourced, the passenger knows the logo on the ticket. Contract manufacturing: the factory makes the product, the brand stays elsewhere. Autonomy is sharper, because what it sells is not a trip but the belief that the system can be trusted. Trust held under another company's sign is hard to transfer.
So nearly every autonomous driving company walks on two legs: orders through a platform, users of its own. The first keeps it alive. The second is the future.
A free ride is not charity
What does one free trip buy?
Data: real roads, passengers, weather, none of which simulation makes. Munich in November: slick surface, low light, thick coats. All different from any Chinese city, and where models fail.
Simulation can generate ten thousand intersections. It cannot generate one fog with moisture in it.
An observation window. A regulator does not let you in because it trusts you, but to watch. Disengagement rates, complaints and near misses all feed the assessment. And reputation: Europe is more cautious than America and more fragmented than China. The first passengers' word of mouth beats any advertisement.
The cost is fuel, a driver, depreciation. The return is three things money cannot buy.
The three arrive in order. Without data, the window is failed. Fail it, no expansion. Without the permit, the best reputation runs in one district.
Free is not a discount. It is a ticket.
What the safety operator is doing
The safety operator is a compliance requirement and a data labeler.
Every takeover is a negative sample: the model judged wrong, judged late, or did not see.
Those samples are worth more than successes. Success tells you what passes. Failure tells you where passing ends.
The role is a cost now, a pen later, drawing the real world's boundary one line at a time. The day he gets out, part of the drawing is done. The takeover itself is data too: why this second, what he saw, what he judged, all annotated afterwards. Those labels say more about weakness than mileage does.
A million kilometers explains less than one takeover.
The order of going abroad has inverted
Chinese companies used to go out in sequence: build the product, find the channel, discuss compliance.
Autonomy reverses it: get the compliance position, enter through someone else's channel, validate the product along the way.
The reversal follows from risk. A car that fails is not a bad review. It is a crash. No regulator lets you run first and talk later.
So "get on the road" becomes "get the ticket." The ticket belongs to someone else, which means accepting their terms: revenue share, data ownership, brand visibility.
Today those look like concessions. At fleet scale they become leverage.
Every business begins by letting the other side earn: the platform wants supply and a story, the car company data and a permit.
The divide comes in year three. If operating data shows the system is steadier than a human driver, the seats swap.
What does one city prove? Not scale. Process: how to apply, how data crosses borders, how incidents are reported. Run it once, and the second city reuses it.
Process travels. The most expensive line was never the car. It is the tuition for entering an unfamiliar regulatory system. Cash only, no installments. So the first city is chosen for clarity of rules, not size of market. Clear rules allow a retry. Vague rules end the file.
The pilot has three stages: operator, geofence, free. Then: no operator, wider area, fares.
Each step is harder by more than a constant. No operator means the takeover rate must satisfy the regulator. A wider area means roads the model has not seen. Fares mean passengers compare with a taxi and stop forgiving.
None of the three is solved by riding on someone else's app.
So watch the conditions, not the car. The more precisely the five limits are written, the more seriously the exercise is taken.
Going abroad is not shipping cars. It is obtaining the sentence that lets you drive, then turning it, year by year, into passengers who choose you.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
