Nvidia is continuing to expand its self-driving car business. The AI model Alpamayo is intended to judge like a human and is used in the Mercedes CLA. A test drive through Berlin.
Mercedes with Nvidia technology: Jensen Huang's company wants to make cars autonomous. Photo: AFP
Nvidia's AI model Alpamayo navigates through Berlin's city traffic using data from human driving behavior and synthetic data and makes autonomous driving decisions.
Mercedes-Benz and Nvidia announced the expansion of their partnership to commercially offer vehicles with the Alpamayo driving system, which is designed for autonomy level 2++, by the end of the year.
Nvidia handles the chips, sensor architecture, software and safety checks for autonomous driving; Mercedes supplies the vehicle and the MB.OS operating system.
The market for car chips is growing strongly: Nvidia achieved sales of $1.69 billion in the car division in 2025, and analysts expect $8 billion for this year.
Berlin, Düsseldorf, Munich. When Nvidia's AI model takes over the wheel, you can only tell by the safety driver's hands. You lie loosely on his lap, palms up, as the AI slides the steering wheel through his fingers.
When a group of young people seem unsure whether they want to cross the street, the car brakes briefly and then drives past them. The AI seems to navigate Berlin's city traffic more confidently than many people. Handelsblatt was taken on a 30-minute test drive through Berlin with Nvidia's head of the automotive division, Ali Kani. The car was controlled by Nvidia's AI model Alpamayo.
For Nvidia, autonomous driving is the first major hurdle to bringing AI from screens into the real world. To do this, they cooperate with car manufacturers such as Mercedes-Benz and Volvo. Before the GTC, Nvidia announced expanded partnerships with the two car manufacturers. “For Nvidia, this is an opportunity worth more than a trillion dollars,” says Auto boss Kani. According to him, every car will one day drive autonomously - and Nvidia is one of the few companies that can accompany this change.
Eleven years ago, at the GTC in March 2015, Nvidia CEO Jensen Huang announced the “big bang” for self-driving cars that was only a few years away. But the Big Bang has so far failed to materialize. That should change with Alpamayo. Huang called the model the “ChatGPT moment” for autonomous driving when it was unveiled in January 2026.
Why did it take so long? Ali Kani cites several reasons: the available computing power, the high cost of sensors, and not enough data on which to train. “We didn’t have the right tools to build a safe product,” he says. According to him, Nvidia first had to design a new type of software development for autonomous driving. Kani has headed Nvidia's automotive division since 2018.
More assistant than autopilot
The practical test of the car AI takes you through the inner-city districts of Kreuzberg and Friedrichshain in a Mercedes CLA. From the parking lot at the Gleisdreieck subway station in Berlin-Kreuzberg, the security driver navigates the car himself. Then he drops his hands and Alpamayo takes over. However, the safety driver's hands must constantly touch the steering wheel.
More than ten cameras, five radar sensors and twelve ultrasonic sensors in the car will provide the model with information about the current traffic situation, explains Ali Kani. Alpamayo concludes how the car can best navigate the traffic situation. “It’s almost like the model is solving a puzzle,” says Kani. The system then translates the solution into an action command for the car. The model does the calculations on the Nvidia chips in the car.
Alpamayo was trained on data from “exemplary human driving behavior,” Kani said. The data comes, among other things, from trips from other car manufacturers who also use Nvidia systems. The model continues to be trained and evaluates its own driving behavior in order to improve. The real data is supplemented by synthetic data.
By “thinking”, also called “reasoning” in AI models, it was not necessary to train in detail for every conceivable situation. "When you're developing autonomous driving, you'll never have all the data in the world. You need a model that can handle situations it hasn't seen before," says Kani.
However, nothing really unforeseeable happens on the test drive. People walk onto the street between parked cars, cyclists push through cars at red lights and construction sites narrow the streets. Alpamayo masters these situations that are more normal for big city traffic.
Nvidia car boss Ali Kani with reporter Lina Knees: test drive through Berlin. Photo: Lina Knees
Then the safety driver has to grab the steering wheel at a construction site. A new route had been drawn on the road in orange paint. Alpamayo seemed to ignore him and turned the steering wheel too hard. The safety driver corrected.
Nvidia needs car manufacturers
From 2027, Mercedes-Benz would like to offer this driving system in Germany together with Nvidia. The system is scheduled to be used in Stuttgart and Munich at the end of 2026. Both companies announced this on Wednesday.
Mercedes provides the car and the MB.OS operating system. And with it the interface between vehicle and AI functions. Nvidia contributes computing chips and sensor architecture for the autonomous driving functions. There is also the software for autonomous driving, the Alpamayo models and a security system called Halos, which additionally checks the driving maneuvers for dangerous and illegal actions.
Even though the driver only intervenes once during the entire test drive, the AI that steers the car through Berlin city traffic is more of a driving assistance. Mercedes and Nvidia speak of Level 2++ autonomy. At level 2, the AI can take over steering, braking, accelerating and changing lanes. However, the driver still has responsibility and must watch carefully. At level 3, the driver would only have to intervene when requested by the system and can otherwise read emails or watch videos.
According to Ali Kani, “Level 2++” is not an official term in the auto industry. The additional plus signs stand for the autonomous capabilities in city and highway traffic.
Alpamayo, which Nvidia makes freely available, is actually designed for autonomy level 4. At this level, the car would drive so independently that the driver could even sleep.
According to Paul Miller, analyst at the market research company Forrester, the focus on Level 2++ with Mercedes could unlock enormous potential: “Precisely because it’s not about replacing every car with a robotaxi,” says Miller. Instead of doing without a driver, as is the case with fully autonomous robotaxis, the next car that customers buy should be just a little more autonomous. This would be well received by customers who occasionally want to drive themselves.
Auto industry
The future of BMW and Mercedes depends on these computer chips
Nvidia relies on both: It equips driverless robotaxis and offers AI assistance when driving. The chip company also maintains several strategic partnerships with car manufacturers such as Jaguar, Land Rover and Hyundai. Volvo also recently announced that it wanted to develop autonomous vehicles with Nvidia.
Beyond the chips, Nvidia offers a lot of what it needs for autonomous driving: from the chips to the training data to the AI model. According to Kani, the automotive industry is currently the only one where Nvidia offers everything. Customers could freely choose what they would like to use from this set. "We don't want to tie anyone down. Whatever suits the partner best, we want to make that possible."
Analyst Paul Miller sees it a little differently: Although the components can be used individually, that is more complicated. “Nvidia offers the entire menu in a way that is difficult to compete with,” says Miller.
The vehicle as a rolling data center
Nvidia earns from Mercedes’ sales of autonomous driving and from the licenses for the software. Normally, car manufacturers partially cover the costs incurred by suppliers for development. However, Nvidia has made advance payments for the chips and is now being paid for them.
It was worth it because demand for car chips is booming and manufacturers are tying themselves to suppliers at an early stage. In 2025, the sales of Nvidia's car division were still at 1.69 billion US dollars, analysts at the US investment bank Jefferies estimate it at eight billion dollars for this year.
chip industry
Billions for competitors: This is how Nvidia turns rivals into allies
The French market researcher Yole estimates that semiconductor companies generated around $77 billion in sales from car chips alone last year. It is a rapidly growing business; according to specialists, it will be worth 160 billion by 2031.
Theoretically, Yole sees Nvidia as being far ahead for automated driving with its latest chip generation, Thor. The group is thus “heralding a new era in the area of AI computing power for the automotive industry”. Thor brings performance like a data center to the automobile, which in the future will have to provide infotainment as well as process camera images and sensor data from the traffic situation.
Related topics
NvidiaBerlinChinaTeslaVolvoChatGPT
However, competition is intensifying, especially in China, Nvidia's most important car market, due to new, in-house solutions from local manufacturers Xpeng, Huawei and Nio.
However, Nvidia also supplies other autonomous driving providers, including in China. The largest automotive customer is currently Tesla, says Kani at the end of the trip. “They buy a lot of our GPUs to train their models,” says Kani. He would currently still drive a Tesla, but would soon like to switch to a Mercedes model - with new driving assistance, of course.
