Mercedes-Benz’s Wang Xin on German Autos’ Tech Turning Point

Summary

• R&D decision-making power is shifting east. Mercedes-Benz China R&D has moved from “testing and localizing” to leading design: it launched the “R&D 3.0” campaign on its own in September 2022, with more than 2,000 engineers in Beijing and Shanghai, the company’s largest overseas R&D operation; China now leads global parking-function development, meaning “even cars designed and built in Germany are running Chinese algorithms.” Wang Xin’s trust formula is “we commit, we deliver, we trust”; the 2026 goal is an “In China for Global” intelligent cockpit platform, with the entire vehicle development process in China “very likely” one day. Zhang Xiaojun said, “I feel like I can already see it.”
• The technology stack is undergoing a major turn. This year’s CLA, built on the NBOS 1.0 architecture, is Mercedes’ first end-to-end, pure-vision model—with no HD maps and no lidar, “the only one besides Tesla.” The logic is that once training is good enough, lidar “doesn’t really add much significance, but it does add cost” and can introduce noise; the decision was based on 20,000-30,000 km of comparative testing with and without lidar. The L3 architecture retains lidar as redundancy, while L4 “may not necessarily have it for now.”
• The real cost of going end-to-end was 7-8 months of wrestling. After the initial redeployment, “you see obvious regression,” and the team briefly considered rule-based driving on highways and end-to-end driving in cities. It ultimately committed to a fully end-to-end stack on the strength of extensive test data. This “battle with itself” was driven spontaneously by the China team and its partners, not mandated by headquarters.
• Twenty years of automotive technology can be divided into four cycles. Before 2004 was about the car itself; 2004-14 was in-car audio, entertainment and navigation; 2014-20 was early intelligent driving; and 2020 to today has been the era of massive compute. In 2020, rule-based systems could never enumerate every long-tail corner case, leaving the industry “in a very broken state”; Nvidia Orin and faster model iteration brought a step-change, while the 2023 HD-map paradox—updates could not keep pace—drove the shift to mapless, pure-vision systems. “In the next cycle, everyone will start competing in L3.”
• Safety is the differentiator. Mercedes should be the only automaker in China using QNX for its intelligent-driving operating system (“we call ours Safety Linux, but it’s still Linux”), with the intelligent-driving safety level basically reaching “S2D” (likely ASIL-D); new vehicles undergo at least nearly 5 million km of testing, 150,000 km of single-function durability testing, 150 real-vehicle crash tests and 15,000 dual-vehicle virtual collisions. “For many families, 0.1% is not 0.1%—it is a family.”
• Luxury’s definition has not changed—that is the moat. “A luxury car is only a real luxury car if it is regarded as one by ordinary people”—138 years of brand accumulation cannot be built overnight. Digital luxury means segmenting user profiles and perfecting the selected functions: “It’s not that you have it and I have it too; it’s that I have it, but I do it better.” Luxury or technology? “Extreme technology is also luxury, so I would choose luxury.”
• AI is the next battleground. Mercedes is looking for the “fourth function” in the car—after navigation, possibly infotainment, and the growing role of ADAS—and believes it “should be related to AI.” An 8-megapixel forward-facing camera plus large models could move perception from identifying “what something is” to understanding “what it wants to do.” Its language-model partnership is with Volcengine’s Doubao; once L3 becomes widespread, car-to-car communication must be redefined, because current V2X is generally not high-precision positioning and “you don’t know which car you sent it to.”
• This is a once-in-a-century inflection point, and he is candid about it. “The focus of technological development is actually not in Germany, but in China or the United States.” The burden is real too: G-Class models from more than 40 years ago are still on the road, spare parts must be stocked for 20 years, and separate teams must maintain multiple platforms. “I wish new automakers would one day have this kind of pain; that would mean they had survived for 100 years.” The bottom line: “We are still a profitable company.”

Deep dive

1. Twenty Years, Four Cycles: From “Four Wheels and Two Sofas” to the Race for L3

• Wang Xin opened with a history lesson: in 2002-03, consumer-electronics heavyweights crowded into auto manufacturing. Broad, marketed as the “fighter jet of mobile phones,” had built cars; Chunlan Air Conditioner had built them too. “Everyone thought four wheels and two sofas were enough.” They used CAD simulations to copy the exterior, but had none of the required rigidity or know-how, could not pass crash tests, and exited within 2-3 years. Zhang Xiaojun pointed out how similar that sounded to today; Wang did not disagree: “The auto industry has gone through several recurring cycles, iterating wave by wave.”
• He divides the past 20 years into four cycles: before 2004, the focus was the car itself; 2004-14 brought audio, entertainment and navigation—6.1-channel Dolby, navigation entering the car, and optical-disc maps weighing 5-6 kg in 2006-07, before maps finally left the disc after 2010; 2014-20 was early intelligent driving; and 2020 to today has been the era of massive compute, with intelligent assisted driving reaching “a mid-cycle stage.”
• The mechanics of the 2020 collapse are worth remembering. Everyone predicted L3 would arrive in 2020, but the deeper rule-based testing went, the more long-tail corner cases appeared, and “you could never enumerate them all.” The industry was “in a very broken state,” while L4 companies collectively retreated after failing to find a commercial loop. Nvidia Orin’s massive compute and faster model iteration brought improvements in 2022-23. Then the 2023 consensus around “lidar plus HD maps” ran into its own paradox: map updates could not keep pace, and by year-end some manufacturers were testing mapless systems and using pure-vision perception to fill the gap. “I believe that in the next cycle, everyone will start competing in L3.”

2. Tier 1 Retreats to the Factory Floor, OEMs Reclaim Design—What Stays Together Splits, What Stays Apart Reunites

• The power shift is driven by data. Cars generate data on the road, and intelligent driving lets that data directly improve performance. The industry is moving from “Tier 1 delivers a black box to the OEM, and the OEM has no idea what is inside” to OEMs operating all the data themselves. Tier 1 suppliers are moving toward white-box delivery and retreating to manufacturing, while design returns to the OEM so “iteration can move faster.”
• The history is cyclical. Delphi and Visteon were once divisions of GM and Ford, spun off around 2000 because keeping them in-house was uneconomic. “What stays together splits, what stays apart reunites.” The force bringing them back today is different: not commercial considerations, but faster response centered on the user. Wang spent nearly 18 years at Delphi and joined Mercedes three years ago: “At my age, a Mercedes-Benz has always been a personal dream.”

3. R&D 3.0: A Self-Started Twelve-Month Campaign During the Pandemic

• When Wang joined, the mission was “home away from home”: bring authentic German design to China for testing and localization. The team soon found the product assumptions in China and Germany were too different, so it launched R&D 3.0 around five pillars—3D navigation, enhanced intelligent driving, longer app usage, parking and more. The campaign ran from September 2022 to September 2023, ahead of the new E-Class entering mass production in December 2023.
• The crucial detail is that this was started by the China team, not assigned by headquarters. During the pandemic, communication between China and Germany was effectively blocked. Shanghai colleagues took German code home, “locked themselves in and studied it,” ran simulations, built demos and sent them to headquarters. “Everyone was very surprised and said, ‘You can really do this?’ We said, ‘We really can.’” The Shanghai team had zero people in 2021; today it numbers in the hundreds.
• Parking illustrates the point. German parking happens mostly above ground, while most Chinese parking happens underground; reflective garage floors create heavy noise for ultrasonic sensors, a problem Germany rarely encounters. The China team solved it with visual compensation, lifting the parking-in rate from 70% to above 95%. The same applies to the 3D navigation engine: China developed it for the domestic market, and “you cannot buy these functions in the German market.”
• The 2023 Shanghai Auto Show was the turning point. Headquarters was “very surprised” that China’s auto industry had not stalled during the pandemic but had moved faster. What they saw was not a PowerPoint deck but working vehicle demos. Ola offered substantial appreciation and praised the team for moving forward with an entrepreneurial spirit. The trust formula was simple: “we commit, we deliver, we trust”—no overcommitment, but ambition backed by on-time, high-quality delivery.

4. Taking Stock of the Voice: What China Leads, Influences and Still Defers to Headquarters

• Beijing and Shanghai now have more than 2,000 engineers—“they should have doubled” in recent years—and form Mercedes’ largest overseas R&D operation. Chinese engineers lead and design much of the vehicle’s interior, exterior and electronic/electrical systems. The rule is clear: anything launching in China is led by the China team; navigation and music applications do not even exist overseas, and road testing must take place in China. At the architecture level, China is “slowly influencing” headquarters, including proposals to put AI elements into the system architecture and launch them in China first, using data iteration to find the killer feature.
• China-developed features must be sent back to Germany or the United States for retesting, fresh data collection and model training, and local regulatory approval—a 6-12 month cycle. Parking has already become a global program led by the China team: cars produced in Germany are running Chinese algorithms.
• The teams meet their global counterparts every week. “We never said China is China and global is global.” What has changed is the posture: headquarters used to issue directives—“you must do it this way”—whereas China now uses data analysis and surveys to prove how something should be done, and headquarters is increasingly adopting those recommendations.

5. Battle Culture: The QNX Outlier, 1.9 Seconds and a UI That Made Them Mad

• Mercedes’ strongest technical position is that it should be the only automaker in China using QNX for intelligent driving. QNX is closed and highly safety-oriented; “behind every instrument cluster people are making now, the operating system is definitely QNX.” Intelligent-driving safety should basically reach “S2D” (likely ASIL-D). The message to competitors is blunt: “We call ours Safety Linux, but it’s still Linux.” “Safety is always our first priority. Even when it is extremely difficult, we still choose QNX.”
• Asked whether he ever felt frustrated, Wang said yes, definitely. UI is the main battleground. When Chinese web interfaces were shown to German colleagues, they said, “I can’t even find the function you want; you have too many buttons.” The German-designed interfaces required so many steps to reach a function that “basically nobody would use it.” The likely answer is a fully China-specific UI.
• Cross-border battles have distinct styles. American companies are direct and want to solve many problems at once, making the process highly intense and the atmosphere difficult to control. German companies are more measured. The only way through is: “Let the data and facts speak. Everyone is an engineer.”
• German thoroughness showed up in two examples. When Mercedes developed in-car karaoke, engineers first asked whether a handheld microphone could cause secondary injury in a crash; the answer was a customized microphone. For the passenger display, hundreds of tests produced one number: if the driver looks to the left for more than 1.9 seconds, the passenger-screen brightness automatically drops—without affecting the passenger’s viewing experience or distracting the driver.

6. CLA: End-to-End and Pure Vision, “the Only One Besides Tesla”

• The CLA launched in China this year after its Rome debut and combines two rare features: it is the only model using QNX and “basically the only one with neither maps nor lidar—apart from Tesla, there is no one else.” The reasoning behind dropping lidar is explicit: once training is good enough, “adding lidar doesn’t really add much significance, but it does add cost.” More sensors do not necessarily increase confidence; sometimes they add more noise. The decision came after 20,000-30,000 km of comparative testing with and without lidar.
• The end-to-end transition was a “battle with itself” and “quite painful—a change in mindset.” After redeployment, “you see obvious regression.” The team wavered: if highway scenarios were not rich enough, should highways remain rule-based while cities went end-to-end? It tested many combinations over 7-8 months. There was no single aha moment; new- and old-architecture vehicles ran side by side, and “as we saw the new architecture getting better and better, we felt this was the direction we should take.” The process was led throughout by the China team and its partners: “We saw some good examples and felt we should be able to do it too.”
• Pure vision is not a final-position doctrine. The L3 architecture retains lidar as “very good redundancy,” while Wang questions the value of seeing beyond 300 meters: “If you see something more than 300 meters away, do you need to make a control decision right now?” L4 “may not necessarily have it for now,” and dropping lidar will not become mainstream in L3 or L4. L3 cannot be skipped: “Only after you do L3 can you capture more data and scene experience, which helps you get to L4.” Mercedes is studying the economics from both directions to find the balance point.
• Mercedes was “the first in the world to receive L3 certification” in Europe and the United States. Beijing has both L3 and L4 testing permits. Products based on the new platform will roll out progressively over the next 1-2 years. The language-model partner is Volcengine’s Doubao.

7. AI Gets in the Car: The Fourth Function, One-Person-One-Interface and Car-to-Car Communication

• The perception upgrade is concrete. The forward-facing camera is moving from 2.5 megapixels to 8 megapixels; combined with large models, this could move perception from “what something is” to “what it wants to do.” If a crossing e-bike rider looks left and right, the system could infer that the rider intends to turn right rather than automatically slowing down every time. Passing a primary school, it could distinguish children from adults and slow more cautiously. A future architecture could connect Android in the cockpit, QNX in intelligent driving and AI agents through SOA so they understand what one another is trying to do.
• Wang is searching for the “fourth function” in the car. Navigation is the most frequently used; the second could be infotainment; ADAS is becoming more common. “Everyone is looking for the fourth, but we believe it should be related to AI.” Once L3 frees up driving time, the car becomes a third space for work and leisure. “Could you livestream from the car? Definitely not now.” UI will also move from heavy homogenization toward one-person-one-interface: the car will truly belong to its owner, with the interface tailored to individual usage habits.
• The judgment on car-to-car communication has a clear trigger. The L3 era will bring mixed traffic, with L3 and non-L3 cars sharing the road. In a cut-in, “how do you understand one another? It will inevitably create many conflicts.” Car-to-car communication must therefore be redefined. The current weakness is practical: V2X is generally not high-precision positioning, so “you don’t know which car you sent it to—you don’t know which of four or five lanes the communicating car is in.” The state should lead the definition, with automakers contributing proposals.

8. The Luxury Moat: Not a Pile of Features, but Perfecting What You Choose

• The core argument is simple: “A luxury car is only a real luxury car if it is regarded as one by ordinary people, rather than something you declare to be luxury.” A 138-year brand history cannot be built in a day. Wang rejects overdesign for gimmicks: backend data shows navigation and music are used frequently, while memory parking sees limited use because many people do not have fixed parking spaces. The differentiator is: “It’s not that you have it and I have it too; it’s that I have it, but I do it better.”
• Digital luxury moves from the analog-era logic of “you want leather seats, so I give you leather seats” to finely segmented user profiles. AMG customers are more aggressive; Maybach customers prioritize comfort; some luxury customers chiefly want “extreme quiet and very good fragrance.” The direction is clear: “Give people what they want, not what I want to give them.” Too much design still comes from knowing what we want to offer, without knowing what customers actually want.
• Extreme performance has hardware costs. To make voice interaction good, “your token count will be higher than other people’s.” A 4-megapixel camera can also support intelligent driving, but its perception is less rich, which is why Mercedes moved to 8 megapixels. Current equipment includes Qualcomm 8295 in the cockpit and Nvidia Orin for intelligent driving. Asked to choose between luxury and technology, Wang answered: “Luxury and technology do not conflict. Extreme technology is also luxury, so I would choose luxury.”

9. The Mercedes Standard: 0.1% Is Not a Number, It Is a Family

• The safety specification is quantified: a new vehicle undergoes “at least nearly 5 million km of testing,” more than 150,000 km of single-function durability testing, testing across extreme cold, heat and humidity in multiple countries, 150 real-vehicle crash tests and 15,000 dual-vehicle virtual collisions, plus a driving simulator modeled on a pilot-training cockpit. His advice to Chinese automakers is to put people first and retain some conservatism: “0.1% or 0.2% may be just a number to us, but for many families, it is a family.” You can keep adding intelligent features, but you cannot forget that it is still a car—and it must be the safest thing it can be.
• Two historical footnotes remain. In the 1980s Prometheus program, a car drove on top of a truck while another truck behind carried all the compute; “today, one chip can probably handle all of that.” Carl Benz exhausted all his money building the car; without his wife serving as his first driver and testing the vehicle, “perhaps we would not have cars today.” The crumple structure designed in 1950-51 became the prototype for every car on the road today. The motto remains “The best or nothing”; Wang’s own formulation for leading his team is “surpassing Mercedes within Mercedes.”

10. A Once-in-a-Century Inflection Point: The Focus Is No Longer in Germany

• The most candid self-diagnosis was this: “We have adapted for more than 130 years, and every decision and idea came from Germany. Today, the focus of technological development is actually not in Germany, but in China or the United States.” For Mercedes, this is a once-in-a-century inflection point. The biggest concern is that people will not understand its products. The narrative that Mercedes is merely a traditional automaker with weak loyalty among younger consumers is misleading, he argues: “Young people grow old too. The luxury I understood when I was young may not be the luxury I understand today.”
• The burden of transformation is real. G-Class models from more than 40 years ago are still on the road; ordinary spare parts generally need to be stocked for 20 years; and dedicated teams must always maintain different platforms. “I wish new automakers would one day have this kind of pain. That would mean they had survived for 100 years.”
• The posture for breaking through is not to grit its teeth and endure. “Every company faces challenges in this fiercely competitive environment in China—we are still a profitable company.” Headquarters should look humbly at competitors’ technology, while the company should “embrace selectively”: “We cannot chase every change that appears outside; that would waste resources.” The priority is still to build the best car for what customers want. The most important task for 2025 is the successful launch of CLA. In 2026, the China team aims to build an “In China for Global” intelligent cockpit platform. Could the entire vehicle design and production process happen in China? “Very likely.” Zhang Xiaojun said, “I feel like I can already see it.” Headquarters’ commitment to investment in China is a firm, unshakable policy.
• The organizational routine is simple: fewer emails, more face-to-face communication, weekly stand-ups and quarterly OKRs. How angry can Ola get? “He just keeps asking questions, asking you to answer.” Wang said: “No, absolutely not. It is a very elegant ‘magic weapon’”—his original wording. Zhang replied: “That is a very good adjective.”