Mercedes-Benz China team uses data to fight for a say with German HQ
The era of Tier 1 black-box delivery is over; OEMs want to operate the data themselves. The China team didn't seize that say — they earned it through commitment, delivery and trust.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Tier 1 goes from black box to white box
Wang Xin lays out the root cause of OEMs taking back design authority very clearly: cars went from mechanical products to data-driven products, a car collects data just by driving on the road, and that data helps intelligent driver assistance iterate continuously. Previously a Tier 1 delivered a black box to the OEM, and the OEM had no idea what was inside; now the OEM says I want to operate all the data inside myself, so my customers get new features faster. So Tier 1 increasingly delivers a white box and gradually withdraws from the design domain into the production domain — manufacturing and cost control can still be done well, but design authority returns to the OEM's hands, and only then can iteration speed and project schedule be kept under control.
— Wang XinTier 1 was originally spun out of the OEMs
Many people assume the OEM's dependence on Tier 1 is a natural industry structure; Wang Xin says it isn't. Before the mid-1990s, today's Tier 1s all came out of the OEMs — Delphi and Visteon were originally a division of Ford and GM. The OEMs found that holding every division in their own hands made design costs too high, and you needed more production-line cost iteration to amortise it, which didn't pay. So around 2000 Ford and GM successively spun these divisions off, forming large component companies. In other words, the OEM doing everything itself had happened before; the business logic is that it only gets split out when the cost is uneconomical. Now that it's being taken back, the consideration isn't cost but user-centric response speed — the kind of cooperative relationship between Tier 1 and OEM is too slow.
— Wang XinThe say wasn't seized, it was earned through delivery
Asked whether the China team and HQ fought over the say, Wang Xin gave a counterintuitive answer: no, it's that we committed, we delivered, and everyone built trust — in English, we commit, we deliver, we trust. He stressed that many things are built on trust, and without trust you can't move forward. When working with HQ, they always committed to what they could do, never over-committed, but did it with ambition, and ultimately delivered on time and to quality, and the trust relationship was established. This explains why the China team could take the lead — not by winning arguments, but by delivering again and again.
— Wang XinMercedes-Benz is the only carmaker using QNX for its ADAS OS
Wang Xin says Mercedes-Benz should be the only carmaker in China using QNX rather than Linux for its intelligent driver assistance operating system. QNX is closed and has a very high safety level — what runs behind the instrument cluster is definitely QNX, because the safety level must at least reach ASIL-B. ADAS also has to basically reach ASIL-B, equivalent to what the medical industry calls a medical device level. Many peers say they are Safety Linux, but it's still Linux; Linux is open, shared by everyone, everyone can contribute, and the understanding of functional safety differs, it won't demand as much as QNX. For Mercedes-Benz safety comes first, so even though it's very hard, they still chose QNX.
— Wang XinBattles in Germany are mild, in the US fierce
Wang Xin has worked at American companies, and the contrast is clear: American companies are more direct, they argue things out on the spot in the meeting, they want to solve many problems in one go, so the process is very fierce and sometimes the atmosphere is hard to control. Battles in German companies are still relatively mild, everyone compromises somewhat, but throughout the cooperation they still want to speak with data and facts — everyone is an engineer, put the data and facts in front of them and communication is more efficient. He also mentions a disagreement over UI design: an interface Chinese colleagues were used to, taken to Germany, had German colleagues saying they couldn't find the function they wanted and there were too many buttons; what Germany gave, Chinese colleagues said took too many steps to find a function and basically nobody would use it.
— Wang XinSwitching from rule-based to end-to-end took seven or eight months of agony
Wang Xin says the transition was quite agonising, because it's a mindset change. At first everyone did rule-based, and actually did it fairly well, but gradually they found end-to-end brings more benefits, it gives you unlimited potential and perception improvement, which rule-based can't. But you have to overturn your previous accumulation, and redeploying to the car shows obvious regression, and that process is agonising — you have to watch it slowly get better through daily training accumulation, until one day it's suddenly much better than rule-based. This transition was led by the China team, mutually encouraged together with partners. The tug-of-war lasted about seven or eight months, and in the end it was settled by massive testing and data, by performance, and they went forward unswervingly.
— Wang XinCLA has neither map nor lidar, except for Tesla
Wang Xin says the CLA is basically the only one with neither map nor lidar, apart from Tesla — not a single lidar, all vision. There are several reasons: first, if training is good enough, adding lidar doesn't really make much sense and just increases cost; second, adding more sensors doesn't necessarily increase credibility, sometimes it adds more noise, and you have to constantly judge whether to trust the data the lidar fills in or the data perception fills in. This decision had battles, but all settled with data and facts — for example, presenting 20,000 km, 30,000 km with and without lidar being the same or even better. But moving toward L3 they haven't given up lidar, because it's a very good redundancy.
— Wang XinA car must run 5 million km before it leaves the line
Wang Xin gives the concrete scale of Mercedes-Benz safety validation: a new car must have at least nearly 5 million km of testing before it leaves the line, a single function must have over 150,000 km of cumulative testing as a mandatory durability test, and it must be tested in extreme cold, extreme heat and extreme humidity in various places, in both China and Europe, passing different laws and regulations. In addition, a lot of simulation testing and real-car testing is done, basically a car is allowed about 150 crash tests, real-car crashes must be done; virtual tests basically 15,000 times must also be done. His logic is: each number may just be a number to us, even 0.1%, 0.2%, but for many people's family it is a family, you can't just see it as a number.
— Wang XinIn their own words · checked verbatim
In the past, T1 delivered a black box to the OEM, and the OEM had no idea what was inside; now the OEM says I want to operate all the data inside myself, so my customers can better enjoy these new features.
从以往的话 T1一个黑盒子交付给主机场 主机场根本不知道里面有什么东西 到现在主机场说 我要自己来运作 operator里面所有的数据 让我的客户更好地能够享受这些新的功能
Wang Xin13:14
It's that we committed, we delivered, and we all built trust — in English, we commit, we deliver, we trust.
是我们承诺了 我们交付了 我们大家产生了信任 英文叫we commit we deliver we trust
Wang Xin27:25
But for us, safety is our number one priority; even though it's very hard, we still chose QNX as our intelligent driver assistance operating system.
但对我们来说的话 安全用的是我们第一位的 即使是非常难 我们还是选择了QX QX作为我们的智能辅助驾驶的操作系统
Wang Xin32:28
American companies are more direct; they basically argue things out on the spot in the meeting.
美国企业比较直接 他基本上就是 就事论事把很多事情在会场上面 去把它battle清楚
Wang Xin34:29
It was actually quite agonising, I think, because it's a mindset change.
其实挺动物的 我觉得因为这是一个mindset的一个change
Wang Xin1:01:59
Each number may just be a number to us, even 0.1%, 0.1%, 0.2%, but for many people's family it is a family, so you can't just see it as a number, you have to imagine it as a living family.
这每个数字对我们来说有可能只是一个数字 甚至来说0.1%0.1%0.2 但它对于很多人的一个家庭来说 那就是一个家庭 所以你不能够只是把它看成一个数字 你要把它想象到它是一个鲜活的家庭
Wang Xin1:10:04
Figures
| Number of Mercedes-Benz China R&D engineers | About 2,000+ in Beijing and Shanghai combined | 28:26 |
| Change in size of Mercedes-Benz Shanghai Digital Technology Center team | Zero in 2021, now several hundred people | 24:22 |
| RD3.0 campaign timeline | Started September 2022, ran to September 2023, new car launched in mass production December 2023 | 23:22 |
| Parking success rate improvement | From about 70% to over 95% | 20:22 |
| Front-view camera resolution change | From about 2.5 megapixels to 8 megapixels | 37:31 |
| Duration of end-to-end switch tug-of-war | About seven or eight months | 1:09:03 |
| Test mileage before a new car leaves the line | Nearly 5 million km | 1:11:06 |
| Durability test mileage for a single function | Over 150,000 km | 1:11:06 |
| Number of real-car crash tests | About 150 | 1:12:08 |
| Number of virtual tests | About 15,000 | 1:12:08 |
Glossary
- Tier 1
- A component system supplier that supplies directly to the OEM, such as Delphi or Visteon.
- QNX
- A closed real-time operating system with a high safety level, often used for instrument clusters and ADAS.
- Rule-based
- An ADAS approach that uses hand-coded rules for planning and decision-making; long-tail scenarios are hard to enumerate.
- End-to-end
- A large-model approach that goes directly from perception to planning and control, replacing rule-based algorithms.
- SOA
- Service-oriented architecture: breaking vehicle functions into callable services to make software iteration easier.
- V2X
- Vehicle-to-vehicle and vehicle-to-road communication, still at an early stage.
How to listen
Founders and engineers watching automotive intelligence, OEM-supplier relations, and the R&D voice of foreign carmakers in China — especially those working on ADAS, cockpit or a Tier 1 transition.
From 1:14:12 to 1:15:12 the transcript is garbled; you can skip it.