Huawei Qiankun hits 2 million units: Should engineering focus on interfaces or mileage?

2 million looks more like a delivery scale threshold.

The synchronized reach of 2 million units equipped with Huawei Qiankun Intelligent Driving and HarmonyOS Cockpit, along with 2 million users of the Huawei Qiankun App, indicates it has shifted from a few showcase models to a universal component across brands, vehicle types, and powertrains. Media Day provided a metric: Qiankun has partnered with over 25 brands and over 50 vehicle models, covering sedans, SUVs, MPVs, off-roaders, BEVs, hybrids, and ICE vehicles. For automakers, this combination is business; for engineering teams, it's a matrix hell.

In the short term, the most direct impact of 2 million is on procurement decisions. When choosing suppliers, automakers used to ask about model capabilities, computing power, sensors, and urban coverage. Now they ask a simpler question: Can you cover us if something goes wrong? 50 vehicle models mean 50 sets of configuration differences: different screen sizes, different cockpit chips, different chassis domain controllers, different steering and braking responses, and different regulatory regions. An intelligent driving version cannot be rolled out OTA to all devices at once like a mobile app. Have tests been written? How large is the regression matrix? What is the gray release ratio? What is the rollback plan? These will be worth more than spec sheets.

I've been obsessing over test coverage for two months and used data cleaning tools for three weeks. The feeling is direct: no matter how strong the model, dirty data, incorrect fields, and inconsistent version naming will drag down subsequent data flows. Huawei's naming seems neat externally—Qiankun Intelligent Driving, HarmonyOS Cockpit, Qiankun Vehicle Cloud—but internally, version naming, vehicle adaptation tags, sensor configuration fingerprints, feature flags, and accident data indices are the real headaches. Poor function naming turns troubleshooting into archaeology later.

Long term, the change brought by 2 million is that data begins to exhibit scale effects. If cumulative mileage reaches the tens of billions of kilometers level, road testing alone is insufficient. The core is whether the cloud can slice real-world scenarios into trainable, evaluable, and traceable samples. Here, distinguish between training and inference. I've said before that separating training and inference is easily disrupted by reinforcement learning; merging them into the same pool is more efficient. On cars, this pool consists of vehicle-side logs, shadow mode, simulation evaluation, and OTA rollbacks. Whoever smooths out this process can turn one accident into a safety use case for the next version.

However, more data doesn't equal cleaner data. In intelligent driving scenarios, the hardest part is the long tail: construction barriers, temporary speed limits, nighttime EVs, roadside door openings, animals crossing. These samples are few, high-risk, expensive to label, and hard to validate. Without clear evaluation criteria, it's like code without unit tests: it looks runnable, but exposes flaws only on the road. When I built an AI dating product, I complained about this point: emotional products lack testing mechanisms. If intelligent driving products also lack executable safety tests, the experience will be extremely fragile.

Products like Cybercab without steering wheels also make me view Huawei differently. I've only used it for a few days, so I'm not a veteran user, but it reminds me that the more aggressive the hardware, the slower software validation must be. Huawei Qiankun entered mass-production cars first, using existing steering wheels, brakes, and redundant backups, rather than building extreme forms first and waiting for the system to catch up. This strategy is engineering-wise more stable. Short term, it won't get the flashiest demos; long term, it accumulates explainable, appealable, and auditable data.

I view 2 million as a shift from feature competition to infrastructure competition. Automakers are unlikely to maintain full intelligent driving teams for each brand in the future. It's more realistic to use full-stack or component solutions, saving energy for cockpits, services, and channels. Among Huawei's five major solutions, names like Vehicle Cloud, Digital Chassis Engine, and Optical Interconnect don't sound like hit features, but they likely determine whether it can continue expanding installations. Because the more cars there are, the more unified interfaces, unified logs, and unified security boundaries are needed.

In the next two to three years, Huawei Qiankun's moat will stem more from turning the 2 million scale into a manageable delivery system. As L3 pilots expand, regulators, insurers, and recalls will ask about versions, data, and liability boundaries. Engineering teams looking at interfaces versus mileage ultimately have to land on whether vehicle adaptation, OTA gray releases, scenario libraries, and accident tracing can be managed. Without this system, the larger the installation volume, the greater the risk.