Nearly a month after the high-profile accident concerning the Xiaomi SU7, the Chinese government has introduced new restrictions on both the deployment and promotion of smart driving systems. According to participants in a recent regulatory meeting, several key directives emerged: automakers are no longer allowed to promote Level 2 systems using the term “takeover,” functions like valet parking that remove full driver control will no longer be accepted, and public road tests are now banned.
In response, on May 4, Xiaomi revised the wording on its SU7 product page, changing “smart driving” to “assisted driving.”
The accident came at a pivotal moment. Just before the crash, several Chinese automakers (including Xiaomi) had completed major equity financing rounds, gearing up to push what they called “smart driving for all.”
In March, Xiaomi and BYD each raised HKD 42.5 billion (USD 5.4 billion) and HKD 43.5 billion (USD 5.6 billion) through share placements, while Nio announced plans to issue 140 million Class A shares to raise HKD 4.03 billion (USD 520 million). Xiaomi and BYD timed their offerings at stock price highs, while Nio opted to raise funds even as its shares were at historical lows.
The market reaction was swift. BYD shares dropped more than 7%, Xiaomi lost nearly HKD 100 billion (USD 12.8 billion) in market cap in a single day, and Nio’s stock also fell more than 8% at one point. Xiaomi’s sharp decline dragged down the Hang Seng Tech Index, which posted a one-day loss of 3.82%.
Smart driving has long been a major cost center for automakers. Data from Gasgoo shows that in 2023, smart driving R&D accounted for 38% of total company expenditures, second only to battery development.
Consider the latest equity deals: BYD plans to spend up to RMB 100 billion (USD 14 billion) on intelligent systems, while Nio will invest RMB 46 billion (USD 6.4 billion) in the same area by the end of 2024. Xiaomi CEO Lei Jun has said the company spends over RMB 2 billion (USD 280 million) annually on smart driving.
To offset these expenses, automakers had begun installing smart driving systems in low- and mid-end models to achieve scale. But just as they ramped up, regulators stepped in, and the capital-intensive nature of smart driving has suddenly turned into a liability.
The price of cutting costs
According to algorithm expert Fu Cong, full-system costs for current smart driving solutions vary from four-figure RMB sums to over RMB 20,000 (USD 2,800), accounting for 5–15% of total vehicle cost based on model and configuration. The addition of advanced features like LiDAR (light detection and ranging) and redundant control only pushes that figure higher.
Automakers like Nio, Li Auto, and Xpeng Motors often take pride in in-house development, but Fu notes that in practice, most rely on a hybrid model of internal R&D and external procurement.
For example, Li Auto uses its proprietary AD Max system for premium models but turns to external vendors for entry- and mid-level models. Similarly, BYD sources lower-end versions of its DiPilot system externally.
This mixed approach allows companies to maintain control over proprietary technology while accelerating deployment with mature third-party solutions. However, external procurement can drive up costs in specific areas. For instance, externally sourced combined navigation systems can cost more than RMB 10,000 (USD 1,400), whereas in-house systems may only cost one-tenth of that figure.
Nio’s management revealed during an earnings call that switching from four Orin chips to its in-house Shenji NX9031 chip could save about RMB 10,000 per vehicle.
LiDAR remains a major hardware cost. Despite prices falling to as low as RMB 1,000 (USD 140), LiDAR stacks still weigh heavily on balance sheets in a price-sensitive market. To cope, carmakers are pursuing differentiated hardware strategies. Base models now lean on cheaper, camera-only systems, while higher trims adopt fusion solutions that combine vision with LiDAR.
Take Xiaomi for example. The standard SU7 uses Xiaomi’s Pilot Pro system with a pure vision approach. Its Pro, Max, and Ultra trims come equipped with the more advanced Pilot Max variant, which adds LiDAR to the sensor suite.
Even Nio, which traditionally champions fusing vision and LiDAR, has released a pure vision version under its new brand, Onvo.
These shifts underscore the mounting pressure to reduce costs. But cheaper hardware doesn’t tell the whole story—pure vision systems come with significant hidden costs.
Pure vision relies on massive data modeling.
“Algorithms must constantly adapt to new cities, scenarios, and regulations. Training models demand expensive servers, computational power, and human talent,” Fu said. “Labeling data alone can cost tens of millions, or even over RMB 100 million for leading firms. Because autonomous systems require high-quality data, especially for multimodal fusion and 3D semantic mapping, each data point might cost hundreds or even thousands of RMB.”
According to Fu, data labeling and algorithm development are the hardest parts to economize in vision-based systems.
No matter the path taken, smart driving continues to burn through capital. And for now, the bet that it will be the differentiator hasn’t paid off.
Xiaomi sold around 136,900 vehicles in 2024 but recorded a loss of RMB 6.2 billion (USD 870 million), marking an average loss of RMB 45,000 (USD 6,300) per car. Nio fared worse, with per-vehicle losses reaching RMB 100,000 (USD 14,000).
Waiting for scale
Years of cost-cutting in hardware have yielded modest returns. Now, scale appears to be the only path to real savings.
Earlier this year, BYD launched a campaign to promote the mainstreaming of smart driving, expanding what used to be confined to high-end vehicles into every segment.
In February, BYD said it would roll out its DiPilot advanced driving system across its Dynasty and Ocean series, standard on all models priced above RMB 100,000 (USD 14,000) and included on many that are priced below that figure. From the RMB 78,800 (USD 11,000) Seagull to the RMB 249,800 (USD 34,972) Song L, the first batch of 21 models came equipped with high-level features, at no extra charge.
This move broke the pricing logic of the industry, where smart driving had mostly appeared in vehicles priced above RMB 200,000 (USD 28,000).
Industry observers say BYD’s strategy is reframing how advanced features are distributed, trickling down into lower-end segments faster than below and driving adoption through volume.
The consensus? Achieving scale appears to be the only viable way to bring costs down. Only when these systems are installed en masse can companies amortize R&D expenses. The most critical market lies in the RMB 100,000–200,000 (USD 14,000–28,000) range.
According to estimates by Everbright Securities, the 2024 penetration rate of smart driving features in urban China rated at Level 2 or higher was about 5–6%. Among vehicles priced between RMB 250,000–400,000 (USD 35,000–56,000), penetration exceeds 20%. But in the RMB 100,000–200,000 range, that figure remains below 0.2%.
The gap is driven by both cost and consumer understanding. “Ordinary users still have limited awareness of what smart driving systems can—and cannot—do,” Fu said.
That knowledge gap becomes more dangerous in lower-end models with downgraded systems.
In the Xiaomi crash, the standard edition of the SU7 was equipped with a pure vision system. “It was driving at night through a construction zone,” one industry insider said. “That’s a tough test for vision-only setups.”
Fu added that most smart driving systems seem up to the mark during demos largely because the scenarios presented are already included in training data. Real-world driving is far more complex. Edge cases, which are rare but high-risk scenarios, are often missing from training datasets and hard to solve with conventional methods.
“Think missing or confusing lane markings, cameras obscured by weather, or unique scenarios like traffic cops directing flow. These are extremely difficult for onboard systems to handle consistently,” Fu said.
Once seen as a make-or-break feature for automakers, smart driving now looks less like a crown jewel and more like a growing burden—at least in the short term, as the path to scale stalls.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Geng Chenfei for 36Kr.