Can Tesla’s self-driving AI survive China’s real-world chaos?
It learned to drive in America. Now it must unlearn—and relearn—everything to survive the streets of China.
Updated: May 23, 2025 10:26 PM ESTTesla Model 3 vehicle drives on the street in Shanghai, China.
Dai Aochen/VCG via Getty Images
Ni Tao is IE’s columnist, giving exclusive insight into China’s technology and engineering ecosystem. His Inside China column explores the issues that shape discussions and understanding about Chinese innovation, providing fresh perspectives not found elsewhere.
In June last year, news that Tesla’s Full Self-Driving (FSD) system was authorized to undergo road testing in China jolted the domestic intelligent driving industry.
At the time, many observers predicted that FSD’s full entry into the Chinese market was not far off.
On February 25 this year, amid growing public anticipation, Tesla quietly began pushing FSD to select models in China—albeit in limited form.
It did not roll out the full-fledged autonomous driving system available in the United States. Instead, Chinese users received a scaled-back FSD “intelligent assisted driving feature.”
Compared to its American counterpart, the Chinese version lacks core functions such as automatic parking and smart summon and still requires full driver supervision.
Moreover, the update only supports Model 3 and Y vehicles produced after February 2024 with the HW4.0 hardware and newer Model S and X editions.
‘A generation ahead’
Even so, FSD’s arrival triggered a wave of excitement—and disruption. Daily rental prices for FSD-equipped used Teslas soared to RMB 3,000 ($416) on some platforms.Whether grudgingly or not, domestic automakers and their suppliers publicly welcomed comparisons between FSD and local intelligent driving systems. Many believed the move would push China’s homegrown players to accelerate technological upgrades.
But beyond the fanfare, FSD’s reputation in China went through a roller-coaster ride.
The interior of a Tesla Model 3. Unsplash
Initially criticized for struggling with local road conditions and frequent traffic violations, the system soon won praise from some industry insiders who saw it as “a generation ahead” of domestic rivals.
At the 2025 China EV100 Forum in March, Wu Yongqiao, President of Bosch Intelligent Driving China, said he was deeply impressed after personally testing FSD.
While he noted that domestic leaders were relatively close in capabilities, he stunned the audience by saying Tesla’s system had a “1-2 year lead” over them.
Veteran industry figures like Zhou Guang, CEO of autonomous driving startup DeepRoute.ai, aired similar views. In its end-to-end structure, Zhou called FSD’s V13 “a full generation ahead” of Chinese advanced self-driving systems.
So, what exactly led to this perception shift?
The pain of localization
As is widely known already, Tesla’s FSD hinges on a camera-based vision-only approach and an end-to-end neural network architecture.In contrast, most Chinese companies rely on multi-sensor fusion perception, combining LiDAR and high-definition maps for an extra layer of redundancy safety. This approach is seen in offerings like Huawei’s ADS, Li Auto’s VLA, and XPeng’s XNGP.
Tesla’s FSD, for its part, ditches LiDAR to imitate human driving. Trained solely on camera data collected from North American roads, FSD delivers driving behavior closely mimicking human decisions on highways and basic urban routes.
However, when this US-trained system was first introduced in China, it quickly ran into trouble. The issues were threefold.
First, failure to recognize local driving scenarios. A test by Beijing-based auto blogger Chen Zhen showed FSD committed seven violations in just two hours, including misidentifying lanes, veering into bike lanes, making illegal lane changes, and going straight in turn-only lanes.
Second, there is a weakness in local traffic negotiation. In dealing with China’s infamously rampant “cutting-in” behavior and jaywalking, FSD reacted conservatively. Tech influencers reported their Teslas coming to a complete halt at the sight of milling pedestrians, refusing to move an inch further.
Its inability to negotiate right of way in pedestrian-heavy scenarios often required frequent manual takeovers—far more than local systems like Huawei’s ADS 3.0.
The interior of a Tesla car. Unsplash
Third, bottlenecks in training data. Due to cross-border data transfer restrictions, Tesla currently trains FSD in China using publicly available road video and simulators instead of real-world data collected during driving.
This significantly limits its understanding of Chinese road signs and intersection rules. Worse still, many “corner cases” are potentially left unresolved. Unsurprisingly, skepticism grew. Critics claimed FSD lagged far behind Huawei and XPeng in actual performance.
But just weeks later, sentiment began to shift. Testers started noticing that FSD was learning quickly, though not always in positive ways. Its neural network began adapting to Chinese driving habits.
In videos shared online, FSD, which was made to follow the rules strictly, eventually learned to mimic reckless local driving tactics, such as squeezing into dense traffic or executing borderline illegal lane changes.
As more real-world tests circulated on social media, FSD’s ability to quickly adapt gained traction. Analysts now expect that with enough on-road mileage, its understanding of China’s complex traffic could improve dramatically.
The compute power conundrum
But even as optimism grows, FSD faces several challenges.The core issue remains data training. While Tesla’s Shanghai data center has passed compliance reviews, it pales in comparison to the power of the US-based Cortex supercomputer cluster that Tesla uses to perfect its algorithms.
The latter has roughly 100,000 sets of Nvidia H100 and H200, versus only thousands of such chips or Chinese-made variants in leased domestic solutions. This hardware gap could slow FSD’s learning curve.
A passenger presses on the touchscreen of a Tesla car for navigation in a photo shot in a Hong Kong tunnel. Unsplash
Efforts to tackle the computing power shortage are underway. Tesla has built local data labeling and operations teams and sent engineers to China to offer training.
The company is also exploring options to build computing infrastructure in China, but that means heavy duplicate investment, and it’s unclear whether domestic computing power can meet Tesla’s needs.
FSD’s algorithms, shaped by American data, reflect an understanding of a relatively orderly traffic environment. China’s roads, by contrast, are often a chaotic mix of e-bikes, pedestrians, rule-bending drivers, and local quirks.
Some industry veterans thus argue that local smart driving systems, fine-tuned to these conditions since their inception, remain better suited, at least for now.
However, the biggest roadblocks for FSD may not be technical.
Hefty price for ‘democratized’ tech
Tesla’s asking price for FSD in China has sparked controversy. It costs RMB 64,000 ($8.883), a 10 percent markup over the US version. That’s more than double Huawei’s ADS advanced package, priced around RMB 30,000 ($4,164), and far above XPeng’s XPILOT, which often comes free with the vehicle.The price reflects Tesla’s broader ambition: boosting software revenue. In 2024, software accounted for over 20 percent of Tesla’s global income.
But in China, consumer willingness to pay for software remains limited. Surveys suggest most are only comfortable paying RMB 12,000–20,000 ($1,666-2,776) for assisted driving features.
The steering wheel and interior of a Tesla Model Y vehicle. Unsplash
Tesla’s subscription model hasn’t caught on either, as many Chinese consumers prefer investing in hardware upgrades over OTA software add-ons.
Meanwhile, Chinese rivals are turning up the heat on Tesla by “democratizing smart tech for all.”
BYD, for instance, has brought LiDAR costs down to just over RMB 1,000 ($139) through scale and made L2+ features standard across its RMB 100,000 ($13,879) models, something competitors struggle to match.
Though Tesla competes in a higher price bracket, alongside Huawei-backed Aito and Li Auto, China’s downward pricing trend may eventually disrupt its business model for FSD.
2025 was expected to be the breakthrough year for L2+ to L3 autonomous driving. However, high-profile accidents, like the fatal Xiaomi SU7 crash in March of this year, caused partly by its cheesy self-driving tech, have triggered a regulatory rethink.
A rose by another name may not smell as sweet
Authorities now require that all intelligent driving products be labeled “advanced driver assistance,” not “autonomous driving,” to avoid misleading consumers.Tesla has complied, rebranding its Chinese FSD as “intelligent assisted driving” in local translations. But a change in terminology carries risks. If FSD cannot deliver truly autonomous driving, the RMB 64,000 price tag could provoke backlash.
Tesla’s updated name did just that. In mid-March, 15 Chinese Tesla owners jointly petitioned the national regulators, accusing Tesla of false advertising and consumer fraud over its FSD promotion.
Naming also carries legal weight. In China, drivers are held liable for accidents involving L2+ systems, whereas responsibility shifts to automakers once vehicles reach L3 autonomy and beyond.
With legal gray areas still unresolved, Tesla’s “driver liability” waiver clauses come off as an attempt to sidestep accountability, especially when local competitors are offering driver insurance for intelligent systems.
In the end, Tesla’s journey in China will depend not just on its technology, but on how well it navigates regulation, public trust, local data laws, and the fiercely competitive market landscape.
Only when all these roadblocks are cleared will FSD’s journey in China truly become a smooth ride.
ABOUT THE AUTHOR
Ni Tao Ni Tao worked with state-owned Chinese media for over a decade before he decided to quit and venture down the rabbit hole of mass communication and part-time teaching. Toward the end of his stint as a journalist, he developed a keen interest in China's booming tech ecosystem. Since then, he has been an avid follower of news from sectors like robotics, AI, autonomous driving, intelligent hardware, and eVTOL. When he's not writing, you can expect him to be on his beloved Yanagisawa saxophones, trying to play some jazz riffs, often in vain and occasionally against the protests of an angry neighbor.L | A
