Summary of Key Points
The Chinese autonomous driving industry appears promising in 2026 (with an L2 level penetration rate of over 70% and L3 systems starting to be deployed on roads), but it faces three core issues: lack of profitability in business operations, stagnation in technological innovation, and severe data constraints. These problems have led to a significant flow of high-end talent and capital shifting from autonomous driving to the field of embodied intelligence, reflecting the cyclical transition of emerging industries from rapid expansion to in-depth adjustment.
Detailed Analysis
1. Business Dilemmas: Autonomous Driving Development as a “Never-Ending Hole”
The autonomous driving industry may seem vibrant, but it is not profitable. The issue lies in the severe mismatch between costs and revenues:
- Costs: Research and development are extremely costly—hiring doctors with annual salaries in the millions, renting expensive computing clusters on an hourly basis, and continuing to invest in software updates after the vehicles are sold (with no additional revenue from existing users).
- Revenues: Users are reluctant to pay for autonomous driving software—costing thousands to tens of thousands of yuan, many consider it unnecessary, and sales personnel complain that payment habits have not yet been established.
After mass production, the algorithm team transforms from a center of innovation to a center of cost. What used to be about creating new things has now become about maintaining existing code and processing data, which is why some say there is no room for further progress.
2. Technological Innovation: On Hold, with Gaps in Critical Abilities
While China has made rapid progress in terms of mass production, breakthroughs in core technologies are slow:
- User Behavior Modeling: Tesla’s FSD can learn users’ parking preferences, but most domestic solutions have not kept up. This is not due to technical limitations; rather, the initial design did not account for these factors.
- Environmental Prediction: Tesla can generate virtual extreme scenarios (such as heavy rain or bumpy roads) from real data, addressing the scarcity of high-quality data. In contrast, domestic companies still rely on 3D Gaussian reconstruction, which falls far short of Tesla’s advanced simulation capabilities.
The reason? After mass production, companies focus on making cost-cutting patches rather than fundamental architectural innovations, lacking the motivation for breakthroughs.
3. Data Challenges: Extreme Scenarios as a Rare Chance, with Limited Influence for Data Teams
Autonomous driving relies on “algorithms + data,” but high-quality data is scarce:
- Scarcity of Extreme Scenarios: There is plenty of data from regular roads, but extreme conditions (such as heavy rain or snow, or unusual vehicles) are rare, making it ineffective to collect enough data.
- Inefficient Data Processing: Out of the terabytes of data collected, only a small portion is usable for training—each step (data collection, preprocessing, and labeling) is costly.
- Low Status of Data Teams: The industry often views algorithms as the core, and data workers are seen as doing repetitive tasks, with limited influence. As Zhou Yiran puts it, “Data workers feel like they are teaching AI to drive, but the company treats them as mere assistants.”
4. Talent Migration: A Shift to Embodied Intelligence for the Next Decade
The departure of autonomous driving talent is not accidental; it reflects a shift in industry focus:
- Clear Capital Signals: Financing for embodied intelligence surged in 2025-2026, with YuShu Technology’s market value exceeding 400 billion yuan in the first half of 2026. The market is investing heavily in what it sees as the “next decade.”
- Skill Alignment: The AI algorithms and data processing skills required for embodied intelligence overlap significantly with those in autonomous driving. For example, half of Chen Moxin’s company’s employees previously worked in autonomous driving, and even basic data labelers have switched to working on humanoid robot data collection.
- Industry Expectations: Autonomous driving is in the “business validation phase,” while embodied intelligence is still in the “technology exploration phase.” Talents believe there is more potential in the robotics field.
5. The Hardship of Basic Data Labelers: Strict Standards and Poor Tools
The pressures in the autonomous driving industry also affect the lowest levels of the supply chain:
- Strict Labeling Standards: Motion trajectories must be smooth, and pedestrian outlines must be precise (e.g., using squares of specific sizes). This limits the labelers’ ability to make nuanced judgments.
- Faulty Tools: Labeling platforms often have bugs, and data from the previous frame remains during frame segmentation. Sensors’ blind spots require manual estimation of targets.
- Ineffective AI Assistance: AI-based pre-labeling tools are either slow to load or prone to errors, making manual labeling more efficient.
Conclusion
The autonomous driving industry has moved from a period of rapid growth to a phase of adjustment, with embodied intelligence taking its place as the new focus. However, it’s important to remember that every emerging industry goes through a similar journey from laboratory development to mass production and commercial validation. Embodied intelligence may also face similar challenges in the future. The real test will only begin when the initial hype driven by capital fades.