Technical Deep Dive
Cao's assertion that 'massive data accounts for only 10%' is a direct challenge to the data-centric paradigm that has dominated autonomous driving since the rise of deep learning. The conventional wisdom, championed by players like Waymo and Cruise, has been that more data—billions of miles of real-world driving—is the key to solving long-tail edge cases. But Cao argues that raw data volume suffers from diminishing returns. The real bottleneck is not data quantity but data quality and algorithmic efficiency.
The 90% Solution: Beyond Data
The remaining 90% of the challenge, according to Cao, lies in four interconnected domains:
1. Algorithm Architecture: The shift from modular pipelines (perception → prediction → planning) to end-to-end neural networks. This includes the adoption of transformer-based architectures, bird's-eye-view (BEV) representations, and occupancy networks. A key open-source reference is the UniAD repository (github.com/OpenDriveLab/UniAD), which implements a unified end-to-end autonomous driving framework. UniAD has garnered over 4,000 stars and demonstrates how planning can be integrated directly into the perception stack, reducing error propagation.
2. System Integration: The ability to seamlessly combine sensor fusion (camera, LiDAR, radar), vehicle control, and fail-safe mechanisms. This is where the 'last 10%' of engineering effort often consumes 90% of the time.
3. Simulation & Validation: The use of neural rendering and generative AI to create photorealistic, controllable simulation environments. This is critical for testing rare events without real-world risk. Repos like NuPlan (github.com/motional/nuplan-devkit) and CARLA (github.com/carla-simulator/carla) are foundational, but newer approaches using diffusion models for scenario generation are emerging.
4. Business Model: The ability to generate cash flow from intermediate products (e.g., L2+ ADAS) to fund the long L4 R&D cycle.
Data Efficiency vs. Data Scale
Cao's 10% figure implies that a company with 1/10th the data but 10x better algorithms could outperform a data-rich competitor. This aligns with recent research on data efficiency. For example, Tesla's FSD V12, which uses a pure end-to-end neural network, reportedly achieved performance parity with V11 using a fraction of the training data by leveraging a more efficient architecture.
| Approach | Data Required (Relative) | Algorithm Complexity | Generalization to Edge Cases | Cost to Scale |
|---|---|---|---|---|
| Data-Scale Dominant (e.g., Waymo) | 100x | Low-Medium | High (with massive coverage) | Extremely High |
| Algorithm-Efficient (e.g., Momenta thesis) | 1x | High | Potentially Higher (via simulation) | Lower (if algorithm works) |
| Hybrid (e.g., Tesla) | 10x | High | Medium-High | Medium |
Data Takeaway: The table illustrates the core trade-off. The data-scale approach has proven effective but is financially unsustainable for most players. The algorithm-efficient approach, if realized, offers a path to L4 with a fraction of the capital, but it requires breakthroughs in world models and simulation that are not yet proven at scale.
Key Players & Case Studies
Cao's thesis is not just theoretical; it reflects Momenta's own strategy and the broader industry dynamics.
Momenta's Dual-Engine Strategy
Momenta has long advocated a 'two-engine' approach: a 'data-driven' engine for L4 R&D and a 'revenue-driven' engine for production ADAS. The company has secured production deals with major OEMs like SAIC, Mercedes-Benz, and Toyota, providing a steady cash flow. This directly embodies Cao's argument: use ADAS revenue to fund the L4 moonshot. Momenta's 'Mpilot' product line for L2+ highway and urban navigation is already generating revenue, while the 'MSD' (Momenta Self-Driving) stack targets L4 robotaxis.
Competing Strategies
| Company | Primary Strategy | Cash Flow Status | L4 Timeline (Est.) | Key Advantage |
|---|---|---|---|---|
| Waymo | Data-scale, full-stack L4 | Negative (Alphabet-funded) | Operational in select cities | First-mover, safety record |
| Cruise | Data-scale, full-stack L4 | Negative (GM-funded, paused) | Uncertain | Vertical integration |
| Tesla | Algorithm-efficient, vision-only | Positive (vehicle sales + FSD) | Ambitious but delayed | Data from fleet, cost structure |
| Momenta | Hybrid: ADAS cash flow + L4 R&D | Positive (from OEM contracts) | 2025-2027 (target) | Revenue sustainability, OEM partnerships |
| Pony.ai | Full-stack L4 with robotaxi ops | Negative (venture-funded) | Operational in China | Government partnerships |
Data Takeaway: The table reveals a clear divide. Companies with negative cash flow (Waymo, Cruise, Pony.ai) are in a race against time, dependent on parent company or investor patience. Tesla and Momenta, with positive cash flow from other businesses, have more runway. Cao's argument suggests that the latter group has a structural advantage in the long L4 marathon.
Industry Impact & Market Dynamics
Cao's statements arrive at a critical inflection point for the autonomous driving industry. After a decade of hype and hundreds of billions in investment, the market is facing a 'reality check.'
The Funding Cliff
According to industry data, global autonomous driving startup funding peaked in 2021 at over $12 billion and has since declined sharply, with 2024 projected to be below $5 billion. The era of easy money is over. This makes Cao's emphasis on cash flow not just strategic but existential.
| Year | Global Autonomous Driving Startup Funding ($B) | Number of Deals | Average Deal Size ($M) |
|---|---|---|---|
| 2021 | 12.4 | 85 | 146 |
| 2022 | 8.1 | 62 | 131 |
| 2023 | 5.8 | 48 | 121 |
| 2024 (Est.) | 4.5 | 35 | 129 |
Data Takeaway: The funding decline is accelerating. Startups that cannot demonstrate a path to revenue will find it increasingly difficult to raise capital. Cao's 'cash flow as ticket' argument is a direct response to this new reality.
The Rise of the 'ADAS Bridge'
Cao's strategy is being replicated across the industry. Companies like Huawei (with its ADS system integrated into AITO and Avatr vehicles) and Mobileye (with its SuperVision product) are generating significant revenue from L2+ systems while developing L4 capabilities. This 'ADAS bridge' model is becoming the dominant paradigm, especially in China, where OEMs are eager to differentiate their vehicles with advanced driver assistance features.
Risks, Limitations & Open Questions
While Cao's argument is compelling, it is not without risks and open questions.
1. The '10%' Claim is an Oversimplification: While data alone is insufficient, it remains a critical moat. Companies like Waymo have spent billions to collect data in specific geographies, creating a high-fidelity map and behavior model that is extremely hard to replicate. Cao's 10% figure may underestimate the defensive value of proprietary data.
2. Algorithm Breakthroughs are Unpredictable: The bet that algorithm efficiency will outpace data scale is a high-risk gamble. If the next breakthrough in world models requires even more data (e.g., video generation models like Sora require internet-scale data), the 'algorithm-efficient' path may prove illusory.
3. Cash Flow vs. R&D Investment: There is a tension between generating cash flow (which requires selling L2+ products) and investing in L4 R&D (which requires long-term, high-risk spending). Companies may be tempted to prioritize short-term revenue over long-term innovation, leading to a 'local maximum' where they get stuck at L2+.
4. The $10 Billion Figure: Is it accurate? Cao's estimate likely includes cumulative R&D, deployment infrastructure (fleet, maintenance, remote operations), and regulatory compliance. For context, Waymo has reportedly spent over $5 billion to date and is not yet profitable. The $10 billion figure may be conservative for a global L4 service.
AINews Verdict & Predictions
Cao Xudong's intervention is a watershed moment for the autonomous driving industry. He has articulated what many insiders have felt but few have dared to say publicly: the data-centric, cash-burning model is broken. The future belongs to companies that can generate revenue from their technology today while investing in the breakthroughs of tomorrow.
Our Predictions:
1. The 'ADAS Bridge' will become the default strategy for all autonomous driving companies not backed by a trillion-dollar parent. Within 18 months, every major L4 startup will announce a production ADAS deal or pivot to a service-based model.
2. A wave of consolidation is imminent. Companies that cannot demonstrate a path to positive cash flow within 24 months will be acquired for their technology or talent at distressed valuations. We expect to see 3-5 major acquisitions in the next 12 months.
3. The next technical breakthrough will come from simulation and world models, not from real-world data collection. Startups that master neural rendering and generative scenario creation will have a decisive advantage. Watch for open-source projects like UniAD and NuPlan to become industry standards.
4. China will lead the 'cash flow + L4' model. With its massive EV market, aggressive OEMs, and government support for autonomous driving, Chinese companies like Momenta, Huawei, and Baidu are best positioned to execute this strategy. The West, with its reliance on venture capital, will face a more difficult transition.
Cao's message is clear: the era of 'move fast and break things' is over. In physical AI, you must 'move sustainably and build things that pay for themselves.' The companies that heed this lesson will be the ones that survive to see the L4 future.