Momenta CEO: L4 Autonomy Needs $10B, Cash Flow Is the Real Ticket to Physical AI

April 2026
physical AIArchive: April 2026
Momenta CEO Cao Xudong has dropped a bombshell on the autonomous driving industry: scaling L4 autonomy requires a staggering $10 billion, and massive data accounts for only 10% of the solution. He argues that cash-flow-generating businesses are the real ticket to the physical AI era, challenging the industry's data-driven orthodoxy and pushing for a survival-of-the-fittest model based on self-sustaining revenue.

In a recent internal presentation, Momenta CEO Cao Xudong laid out a stark vision for the autonomous driving industry: achieving large-scale L4 deployment demands a capital commitment of up to $10 billion, with the value of massive data sets contributing merely 10% to the overall solution. This directly challenges the prevailing industry narrative that data scale is the primary moat. Instead, Cao argues that the remaining 90% must come from breakthroughs in algorithm architecture, system integration, simulation validation, and—most critically—a sustainable business model.

Cao's central thesis is that cash-flow-positive operations—such as production ADAS (Advanced Driver-Assistance Systems) contracts, robotaxi fleet services, or logistics automation—are the non-negotiable entry ticket to the physical AI era. Without a self-sustaining revenue engine, no company can afford the decade-long, multi-billion-dollar R&D cycle required for true L4 autonomy. This is a direct rebuke to the 'burn cash for data' approach that has characterized many autonomous driving startups.

The implications are profound. Cao's logic suggests that the next phase of competition will not be about who collects the most miles of driving data, but who can architect the most efficient learning systems—leveraging world models, video generation, and reinforcement learning—to generalize from limited data. The winners will be those who can monetize their technology incrementally through ADAS while funneling profits into the long-term L4 moonshot. AINews analysis confirms that this represents a fundamental shift from a technology-first to a business-first mindset, with cash flow as the ultimate gatekeeper.

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.

Related topics

physical AI22 related articles

Archive

April 20262543 published articles

Further Reading

World Model on a Chip: How 500 TOPS Rewrites the Rules of Autonomous DrivingQcraft has become the first autonomous driving company to officially enter the physical AI arena, compressing a world moHow a Table Tennis Robot's Victory Signals Embodied AI's Leap into Dynamic Physical InteractionA table tennis robot has decisively defeated a human professional player, an achievement far more significant than a spoDexWorldModel's Rise Signals AI's Pivot from Virtual Prediction to Physical ControlA leaderboard change in a world model benchmark is signaling a tectonic shift in AI priorities. Crossdim AI's DexWorldMoGoogle's Embodied AI Breakthrough Gives Robots Spatial Common SenseA new class of AI models is bridging the gap between digital intelligence and physical action. By endowing robots with s

常见问题

这次公司发布“Momenta CEO: L4 Autonomy Needs $10B, Cash Flow Is the Real Ticket to Physical AI”主要讲了什么?

In a recent internal presentation, Momenta CEO Cao Xudong laid out a stark vision for the autonomous driving industry: achieving large-scale L4 deployment demands a capital commitm…

从“Momenta L4 autonomous driving cash flow strategy”看,这家公司的这次发布为什么值得关注?

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 p…

围绕“Cao Xudong 10% data claim analysis”,这次发布可能带来哪些后续影响?

后续通常要继续观察用户增长、产品渗透率、生态合作、竞品应对以及资本市场和开发者社区的反馈。