DeepRoute.ai Bets on Physical AI, Using Foundation Models to Chase 1,000-km MPCI

Introduction: An Unconventional Player in the Smart Driving Arena

Since 2024, China's autonomous driving industry has been swept up in a massive "city expansion war," with manufacturers competing to cover as many cities as possible with urban NOA capabilities. Small models and lightweight solutions have become the industry's dominant narrative. Yet amid this seemingly white-hot competition, DeepRoute.ai has chosen a decidedly different path — eschewing small models and city expansion battles in favor of betting on a Physical AI foundation model, with its sights set squarely on achieving 1,000-kilometer MPCI (Miles Per Critical Intervention).

This strategy may appear to "go against the tide," but it conceals a deeper contemplation of autonomous driving's endgame.

The Core: Technical Ambitions of a Physical AI Foundation Model

MPCI (Miles Per Critical Intervention) is a key metric for measuring autonomous driving system reliability, representing the average distance a system can drive autonomously before requiring human critical intervention. A 1,000-kilometer MPCI means the vehicle only needs human intervention once every 1,000 kilometers on average — a benchmark that far exceeds the current industry mainstream and approaches the safety threshold of L4-level autonomous driving.

DeepRoute.ai's core weapon is its self-developed Physical AI foundation model. Unlike the popular "small model plus rules" approach prevalent in the industry, the Physical AI foundation model emphasizes deep understanding and reasoning about the real physical world. It is not merely a perception model or a decision-making model but an attempt to build a unified, end-to-end world model that enables the system to understand causal relationships in road scenarios the way one understands the laws of physics.

According to available information, DeepRoute.ai's Physical AI foundation model integrates multiple core capabilities including multimodal perception, spatiotemporal reasoning, and physics simulation. The system can make physics-level predictions about complex traffic scenarios — for example, estimating the braking distance of a vehicle in an adjacent lane on a wet road surface, or predicting the potential trajectory of a pedestrian whose line of sight is obstructed. This kind of physics-based reasoning capability is precisely what traditional rule engines and small model approaches struggle to match.

Analysis: Why Skip the City Expansion War?

The logic behind DeepRoute.ai's decision to "fight a different battle" is not difficult to understand.

First, the city expansion war is essentially a war of resource attrition. Opening each new city requires extensive HD map collection, localization adaptation, and road testing verification. For resource-constrained startups, spreading efforts across dozens of cities often means compromising on core technology depth. DeepRoute.ai clearly believes that rather than competing on breadth against well-funded industry giants, it is better to build barriers through technological depth.

Second, while small model solutions offer low deployment costs and fast inference speeds, their capability ceiling is relatively limited. Small models excel at handling common scenarios but often fall short when facing long-tail corner cases. The real challenge of autonomous driving lies precisely in those extreme scenarios that account for less than 1% of situations yet determine system safety. The advantage of a Physical AI foundation model is that through deep understanding of the physical world, it possesses stronger generalization and scenario transfer capabilities, theoretically enabling better handling of unknown situations.

Third, from a business logic perspective, MPCI is the core metric that determines the commercialization trajectory of autonomous driving. Whether for Robotaxi operations or the user experience of advanced driver assistance systems, everything ultimately hinges on system reliability and safety. If 1,000-kilometer MPCI can be achieved, it would represent a qualitative leap in the safety of DeepRoute.ai's solution — far more persuasive than "how many cities are covered."

Notably, DeepRoute.ai's strategic direction also aligns with global AI development trends. NVIDIA CEO Jensen Huang has repeatedly emphasized that "Physical AI" will become the next frontier of AI, and Tesla's FSD is also evolving toward an end-to-end large model approach. DeepRoute.ai's Physical AI foundation model strategy can be seen as the Chinese autonomous driving industry's proactive response to this global technological trend.

Challenges and Risks

Of course, DeepRoute.ai's chosen path is not without risks. Foundation model R&D demands enormous investment, with requirements for computing power, data, and talent far exceeding those of small model approaches. In today's increasingly cautious capital market, securing sustained and sufficient resource support is a practical concern.

Moreover, while the 1,000-kilometer MPCI target is exciting, the verification cycle is lengthy. With competitors already building brand awareness and market share through city expansion, DeepRoute.ai needs to find suitable commercialization pathways before its technology fully matures, avoiding the trap of "leading in technology but lagging in the market."

Outlook: Physical AI May Redefine the Smart Driving Endgame

DeepRoute.ai's choice ultimately points to a fundamental question in the autonomous driving industry: What is the true endgame of smart driving competition?

If the endgame is about who covers more cities first, then the city expansion war is undoubtedly the right strategy. But if the endgame is about whose system is safer, more reliable, and more capable of generalization, then a Physical AI foundation model may be the correct path to the finish line.

From a broader perspective, Physical AI is becoming the critical bridge connecting digital intelligence to the physical world. Autonomous driving is just one of the most representative application scenarios for Physical AI. Once the underlying technological paradigm matures, it could extend to robotics, industrial automation, and other far-reaching domains.

DeepRoute.ai's choice to fight a different battle with Physical AI is not merely a bet on the technical benchmark of 1,000-kilometer MPCI — it is a strategic judgment on the future direction of autonomous driving and Physical AI as a whole. The outcome of this bold wager will likely be revealed within the next two to three years. Regardless of success or failure, the courage to break free from industry-wide involution and pursue the essence of technology deserves the attention and respect of the entire industry.