The Technology Vision
No Maps. No Shortcuts. Just AI.
The detail that separates DeepRoute.ai from most of its competitors is deceptively simple: the system works without HD maps. That matters enormously in practice. HD maps require constant maintenance, geographic coverage, and update cycles that break down in construction zones, at night, and at the edges of where anyone has bothered to survey. DeepRoute IO navigates city streets through pure AI inference - perception, decision, and control fused end-to-end into a model that learns from every kilometer driven.
Zhou's critique of the previous generation of ADAS systems is sharp and specific. "Traditional small-model autonomous driving suffers from a 'seesaw effect,'" he explained in early 2026. "Performance varies wildly across scenarios." Fix one edge case and another breaks. The only solution, he argues, is the same move that transformed language AI: foundation models. A 40-billion-parameter Vision-Language-Action model that reasons across driving, analysis, and evaluation simultaneously.
The Seesaw Problem and the Foundation Fix
The seesaw metaphor is precise. Small, scenario-specific models require constant retuning - optimizing one behavior degrades another. Every model update becomes a balancing act. Zhou pushed DeepRoute.ai toward a unified foundation model that avoids the tradeoff by learning a richer representation of the world in the first place. The same shift that made GPT-4 better than a thousand fine-tuned chatbots is what he believes will make DeepRoute IO better than a thousand hand-coded driving rules.
The infrastructure reads: DeepRoute IO supports compute solutions ranging from 100 to 1,000+ TOPS, meaning the same software stack can sit on commodity hardware or on Nvidia's Thor chip - which DeepRoute.ai was among the first Chinese companies to receive. The system also integrates with Nvidia DRIVE Orin and previously DRIVE Hyperion, giving it a clear technology partnership lineage rather than a go-it-alone chip strategy. Zhou is explicit about why: "From an economic standpoint, self-developing chips is simply untenable."
What "AI-Native" Actually Means
Zhou draws a hard line between companies that use AI and companies that are AI-native. His distinction goes past tooling. "AI capability is closely linked to organizational structure, culture, talent, and fundamental cognitive frameworks." He has recruited what he describes as an AI-native team - not engineers who learned AI, but people for whom AI is the default lens. The most visible proof: hiring Chong Ruan, the former Head of R&D at DeepSeek, as Chief Scientist. A multimodal AI researcher leading the technical direction of an autonomous driving company is not the conventional move.
The operational evidence is in the numbers. DeepRoute.ai's data closed-loop - the cycle from collecting a driving scenario to updating the model based on it - dropped from over five days to roughly 12 hours. That's not an incremental improvement. It's a different category of learning speed. Every car on the road becomes a faster feedback loop. At 300,000 vehicles, the compound effect is significant.
The Physical World Bet
Zhou's long-horizon ambition is not limited to cars. He told Bloomberg in March 2025 that DeepRoute.ai is already expanding into robots and other physical platforms - testing whether the same foundation model principles that work in an urban driving scenario can generalize to other real-world movement problems. His timeline for "a lot of general robots" is five years. For the full era of robots: a decade. He's not hedging: "It will definitely happen."
The robotaxi piece fits into this. Commercial operations planned for Wuxi and Shenzhen in 2026 represent DeepRoute.ai's return to the L4 ambition Zhou started with in 2016 - now running on a foundation of scale, data, and operational experience that didn't exist when Roadstar.ai tried first. The circle closes at a very different altitude.
