China’s smart EV industry sent a clear message at the 2026 Beijing Auto Show: the next competitive battleground is no longer just electrification, but physical AI—the fusion of autonomous driving, world models, vehicle intelligence, and connected infrastructure. At the show on April 25, autonomous driving firms DeepRoute.ai (Yuanrong Qixing) and QCraft laid out their next-generation strategies for scaling assisted driving, while Quectel and UNISOC highlighted how 5G and V2X hardware are becoming foundational to future Chinese EV architectures. Together, these announcements show how China’s EV market is moving from feature-led smart driving to a broader AI-native mobility stack.
Physical AI Becomes the New Industry Theme
The strongest takeaway from Beijing was that “smart” is no longer an optional selling point for Chinese EV brands—it is becoming a baseline capability for brand competitiveness. Across the show floor, the conversation shifted from standalone ADAS functions to a much bigger question: how do cars become intelligent agents that understand and interact with the physical world?
That framing explains why both DeepRoute.ai and QCraft emphasized physical AI rather than simply using familiar labels like L2, L3, or L4.
Why the concept matters
Physical AI in the automotive context generally refers to systems that can:
- Understand complex real-world driving scenes
- Predict motion and behavior in dynamic environments
- Make safe driving decisions in long-tail scenarios
- Learn continuously from large-scale real-world fleet data
- Extend beyond driving into cockpit, user interaction, and embodied intelligence
This is a significant evolution from the earlier generation of rule-heavy or narrowly trained driver-assistance stacks.
DeepRoute.ai: From Urban NOA to an AI Infrastructure Vision
At its Beijing Auto Show event, DeepRoute.ai CEO Zhou Guang framed the company’s long-term goal in unusually ambitious terms: to become part of the AI infrastructure of the physical world, comparable in importance to communications or electricity.
The company’s near-term argument is grounded in autonomous driving progress. Zhou said current urban assisted driving is still far from perfect, with MPI (mean takeover distance) in city environments still only in the tens of kilometers. Even so, he argued that existing autonomous driving systems are already several times safer than purely human driving, and that larger AI models could unlock truly safe autonomous driving within the next two to three years.
That pitch was backed by hard deployment numbers.
DeepRoute.ai key data points
- 300,000+ production vehicles equipped with DeepRoute.ai urban NOA-assisted driving solutions
- 1.3 billion km of cumulative real-road mileage from vehicles equipped with its active safety system over the past year
- 44.8 million hours of user driving companionship time
- Data-loop iteration cycle reduced from about 5 days to 12 hours under its new foundation-model architecture
- 2026 target of 1 million production deliveries for assisted driving systems
- 2026 goal to raise MPCI/MPI-related performance to above 1,000 km
- Target user high-frequency usage rate of more than 50%
These figures matter because scale remains one of the biggest advantages in Chinese autonomous driving development. A large production fleet generates the real-world edge cases needed to improve model robustness, especially in dense urban traffic.
Foundation Models Replace the Small-Model Era
One of the more important technical disclosures came from Ruan Chong, DeepRoute.ai’s chief scientist and former DeepSeek R&D lead, who gave his first public talk in the new role.
Ruan argued that the previous generation of assisted driving development—built around many smaller models—has struggled to deliver stable gains in system reliability and user trust. In difficult long-tail scenarios, performance can still fluctuate too much. DeepRoute.ai’s answer is to build around a foundation model that unifies:
- Driving decision-making
- Scene understanding
- Behavior evaluation
within a single architecture.
This matters because the industry increasingly believes fragmented model stacks are reaching diminishing returns. A larger, more integrated model can potentially improve consistency, accelerate training, and make data reuse more efficient.
DeepRoute.ai’s architectural shift
| Area | Previous approach | New direction |
|---|---|---|
| Model structure | Multiple smaller specialized models | Unified foundation model |
| Iteration cycle | About 5 days | About 12 hours |
| Core challenge | Long-tail instability | Scalable continuous learning |
| Product focus | Assisted driving functions | AI brain for vehicle and user interaction |
DeepRoute.ai also previewed an integrated cabin-driving agent, which it says is not just another voice assistant or infotainment layer. Instead, the system is positioned as an AI brain capable of understanding user intent and proactively responding in complex scenarios.
That signals a broader industry trend: smart driving and smart cockpit are converging into a single compute and software stack.
QCraft: Autonomous Driving Is the Best Entry Point for Physical AI
If DeepRoute.ai focused on scale and architecture, QCraft focused on first principles. In a media interview during the Beijing Auto Show, co-founder, chairman, and CEO Dr. Yu Qian argued that autonomous driving is the most realistic and scalable entry point for physical AI.
His core claim was straightforward: if an AI system cannot handle autonomous driving—a relatively structured but still highly dynamic real-world problem—it will struggle even more with robotics and other embodied AI applications.
According to QCraft, autonomous driving has several advantages over other physical AI fields:
- Large volumes of relatively homogeneous data
- Mature engineering and productization pathways
- Existing deployment at scale in production vehicles
- Real commercial incentives tied to safety, convenience, and efficiency
CTO Li Dong added that while language models benefit from highly standardized tokenized text data, physical-world data is much scarcer and more fragmented. That gives autonomous driving a unique edge because road data is abundant and consistent enough to train world models effectively.
No Open-Source Shortcut, but World Models Are Near Production
QCraft’s comments were especially notable for pushing back on hype. The company said there is no true open-source world model that companies can simply adopt off the shelf for autonomous driving.
Instead, the current reality is a patchwork approach: companies may borrow capabilities from different open-source models for specific functions such as image generation or scenario synthesis, then combine those elements into their own production stack.
More importantly, QCraft said world models in autonomous driving are already moving beyond the lab.
QCraft’s view on world models
- World models in autonomous driving are already in the application stage
- Both cloud-side and vehicle-side world models are being used in production workflows
- Cloud models can generate rare or scarce training data
- Vehicle-side systems remain essential for all safety-critical functions
- In higher-freedom environments, such as more generalized robotics, the technology is still in the research phase
This is a useful distinction for investors and consumers alike. The term “world model” can sound futuristic, but in China’s EV market it is already becoming a practical engineering tool for scaling smart driving systems.
Domestic Chips and Cost-Efficient Performance Are a Major Theme
Another important signal from QCraft was its strong emphasis on Chinese semiconductor localization. Yu said QCraft was the first among L4 autonomous driving firms to fully support domestic automotive chips, especially solutions based on Horizon Robotics’ Journey 6 (J6).
His claim is that QCraft is not chasing raw compute for its own sake, but trying to deliver a “leapfrog” user experience on more efficient hardware.
QCraft compute roadmap and market positioning
| Metric | QCraft commentary |
|---|---|
| Current mainstream urban NOA | 100-200 TOPS becoming mainstream this year |
| Next step | 500+ TOPS expected to become standard next year |
| Claimed target experience | 500+ TOPS solution aiming to deliver 1,000-2,000 TOPS-class experience |
| Price band target | 150,000-200,000 yuan segment |
| Hardware philosophy | Focus on user value, not compute arms race |
That is a notable strategic contrast to some rivals that rely on premium imported silicon or emphasize massive onboard compute as a selling point.
QCraft also downplayed concerns that chip bandwidth is the main bottleneck for L4 deployment, saying technical optimization can work around memory and bandwidth constraints on platforms such as NVIDIA Thor. In its view, the bigger bottleneck remains the underlying driving intelligence.
L2, L3, L4: China’s Smart EV Debate Is Shifting
Both QCraft executives were skeptical about overemphasizing autonomy labels. Their argument was that consumers care more about real-world value—safety, comfort, convenience, and reduced fatigue—than whether a function is branded L2 or L3.
That reflects a broader shift in China’s EV industry. After years of autonomy marketing battles, many companies are now focusing on:
- Takeover rate and intervention frequency
- High-frequency user adoption
- Insurance and safety implications
- Cost of scaling the technology into mass-market vehicles
QCraft’s position was especially blunt: an L3 function limited to narrow scenarios may have less social and commercial value than a highly capable L2+ system used every day by a much larger customer base.
For robotaxis, the company also argued that commercialization must be built on mass-production validation rather than isolated pilot deployments. The key metric, according to Li, is not a headline autonomy level but the operational economics: for example, how many vehicles one remote operator can manage, and whether intervention frequency can fall to a viable level.
Connectivity Moves to the Center of EV Architecture
While software and AI drew most of the spotlight, the Beijing Auto Show also showed how enabling hardware is moving up the value chain. Quectel, together with UNISOC, launched the new AR59xUB automotive-grade 5G communication module series.
On paper, the specifications are solid:
- Built on UNISOC’s automotive-grade A7726 5G platform
- Supports 3GPP Release 16 and upgradable to Release 17
- Peak downlink speed of up to 5.0 Gbps
- Integrated quad-core Arm Cortex-A55 processor
- Compute performance of 22K DMIPS
But the bigger story is strategic rather than purely technical.
Why the Quectel-UNISOC launch matters
First, the module’s four key components—baseband, RF, V2X chip, and memory—all come from the domestic supply chain. That is important at a time when China’s auto industry is pushing harder to localize critical technologies.
Second, the module’s product definition aligns closely with incoming policy and regulatory changes.
Regulation Is Rewriting the Smart EV Supply Chain
Policy developments are helping explain why connectivity modules are suddenly far more strategic.
Two specific developments stand out:
- The 2027 version of C-NCAP will include C-V2X in its evaluation system
- China’s Ministry of Transport has said it aims to complete technical standards for vehicle-road-cloud information interaction by 2027
That means components like 5G/V2X modules are no longer just cost-sensitive commodity parts. They are becoming part of the vehicle’s compliance architecture.
Market shift underway
| Old view | New view |
|---|---|
| Communication module as a low-profile component | Communication module as strategic vehicle infrastructure |
| Performance-driven planning | Regulation-driven and compliance-driven planning |
| Isolated supplier role | Expansion into adjacent value-chain segments |
Quectel’s broader booth display reinforced this point. Beyond communication modules, it also showcased:
- Cabin-connectivity integrated solutions
- Millimeter-wave radar for central computing and satellite architectures
- In-vehicle AI robot concepts
This suggests a classic move seen elsewhere in China’s EV supply chain: component specialists expanding into neighboring system-level opportunities as intelligence and connectivity become central to vehicle design.
Why This Matters Globally
The developments in Beijing are significant well beyond China.
1. China is redefining the smart EV stack
The competitive frontier is shifting from batteries and infotainment toward a full-stack model that combines:
- Urban NOA and advanced ADAS
- Foundation models and world models
- Domestic chips and compute optimization
- 5G, V2X, and connected infrastructure
- Cabin-driving integration under a single AI brain concept
2. Scale is becoming a decisive advantage
DeepRoute.ai’s 300,000-vehicle deployment base and 1.3 billion km of real-road data illustrate how Chinese autonomous driving players are building feedback loops at speeds that many global peers may struggle to match.
3. Mass-market pricing remains central
QCraft’s focus on the 150,000-200,000 yuan segment shows that smart driving leadership in China will not be defined only by premium halo cars. The bigger prize is delivering advanced assisted driving to mainstream consumers.
4. Supply-chain localization is deepening
From domestic autonomous driving chips to fully localized 5G module components, China’s EV ecosystem is steadily reducing dependence on overseas suppliers in critical intelligent-vehicle domains.
5. Regulation and technology are converging
The inclusion of C-V2X in future NCAP testing means smart connectivity is moving from “nice-to-have” to “must-have.” That could accelerate adoption of vehicle-road collaboration technologies and change how OEMs prioritize electronic and electrical architectures.
The Road Ahead
The 2026 Beijing Auto Show made one thing clear: China’s EV sector is entering a new phase in which physical AI may become the organizing principle for the next decade of automotive innovation. DeepRoute.ai is betting on foundation models and large-scale data loops to build a vehicle AI brain. QCraft is betting that autonomous driving is the first true industrial-scale application of physical AI, powered by world models and efficient deployment on domestic chips. And suppliers like Quectel are showing that connectivity and V2X are becoming core strategic enablers rather than background hardware.
The next milestones to watch are straightforward but consequential:
- Whether DeepRoute.ai can push toward its 1 million vehicle production target by 2026
- Whether QCraft and peers can make 500+ TOPS city NOA mainstream in lower price segments
- Whether China’s V2X and vehicle-road-cloud standards reshape OEM architecture planning by 2027
- Whether physical AI can move from promising demos to truly trusted, high-frequency daily use
For global automakers, suppliers, and software firms, the message from Beijing is unmistakable: the Chinese EV race is no longer just about building electric cars. It is about building cars that can perceive, reason, connect, and act in the physical world at scale.
