At the recent 2026 WICV (World Intelligent Connected Vehicle Conference), there was a "Key Technology Forum" for intelligent vehicles. During the one-day forum, three of the speakers were from intelligent driving assistance companies.
Han Xu, CEO of WeRide, used the CNY 100,000-level Guangqi Aian i60 as an example, stating that the ADAS of this vehicle only uses about 200 TOPS of computing power.
Zhou Guang, CEO of DeepRoute.ai, presented a bar chart showing DeepRoute.ai's urban NOA installations reached 62,100 units in August, ranking second in the industry for the month. He also showed a slide of an experiment where the general large model GPT-6 Astra directly controlled a real vehicle to complete a route.
Jiang Jingfang, President of neueHCT, talked about running a one-stage end-to-end model on a 128 TOPS Journey 6M chip, and plans to deploy their products in over 30 countries from 2026 to 2027.
All three companies started with "Physical AI," but they referred to very different things. WeRide talked about the long-term operation of L4 autonomous vehicles in over a dozen countries; DeepRoute.ai talked about a 40B driving foundation model trained using large model methodologies; neueHCT talked about a lightweight one-stage solution that can be sold to dozens of OEMs. Stripping away the promotional rhetoric, the judgment of this article is: with urban NOA having penetrated the CNY 100,000-level market and the penetration rate approaching 28% in August, the three companies are actually answering the same question: where will the next round of differentiation come from. WeRide is betting on data sources, DeepRoute.ai on model scale, and neueHCT on replication speed.
This article is based on these three speeches to analyze and summarize the development trends of the ADAS industry, and then look at what each company said and why they said it. Hope to provide some information and inspiration for the industry.
Five Industry Signals Revealed by the Three Speeches
The Ruler for Evaluating Intelligent Driving is Changing
Over the past two years, chip computing power has almost been synonymous with the tier of intelligent driving. This time, all three companies bypassed this ruler.
Han Xu proposed "intelligence generated per unit of computing power," borrowing the concept of battery energy density and calling it "intelligence density."
Zhou Guang used MPCI (Miles Per Critical Intervention, how many kilometers before the driver needs to take over a critical hand) and "generalization time" (how many days it takes to drive well in a new country).
Jiang Jingfang talked about achieving 200 to 500 TOPS-level performance with 128 TOPS, and latency below 200 ms.
Actually, there are two reasons for the change in the ruler.
First, as urban NOA moves towards the CNY 100,000-level, vehicles cannot afford 500 TOPS, so suppliers can only prove they do better with limited computing power.
Second, as we shared in our previous article "Where Exactly Has China's ADAS (Intelligent Driving) Developed to Now?", consumers are also starting to evaluate the experience using intervention frequency, and the persuasiveness of computing power parameters is declining.
However, these three new rulers, due to time and level constraints at this conference, remained at the conceptual level: WeRide did not provide a calculation method for "intelligence density," DeepRoute.ai's MPCI curve was a schematic without their own measured values, and neueHCT's "comparable to 200 to 500 TOPS" did not say compared to whom or in which scenarios.
However, the overall revelation is that the industry is looking for indicators to replace TOPS. Whoever first turns intervention miles and generalization time into auditable metrics will hold the discourse power for the next round.
The Main Battlefield for Urban NOA Falls in the 100 to 200 TOPS Range
The three companies speaking this time actually have their main solutions in this range:
WeRide uses about 200 TOPS on the Aian i60, and Han Xu verbally mentioned 20,000 orders.
neueHCT's Astra was simplified from dual Journey 6M to a single chip with 128 TOPS. Its product gradient is drawn very straightforwardly: the entry-level Luna uses Journey 6B for integrated driving and parking, and the affordable Astra does urban NOA. Only the Astra Max above 500 TOPS does parking space to parking space, and the Astra Ultra above 600 TOPS is reserved for integrated cockpit and driving and L3-level experience. In August, national urban NOA installations were about 427,000 units, with a penetration rate of 27.9%.
After the volume picks up, OEMs' requirements for solutions shift from "whether the function exists" to "how much more each vehicle costs." The high computing power tier mostly undertakes the reserve for flagship models and L3.
Urban NOA has become a mass-produced component competing on cost. The computing power tier that CNY 100,000-level vehicles can afford will determine who gets most of the installations in the next two years.
The Data Battle Shifts from "How Much" to "What is Missing"
Han Xu explained this point most clearly: when L2 encounters dangerous scenarios, the driver has usually already taken over, so how the system handles it on its own is missing in the data; L4 autonomous vehicles have no one to take over, so scenarios like ghost probing and accident detour exits can leave a complete process of recognition, decision-making, and execution. Therefore, WeRide uses the L4 fleet to supplement the long tail of L2.
DeepRoute.ai starts from intervention data, treating interventions as signals to locate shortcomings, and uses RSI (Recursive Self-Improvement) to let AI participate in the data ratio for post-training.
neueHCT talks about a localized closed loop of local collection, local compliant processing, and local iteration.
The cost structures of these three paths are also different.
L4 fleets require real money to maintain and operate vehicles; long-tail samples are expensive but complete.
Intervention data comes from sold mass-produced vehicles; it is cheap but requires strong mining capabilities.
Overseas localized data is constrained by regulations, often requiring local teams and data centers. Which path to choose depends on what the company already has.
Mass-produced mileage is no longer scarce; what is scarce are complete handling samples, intervention signals that can locate shortcomings, and compliant available overseas data. The assessment of data teams should also change from "how much was collected" to "what kind of gaps were filled."
After End-to-End, the Routes Begin to Diverge
One-stage end-to-end puts perception, planning, and vehicle control into one network. Sensor data and navigation instructions go in, and trajectories come out. As we shared in our previous article "The Battle of Intelligent Driving Shaping: A Comprehensive Look at the Underlying Logic and Architecture Evolution of Autonomous Driving 'End-to-End'," end-to-end has become the mainstream for mass production (at least in slogans).
How the end-to-end model moves forward saw a divergence on stage. Zhou Guang believes that current end-to-end is a "small model," with training data of about 10 million 10-second clips, equivalent to 20,000 to 30,000 hours, only one-thousandth of a generative large model, and without large-scale pre-training. In his view, the MPCI upper limit of such models is in the tens to hundreds of kilometers, and to go further requires foundation models trained according to the large model paradigm.
neueHCT and WeRide maintain the scale of the vehicle-end model, superimposing multimodal, reinforcement learning, and physical priors on the one-stage model to make full use of every bit of computing power.
The two routes are not completely opposed.
A 40B model cannot fit into a 128 or 200 TOPS vehicle-end chip. As we shared in our previous article "2026 Bainianhui Forum: DeepRoute.ai | The End of the Small Model Era, Zhou Guang Reconstructs Autonomous Driving with a 40B Foundation Model," DeepRoute.ai will go with cloud training plus distillation to the vehicle end.
The vehicle-end efficiency camp also needs stronger cloud models for data mining and evaluation.
The difference lies in where the money is spent: one side bets on cloud pre-training and talent, and the other side bets on vehicle-end engineering and mass-production adaptation.
The focus of competition is shifting from the vehicle-end model structure to the cloud training system. Companies without large-scale training clusters and large model background teams will find it hard to maintain a gap through small structural improvements.
The Cost of Going Global Becomes the Cost of Generalization
All three companies have put overseas expansion in their main lines.
Zhou Guang's numbers are the most specific: in September 2025, generalization in Berlin took 4 months; in October 2026, it only took about 4 days in Paris.
WeRide has business in over 60 cities in 13 countries. In the UAE, over 100 vehicles are doing fully driverless commercial operations. What it supplements is the operation layer: from having safety drivers to driverless in the car, and then to remote assistance and scaled operations, every step requires local regulatory approval.
neueHCT's global headquarters is in Singapore, with an R&D center in Germany. The four things Jiang Jingfang listed are basically an overseas expansion cost checklist: generalization training for global roads, EU GSR II, E-NCAP, and GDPR, driving habits like left-hand/right-hand drive, European roundabouts, and Southeast Asian motorcycles, and local compliant data closed loops.
How many days, how many people, and how much compliance investment it takes to enter a new country will be directly written into the quotation for overseas OEMs.
General Large Models Begin to Touch the Steering Wheel
Zhou Guang said he is "getting more and more anxious." He cited DrivingBench, which lets general models directly control real vehicles through tool interfaces. GPT-6 Astra completed the whole route on the second attempt, while the completion rate of other models was between 6% and 45%. On another page, the robot simulation leaderboard RoboDojo, GPT-6 Astra, which has not undergone specific robot training, ranked first. He worries that in a few years, the companies that achieve autonomous driving will be companies like OpenAI, Anthropic, and DeepSeek. I agree with this; the future of autonomous driving companies doing large computing power models will face a huge impact if these companies strike down a dimension.
Calculated from the screenshot, this time 134.7 meters took 5 minutes and 22 seconds, about 1.5 km per hour, costing USD 7.74, equivalent to about USD 57 per kilometer, which is still far from being usable. The significance of this slide is that the most difficult "scenario understanding" is being gnawed at by general models from the other side.
The long-term barriers for intelligent driving companies must fall on real fleet data, vehicle-grade engineering, and safety responsibilities, which general models cannot take away.
WeRide: Feeding L2 with L4 Long Tail
Han Xu's main line is "L4 capabilities entering L2 mass-produced vehicles." He first showed videos of Aian i60 commuting and extreme evasion, then mentioned "intelligence per unit of computing power," and finally used 3,600 L4 fleet vehicles, 100 million km of mileage, and fully driverless operations in the UAE to explain where the capabilities come from.
He also made two points: data flows in both directions, and the data and model progress accumulated by L2 mass production will also feed back into L4; AI hallucinations might amuse people in the digital world, but they are fatal in the physical world, so safety investment should be seen as a highly cost-effective brand building.
There are realistic reasons for saying this. As we shared in our previous article "WeRide's 2026 Q2: Earning Money from L2, Profitability is in Sight," WeRide started with L4, and its revenue scale is still small, with Q2 revenue at CNY 230 million. L2 mass production is the most realistic volume business, and this track is very crowded. "L4 long-tail data" is exactly the reason it is different from others. Benchmark models can only be equipped with about 200 TOPS, so it makes sense to turn hardware constraints into efficiency advantages.
How much data can be migrated? The sensors and computing power of a Robotaxi are very different from a CNY 100,000 SUV, and no numbers were given for the migration effect.
Intelligence density has no algorithm, only a concept, and cannot be compared horizontally.
Fleet numbers: looking at the point-in-time financial report caliber, as of July 31, it was about 3,400 vehicles; the forum said 3,600 units. The 20,000 orders for i60 only appeared verbally.
WeRide needs to provide comparative numbers of intervention rates before and after L4 data migration; otherwise, "L4 same source" is just a brand label.
DeepRoute.ai: Betting the Next Generation Paradigm is a Foundation Model
Zhou Guang barely talked about product features; he talked about paradigms: from rules and modularity to end-to-end, and then to foundation models deeply researched since 2026. The threshold is data, talent, computing power, and funds. DeepRoute.ai's progress is a 90% improvement in the decision-making capability of the 40B foundation model, a 10x improvement in data closed-loop efficiency, and the establishment of Superfluid Lab, with the chief scientist coming from DeepSeek. On the mass production side, IO 4.0 adds spatial capabilities like reverse escape and multi-point U-turns.
Why say this? Look at two sets of numbers. DeepRoute.ai's installations from January to August totaled about 314,000 units, and August accounted for about 14.5% of the national total. Ranking second in a single month, with Huawei ahead. The functions of small model end-to-end are converging, and Zhou Guang himself said that it is not very meaningful to continue to compete on "you are long, I am short."
Betting on the next generation paradigm is a common choice for the second place to change the ranking, and three of the four thresholds require money. This statement is also a shout to the capital and talent markets. Zhou Guang mentioned two details: the first thing the chief scientist said after joining was "the talent density is not enough," which is far lower in the autonomous driving field than in large model companies; the scale of 1,000 cards for end-to-end is not enough to support the breakthrough of the intelligence upper limit.
These two sentences are both industry judgments and explanations of why DeepRoute.ai needs to continue financing and why it needs to organize teams in the way of large model companies.
Mass-produced vehicles haven't used it yet. Zhou Guang explicitly said IO 4.0 is not fully driven by the foundation model; the native version will wait until next year.
90% has no baseline; it didn't say relative to which model or using what metrics.
MPCI is a schematic showing three levels of tens of kilometers, hundreds of kilometers, and ten thousand kilometers as directions; the graph does not have DeepRoute.ai's own measured values.
How it runs on the vehicle end was not mentioned. How the 40B model falls on the vehicle-end chip determines cost and mass production rhythm, which was not explained in the whole session.
Whether the intervention mileage of next year's native foundation model version can pull apart by an order of magnitude is the only standard to test DeepRoute.ai's judgment.
neueHCT: Turning Intelligent Driving into Replicable Goods
neueHCT is a joint venture of Horizon Robotics and AUMOVIO (formerly the automotive electronics business of Continental AG), and changed its name from "Zhijia Dalu" in April this year.
Jiang Jingfang provided a gradient of algorithm same source: Journey 6B for integrated driving and parking, a single Journey 6M for 128 TOPS universal urban NOA, and above that is parking space to parking space and integrated cockpit and driving. The same set of capabilities also extends to two-wheeler riding assistance and L4 logistics vehicles. Her judgment is that the watershed of the next generation of competition is faster cross-scenario deployment capability.
This is related to its shareholders and customers. Horizon Robotics provides chips and teams, AUMOVIO provides global supply chain, channels, and quality systems; customers include Honda, Toyota, Nissan, and Hyundai. They need a solution that can pass regulations and be mass-produced in China and overseas simultaneously.
The company positions itself as a "software T2 or software T1," listing Desay SV Automotive, Hyundai Mobis, and other Tier 1s as partners. What it sells is standardized software and reference designs that can be integrated by others. For such a company, replication speed is scale, and scale is gross profit, so it values the number of designated vehicle models and covered countries the most.
The AUMOVIO relationship also brings hardware resources. The slide explicitly states it has secured the first launch cooperation for AUMOVIO's 6th generation millimeter-wave radar. In the eyes of overseas customers, a Chinese software company with a German supply chain and quality system is easier to enter the procurement list than a pure local supplier.
Performance statements lack baselines. "Comparable to 200 to 500 TOPS," "100% behavioral consistency," and "over 90% reduction in inching and weaving" have no comparison objects and test conditions.
Designation does not equal mass production. Over 16 OEMs and over 70 vehicle models are designation numbers. Jiang Jingfang said this year and next year are the concentrated mass production and delivery period; how much lands will depend on subsequent insurance data.
Caliber needs to be aligned. The verbal "16 countries" should be 16 OEMs; overseas vehicle coverage verbally said over 50 countries, the slide wrote 2026 to 2027 expected over 30, long-term over 50.
The risk for neueHCT is not in the technical upper limit, but in whether over 70 designated models can be delivered on time and with quality in multiple countries simultaneously.
Conclusion
Looking at the three speeches together, intelligent driving assistance has crossed the chasm and entered the stage of rapid mass production and popularization. Urban NOA has changed from a technical issue to a cost and replication issue. Intelligent driving assistance is no longer a halo function, but has become a robust chassis safety feature like AEB/ESP. At this moment, its products, positioning, and market methods have all changed.
Practitioners may need a new way of thinking and operation in strategy, products, and market.
Reference materials and pictures
WICV 2026 DeepRoute.ai speech and PPT;
WICV 2026 WeRide speech and PPT;
WICV 2026 neueHCT speech and PPT.