Has the turning point for autonomous driving really arrived?
```
Author | Wang Xiaojuan
Editor | Zhou Zhiyu
On March 23rd, China's autonomous driving industry saw two major news events on the same day.
On this day, Qingzhou Zhihang announced the completion of a new round of Series D financing, raising $100 million. The lineup of investors is quite striking—including a leading domestic OEM, a top automotive electronics components company, as well as several industry funds.
This company, which started with L4 level autonomous driving and later established a foothold in the L2++ mass production market, has explicitly directed the funds towards R&D in world models and reinforcement learning. Its CEO, Yu Qian, said: "Autonomous driving is the best entry point to physical AI."
On the same day, XPeng Motors announced the establishment of a Robotaxi business unit.
According to Wallstreetcn, this first-level organization will coordinate product definition, project integration, R&D testing, and operation work of Robotaxi. The person responsible will be Yuan Tingting, Senior Director of Product at XPeng's Autonomous Driving Center. Empowered by the second-generation VLA (Vision-Language-Action) large model, XPeng plans to start passenger demonstration operations in the second half of this year and roll out three Robotaxi models in 2026.
At the recent Nvidia GTC conference and various industry summits, CEO Jensen Huang repeatedly said: "The next wave of artificial intelligence is physical AI. AI will understand the laws of the physical world, and autonomous vehicles are currently the largest and most mature embodied intelligent robots we can see."
It's not just Qingzhou and XPeng.
In early March, WeRide and Geely Remote deepened their cooperation, planning to deliver 2,000 pre-install mass-produced Robotaxi vehicles in 2026; Pony.ai partnered with Toyota and GAC Toyota, aiming to deploy thousand-unit fleets in major Chinese cities. GAC's Ruqi Mobility completed deliveries at the hundred-unit level, doubling fleet size to 600 vehicles; Caocao Mobility has set up over 3,600 Robotaxi virtual pick-up and drop-off points in Hangzhou.
After two years of quiet, the autonomous driving track is collectively moving in the spring of 2026.
Three Bottlenecks Loosened Simultaneously
Two years ago, the scene of L4-level autonomous driving was completely different.
Around 2024, this track was still burning money and telling stories. The solution stack of lidar, high-performance chips, and high-precision maps drove single vehicle retrofitting costs to hundreds of thousands of yuan, and Robotaxi demonstration operations in select cities were more like money-burning shows, far from true commercialization.
The capital market's patience for technological storytelling was exhausted, and a batch of L4 startups reached the brink of life and death. The core industry logic shifted from running faster to surviving longer.
Recently, autonomous driving practitioners told Wallstreetcn that since 2025, changes in technology, hardware costs, and policy have pushed the industry to a new node.
First, the technology path has converged.
End-to-end large models have become industry consensus, and the L4 technological route has been redefined. Tesla’s CyberCab went offline in February 2026, proving that pure vision solutions and end-to-end large models can support the vehicle to operate for a long period and distance without human intervention. XPeng’s second-generation VLA model also achieves direct end-to-end generation from visual signals to motion commands.
More importantly, Nvidia released the world’s first open-source autonomous VLA model, Alpamayo, with thinking and reasoning capabilities at CES, and simultaneously opened a high-fidelity simulation framework, AlpaSim, and a large-scale driving dataset, building an open ecosystem of "Model-Simulation-Data," greatly reducing the R&D threshold for advanced autonomous driving.
Second, the establishment of the pre-install mass production route.
Compared to early retrofitted Robotaxi, the industry now generally turns to the pre-install mass production route. WeRide’s GXR is equipped with the latest GEN8 autonomous driving kit, relying on Geely Remote’s AI-controlled chassis, supply chain, and production management system. The unit off-line pace has been dramatically reduced from 1 hour to less than 10 minutes.
Pony.ai’s seventh-generation Robotaxi autonomous driving kit costs 70% less than the previous generation. The vehicle computing unit cost dropped by 80%, the lidar cost decreased by 68%, and the selected vehicle's price fell to the 100,000 to 150,000 yuan range.
At this cost level, the single-car profitability model of Robotaxi can start to work.
Third is the substantial breakthrough in policy bottlenecks.
In December 2025, the Ministry of Industry and Information Technology announced the first batch of L3-level autonomous vehicle license approvals. BAIC ArcFox Alpha S and Changan Deep Blue SL03 models started road test pilots in designated areas in Beijing and Chongqing. By mid-January 2026, Deep Blue L3 vehicles had accumulated over 70,000 km of autonomous driving mileage.
More importantly, the pilot clarified the car company's primary responsibility during system activation, solving the long-standing issue of liability division in the industry. Though there are still differences between L3 and L4 legally, this signal has opened the door for L4 autonomous driving technology commercialization.
Academician Ouyang Minggao of the Chinese Academy of Sciences predicted at the 2026 High-level Forum on Intelligent Electric Vehicle Development: By 2030, L4-level autonomous driving based on advanced end-to-end large models will achieve mass commercialization in mid-to-high-end passenger vehicles.
Physical AI Is No Longer Just a PPT
"In 2026, humanity is at a critical watershed in AI development. In the next 5 to 10 years, even greater opportunities lie in the physical world." Qingzhou Zhihang CEO Yu Qian,whenannouncingthe financing,associatedthe development of autonomous driving closely with AI development.
Physical AI has become the hottest new story in the industry.
The reason why this concept is attractive is that it elevates autonomous driving from a vertical track to a general entry point for AI into the physical world.
Last year, XPeng upgraded its brand positioning to 'Explorer of Mobility in the Physical AI World' at its Tech Day. Its VLA model can drive four carriers: cars, Robotaxi, humanoid robots, and flying cars. Nvidia’s Alpamayo model released at CES also featured the banner of physical AI.
But in 2026, what’s different is that the narrative is rapidly turning into quantifiable business indicators.
From a technical perspective, the computing power race hasn't disappeared but entered a new stage. XPeng Robotaxi plans to equip four Turing AI chips, delivering up to 3,000 TOPS of onboard computing power, and using a pure vision solution without relying on lidar or high-precision maps.
Nvidia launched the Alpamayo platform, trying to package "brain + skull" for car makers, lowering the deployment threshold for advanced autonomous driving. End-to-end latency is reduced by 50%, traffic efficiency increases by 20%, and hard-braking rate drops by 30%—these become hard benchmarks for algorithm capability.
Production scale is also expanding.
WeRide plans to deliver 2,000 pre-install mass-produced Robotaxi vehicles in 2026, Pony.ai aims to expand its fleet to 3,000 vehicles. Qingzhou Zhihang's "Chengfeng" intelligent assisted driving models have exceeded 1 million units, with nearly 10 head car makers as partners. In 2026, the number of new cooperating models is expected to surpass 50.
The premise of scale is controllable cost. Some institutions predict that if annual output reaches 100,000 vehicles, the manufacturing cost per Robotaxi can drop to $10,000.
From a business operation perspective, Robotaxi can finally calculate its accounts.
Waymo's paid rides reach 450,000 times per week, and its operational scope has expanded to Houston, Miami, as well as Tokyo, London and other international markets. Domestically, Qingzhou has officially entered the unmanned logistics vehicle track with industrial partners, operating in Jinhua, Wuhu, and Ningbo, pioneering the model of production equals operation. It is projected that China's Robotaxi market size willgrowto $8.655 billion by 2033, with an annual compound growth rate of 74.0% from 2025 to 2033.
Deutsche Bank judged in a report after CES: "2026 will be the year autonomous vehicles transition from testing/validation to scale, and humanoid robots shift from lab experiments to small-scale deployments."
This round of recovery is not justtechnology-driven, but propelled by the synchronous evolution of technology, policy, cost, and user cognition.
Users no longer blindly believe in increasingly inflated technical specs, but pursue smart driving features that are useful, trustworthy, and affordable. City NOA penetration rate reached 15.1% from January to November 2025 and is accelerating its extension to models under 200,000 yuan. The L3 pilot accumulated over 70,000 km of autonomous driving mileage in 19 days, covering interchanges, congested roads, and other complex urban scenarios.
Of course, the industry is not without worries. Industry insiders told Wallstreetcn that policy and legislative pace remain uncertain. Systemic safety events may trigger regulatory tightening, and whether cost reduction speed can match the intensity of price wars remains unknown.
The survival space for third-party smart driving solution providers is being squeezed by head car makers’ full-stack self-development, and the threshold for "technical moats" has risen rather than fallen.
In 2026, the autonomous driving industry no longer relies on stories for financing, but begins to rely on numbers. Physical AI has moved from exhibition stands to city streets, and L3 from documents to daily commutes. After a decade-long marathon, the turning point may really have arrived.
Risk Warning and Disclaimer ClauseThe market has risks; investment needs caution. This article does not constitute personal investment advice, nor has it taken into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinion, viewpoint, or conclusion in this article suits their circumstances. Invest accordingly at your own risk. ```