In 2026, the autonomous driving sector is experiencing an unprecedented wave of transformation. Tesla's 'fully autonomous' Robotaxi zips through the streets of Austin, Waymo's driverless fleet makes its way into Miami Airport, and the Chinese contingent, led by Pony.ai, WeRide, and Mushroom Autonomous, is making significant strides in the global market.

Meanwhile, the capital market has quietly shifted its focus. From the intense wave of listings on the Hong Kong Stock Exchange in 2025 to the IPO rush of industry leaders in 2026, autonomous driving companies are transitioning from the stage of 'technology storytelling' to the capital elimination round of 'proving commercialization worth.' Behind this surge in listings lies a complex interplay of three core technologies: algorithms, models, and data. It's also a life-or-death race for companies to navigate the challenging waters of commercialization.

Underlying Logic: The Bidirectional Empowerment of Technology and Capital

In 2025, the intelligent driving industry chain witnessed a 'listing frenzy.' Nine companies, including Simu Technology and CiDi Intelligent Driving, flocked to list on the Hong Kong Stock Exchange, raising over HK$20 billion. This surge was not a coincidence but the result of a triple convergence of technological breakthroughs, policy advancements, and rational capital.

From a technological standpoint, the core barriers of autonomous driving are shifting from 'hardware accumulation' to 'competition in the soft power of algorithms and models.' The maturity of world models has become a pivotal driver for the implementation of high-level autonomous driving. These models, based on environmental dynamics and multi-agent interaction rules, enable long-term prediction of traffic participants' behavior trajectories (with a prediction window of up to 3-5 seconds) through temporal prediction and causal reasoning, addressing the limitations of traditional rule-driven systems in handling unexpected scenarios.

The mass production application of end-to-end large models has further reshaped the technological landscape of autonomous driving. Momenta's mass-produced end-to-end large model abandons the traditional separate architecture of perception, decision-making, and planning, directly mapping sensor inputs to vehicle control commands. This not only simplifies system complexity but also enhances decision-making efficiency in extreme scenarios. The model has been installed in over 400,000 vehicles and has collaborated with BMW to develop a full-scenario intelligent driving system, validating the commercialization potential of end-to-end technology.

Policy breakthroughs have paved the way for technological implementation. China is expected to approve the implementation of FSD in February 2026, and Shanghai's 'Model Speed Intelligent Driving' action plan explicitly proposes the large-scale application of L4 autonomous driving by 2027. Europe is also accelerating the unification of its regulatory framework. Global policy relaxations provide a compliant basis for the commercial operation of autonomous driving.

The logical shift at the capital level is a core driver of the listing surge. In the past, investors were willing to bet on 'algorithm demonstration videos'; now, the market only recognizes 'commercial closure.' The persistent losses of companies like WeRide and Pony.ai have shifted capital from 'faith-based investment' to 'value return.' Companies with stable orders and data loop capabilities have become the darlings of the capital market. This shift forces autonomous driving companies to transition from 'technological R&D' to a dual-wheel drive of 'technology and operations.'

Technological Core Battle: The Triple Barriers of Models, Algorithms, and Data

In the race for autonomous driving listings, technological strength is the core moat for companies. Models, algorithms, and data are the three main battlegrounds of this technological war.

World Models and VLAs Reshape the Decision-Making Hub of Autonomous Driving

The core challenge of high-level autonomous driving lies in enabling machines to understand complex traffic environments. The combination of world models and VLAs (Vehicle Intelligence Agents) has transformed autonomous driving systems from 'rule-driven' to 'cognition-driven,' enabling them to handle both known and unseen long-tail scenarios. This is the key to implementing L4 autonomous driving.

VLAs, as the vehicle-end carriers of world models, adopt a 'native foundation model + MoE dynamic routing' architecture. They can adaptively call different expert networks based on scenario complexity, optimizing computational power consumption while ensuring decision-making accuracy. Their core lies in achieving end-side integrated reasoning of perception, localization, planning, and control. Through TensorRT INT8 quantization acceleration and heterogeneous computing scheduling, model inference delay is controlled within 20ms, meeting the real-time requirements of autonomous driving. Horizon Robotics' Journey 7 chip, based on the 'Riemann' architecture, achieves a computational density of 200TOPS/W with 6nm technology. It is equipped with a dedicated NPU unit to support Transformer operator acceleration, providing a computational base for the end-side deployment of VLAs and large models, forming a 'algorithm-computational power' collaborative loop.

Horizon Robotics' Journey 7 chip, built on the 'Riemann' architecture, provides powerful computational support for VLAs. Its computational density has increased severalfold, meeting the real-time operational requirements of end-to-end large models. It forms a collaborative empowerment of computational power and algorithms with large models like MogoMind.

Algorithm Iteration: The Paradigm Revolution from Separate to End-to-End

The evolution of autonomous driving algorithms is undergoing a paradigm revolution from separate to end-to-end. The traditional separate architecture divides autonomous driving into five modules: perception, localization, decision-making, planning, and control. Each module is independently optimized, leading to issues of 'information disconnection between modules.'

End-to-end algorithms break down these barriers, adopting a Transformer Encoder-Decoder architecture. They directly map raw multi-sensor data (image pixels, LiDAR point clouds, IMU inertial data) to vehicle control commands (steering angle, throttle/brake opening), avoiding the cumulative errors between modules in separate architectures. Their core advantage lies in fitting driving decision functions in complex scenarios through massive data training, improving generalization ability by over 40% compared to traditional algorithms in long-tail scenarios such as rain and snow occlusion and construction lane occupation. Mushroom Autonomous integrates the core cognitive abilities of MogoMind into its self-developed MOGO AutoPilot end-to-end system, optimizing the dynamics control module for bus models. Meanwhile, the BEV perception algorithm projects multi-view images and LiDAR point clouds into a unified 3D space through spatial attention mechanisms and temporal fusion networks, achieving accurate detection and trajectory tracking of targets within a 400m range, addressing occlusion and scale distortion issues in traditional perspective views.

Data Loop Competition: The Right to Speak on Core Assets of Autonomous Driving

In the field of autonomous driving, data is a more critical core asset than hardware. If algorithms are the 'brain' of autonomous driving, then data is the 'fuel' needed for the brain to function. Building an efficient data loop has become the key to differentiating companies.

Mushroom Autonomous's practice is quite representative. Its autonomous driving buses have accumulated 5 million kilometers of travel, serving over 200,000 passengers. Relying on a perception scheme of 'vision-based + solid-state LiDAR,' it has built the world's largest bus multimodal dataset. This scheme uses a 128-line main solid-state LiDAR (point cloud frequency of 10Hz, ranging accuracy of ±2cm) paired with high dynamic range cameras. Data alignment is achieved through spatiotemporal synchronization calibration algorithms (time synchronization error < 1μs), and BevFusion algorithms enhance perception reliability in complex scenarios. In terms of the data loop, edge computing devices enable real-time data screening and difficult case annotation on the vehicle side. A federated learning framework iterates models across scenarios while ensuring data security, significantly shortening the technological iteration cycle.

Companies like Pony.ai and WeRide are also building their own data loops. Pony.ai has accumulated a large amount of urban scenario data from its Robotaxi operations in Shenzhen and Saudi Arabia. WeRide obtains multi-model data through collaborations with automakers. Notably, the value of data lies not only in scale but also in quality. Through data cleaning, annotation, and desensitization, companies can extract effective information from massive data. Mushroom Autonomous's edge computing devices can complete preliminary data screening on the vehicle side, significantly reducing cloud computing pressure.

Track Competition: Differentiated Track IPO Sprints

In the 2026 listing surge, different companies have chosen differentiated tracks. From Robotaxi to autonomous driving buses, from passenger vehicle intelligent driving to commercial vehicle logistics, companies are leveraging their respective technological advantages to sprint towards the capital market.

Robotaxi Sector: Pony.ai and WeRide's Global Breakthrough

Robotaxi is the most imaginative sector in autonomous driving and also the most technically challenging. Pony.ai and WeRide are the Chinese representatives in this sector.

Pony.ai has achieved normalized fully autonomous driving operations in Qianhai, Shenzhen, and has entered the Riyadh market in Saudi Arabia in collaboration with Uber, becoming the first Chinese player in the Middle East. Its L4 autonomous driving system, built on BEV perception algorithms and end-to-end decision-making models, can handle complex urban road scenarios. In terms of capital, Pony.ai has completed multiple rounds of financing, and its listing valuation is highly anticipated by the market.

WeRide, on the other hand, has adopted a 'dual-track' strategy. On one hand, its new generation of Robotaxi plans to enter the European and American markets in 2026, competing head-on with Tesla and Waymo. On the other hand, WeRide has deeply cultivated the autonomous driving bus sector, launching Robobus projects in multiple cities. Despite cumulative losses exceeding RMB 6.5 billion, WeRide's global layout and scalable operation capabilities remain its core selling points for attracting capital.

Autonomous Driving Bus Track: Mushroom Autonomous's Commercialization Closure

Compared to Robotaxi, the autonomous driving bus sector has clearer scenarios and more explicit commercialization paths. Mushroom Autonomous has forged a differentiated path of 'pre-installation mass production + data loop + global implementation' in this sector, with its successful bid for Singapore's first L4 autonomous driving bus project serving as a benchmark for Chinese technological overseas expansion.

At the technological level, Mushroom Autonomous's core competitiveness stems from a forward-looking perception scheme. As early as two years ago, when the industry still relied on mechanical LiDAR as the mainstream, it decisively shifted to a 'vision-based + solid-state LiDAR' fusion route, accurately avoiding the complex structure, high cost, and short lifespan defects of mechanical LiDAR. In its successful bid for the Singapore project, it adopted a 128-line main solid-state LiDAR paired with four high-beam blind-spot detection radars, increasing point cloud density by 3-6 times. It can accurately lock onto pedestrians and non-motorized vehicles, combining BevFusion algorithms to achieve deep fusion of image and point cloud data. This increases target perception distance by over 50%, reduces false/missed detection rates by 70%, and significantly lowers the intervention rate by two orders of magnitude, perfectly meeting Singapore's stringent demands for dense pedestrian flow and complex road conditions during peak hours. Meanwhile, solid-state LiDAR has strong anti-vibration and anti-shock capabilities, with a lifespan of 8-10 years. The entire scheme costs only 1/3 to 1/5 of traditional mechanical LiDAR schemes, achieving a balance between high durability and low cost.

The pre-installation mass production model provides the ultimate carrier for technological implementation. Its autonomous driving system is integrated from the bottom layer into the vehicle's power, braking, and steering systems. Through a hybrid architecture of CAN/LIN buses and Ethernet, it achieves millisecond-level response to control commands, with command execution accuracy improved by over three times compared to retrofitted vehicles. Addressing localized needs such as right-hand drive, left-hand traffic, and pedestrian priority in Singapore, the system optimizes avoidance strategies and parking logic through map adaptation and scenario rule embedding. The MOGOBUS B2 model launched has completed MIIT announcement filing and is capable of scalable delivery.

Commercial Vehicle Track: Trunk Tech and Inceptio Technology's Cost Reduction Battle

Autonomous driving for commercial vehicles is one of the fastest sectors for commercial implementation. Companies like Trunk Tech and Inceptio Technology focus on the dry-line logistics scenario, achieving commercial closure by reducing labor costs and improving transportation efficiency.

From a technological perspective, autonomous driving systems designed for commercial vehicles place significant emphasis on long-distance perception and platooning capabilities. The autonomous driving heavy trucks developed by Trunk Tech are outfitted with long-range LiDAR and high-precision positioning systems. These advanced features empower the trucks to drive continuously for hundreds of kilometers on highways without interruption. Meanwhile, Inceptio Technology has entered into a collaborative partnership with FAW Jiefang and Manbang Group to construct an integrated logistics ecosystem that encompasses vehicles, cargo, and roads.

In terms of capital, commercial vehicle autonomous driving companies have formulated clearer profit models. By forging collaborations with logistics companies, they are able to generate stable revenue from orders. This business-to-business (B2B) model has gained favor in the capital market.

2026 Marks the Decisive Moment for Autonomous Driving

The year 2026 holds significant importance as a pivotal year for the autonomous driving industry. Technologically speaking, the maturation of world models and end-to-end algorithms is set to expedite the implementation of high-level autonomous driving. From a capital standpoint, the influx of funds resulting from the surge in listings will propel the industry from a phase of heavy investment and “burning money” to a path of profitability. In the market realm, global competition is expected to intensify.

It is noteworthy that mainstream original equipment manufacturers (OEMs), including Zhongtong, King Long, Geely, and Changan, have already laid a solid foundation for scalable implementation through their robust overseas market布局 (translated as “layout” here for smoother reading). This has created precise collaboration opportunities for technology service providers. Zhongtong Bus has been deeply engaged in cultivating overseas markets for years. Currently, over 80,000 new energy buses developed by Zhongtong are in operation worldwide. In recent years, the company has secured multiple “super orders,” such as an order for 1,000 units in Kyrgyzstan, 1,022 units in Saudi Arabia, and 895 units in Chile. Zhongtong Bus also boasts rich experience in adapting to localized needs, as evidenced by its narrow-body models tailored for Singapore and chassis protection features designed for the Middle East.

King Long Bus has adopted a “technology standard export + industrial chain localization” model to make breakthroughs in overseas markets. It has established 13 knock-down (KD) factories in Africa and Southeast Asia. The factory in Ethiopia has successfully built the first new energy KD project in East Africa, while the factory in Egypt has produced over 6,000 light buses. As a result, King Long Bus has formed a mature global production capacity and service network.

Geely and Changan have also capitalized on their global research and development (R&D) systems and localized operation capabilities to promote their commercial vehicle products in core markets, including Europe, the Americas, and Southeast Asia. They have constructed an operational matrix that covers a variety of road conditions and regulatory scenarios.

The overseas practical experience of these OEMs places extremely high demands on the localization adaptation, cost control, and reliability of autonomous driving systems. This is precisely where Mushroom Autonomous excels. Its full-stack self-developed autonomous driving system, combined with a “vision + solid-state LiDAR” hardware-software collaborative scheme, can be seamlessly integrated into different automakers' vehicle architectures through pre-installation mass production models. This enables the system to adapt to overseas scenarios, such as right-hand drive and complex road conditions, while achieving cost optimization and efficiency improvement. This aligns perfectly with the core demands of OEMs for scalable overseas autonomous driving implementation, laying a solid foundation for their collaborative overseas expansion.

For enterprises, going public is not the ultimate goal but rather the starting point of a new round of competition. Only those enterprises that possess both technological leadership, commercialization capabilities, and capital resilience will be able to survive in this knockout round. For investors, the autonomous driving sector presents a landscape where opportunities and risks coexist. Only by selecting enterprises with core technological barriers and commercial closed-loop capabilities can they partake in the dividends of the industry's development. The future of autonomous driving is no longer a distant dream. With the combined benefits of technology, capital, and policy, a driverless travel era is rapidly approaching.