WeRide Q1 Revenue Hits $15.7M, Incubates Jingshuo to Bet on Embodied AI Infrastructure

In mid-2026, WeRide (Nasdaq: WRD, HKEX: 0800), the world's first Robotaxi stock, delivered an impressive Q1 earnings report and simultaneously unveiled a far more ambitious move: spinning off Jingshuo Technology to bet on the most scarce resource of the embodied AI era—data infrastructure.

This is not a routine personnel reshuffle or business extension. While most embodied intelligence players are still stacking "brains" and chasing demo videos, WeRide is leveraging a full decade of Robotaxi engineering experience to answer a starker question: where exactly is the "ChatGPT moment" of physical AI being held up?

Two Key Numbers from the Earnings Report

In Q1 2026, WeRide posted total revenue of 114 million yuan, up 57.6% year-over-year. Product revenue alone surged 115.8% to 20.5 million yuan. Gross margin held at 34.7%, and the company maintained 6.225 billion yuan in cash and liquid reserves—plenty of ammunition for the road ahead.

On the operational side, as of April 30, WeRide's global autonomous fleet reached approximately 2,800 vehicles, including around 1,300 Robotaxis, of which roughly 1,000 are deployed in China. At the current pace of adding about 8 vehicles per day, hitting the year-end target of 2,600 Robotaxis globally looks well within reach.

Even more telling is Abu Dhabi—WeRide's service there now covers approximately 70% of the city's core area and has reached unit economic breakeven. This is the first time a Chinese autonomous driving company has built a profitable model in an overseas market, validating that the "technology-to-revenue" flywheel is finally turning.

A Truth That Has Been Overlooked

When embodied intelligence is described as "the next ten-trillion-yuan track," capital is flooding in at unprecedented speed. In the first half of 2026 alone, Chinese embodied intelligence funding reached approximately 43.8 billion yuan, with more than half—around 24.1 billion yuan—directed at "embodied brain" startups.

But Huo Da, CEO of Jingshuo Technology (WeRide employee No. 001, with 12 years of autonomous driving data infrastructure experience), points to a frequently overlooked truth: what constrains embodied AI from scaling beyond demos is neither algorithms nor compute, but data.

Large language models achieved their ChatGPT moment because they were built on top of decades of accumulated internet text. Physical AI has no such historical dividend—robotic behavior data must be collected from scratch. The industry consensus is that a general-purpose embodied model requires at least tens of millions of hours of high-quality interaction data, potentially hundreds of millions.

"Without data, the model cannot iterate. Embodied intelligence, apart from demos, can hardly land," an industry observer told QbitAI. The binding constraint is the absence of a high-quality data infrastructure forming a closed loop to act as the model's kick start.

Jingshuo's Three-Layer Product Stack

Operating from this premise, Jingshuo Technology offers a three-layer packaged solution, each layer interconnected yet independently functional.

Layer One: WorldEngine—The Data Closed-Loop Engine Powered by a unified world model, WorldEngine covers the full pipeline: collection → curation → annotation → synthesis → evaluation → deployment. Real-world data post-deployment flows back to the collection side, calibrating the next round of collection strategy.

To support this closed loop, Jingshuo built EGOK, a fully self-developed collection device: binocular cameras outputting 4K@60fps depth data in real time, near-infrared arrays enabling sub-millimeter hand tracking, weighing 280g, running continuously for over 5 hours, with 40% lower power consumption than traditional solutions. The crucial point is that from the moment of collection, "hand–object–scene–action" are already aligned—no post-hoc stitching required.

Layer Two: GENESIS-Robotics—The World Model Core GENESIS-Robotics takes a route called Transfusion—within a single Transformer, language, policy, image, and video follow separate computational paths while sharing parameters. It performs three tasks simultaneously: world understanding (given current state and action, predict the next physical state), data synthesis (based on world understanding, generate physically plausible new scenes and interaction data), and policy generation (given state and goal, directly output actions).

With shared parameters, the flywheel turns: stronger models → better synthetic data → stronger downstream models → more precise collection strategies → higher-quality real data. Once this flywheel spins up, the competitive moat grows exponentially.

Layer Three: SkillForge—Out-of-the-Box Skill Packages SkillForge covers scenarios including kitchens, living room tidying, and industrial operations. Each skill package includes a complete task chain, 4D spatial annotation, quality evaluation, and model validation report—customers can begin training immediately upon receipt.

SkillForge is not a "dataset marketplace" but a system that organizes skill packages according to model training needs—specific state distributions plus action distributions plus evaluation standards, designed cross-embodiment, validated through the full WorldEngine pipeline, and tagged with L1/L2/L3 evaluation results. Currently, it holds 500K+ hours of real interaction data, 50M+ task fragments, 200+ standardized skill packages, covering four major domains: household, manufacturing, retail, and education.

The NVIDIA and CATL Playbook

The reference framework for this business logic is precisely the 2017 versions of NVIDIA and CATL.

Back then, the industry was chasing algorithms and tools, with compute viewed as "one link in the supply chain" rather than a strategic chokepoint. But the players who chose to provide infrastructure at that inflection point ended up becoming unavoidable presences in the AI wave. Physical AI is now entering the same inflection point. Jingshuo's choice is to stand at that position early—neither building complete machines, nor building brains, but doing the one thing that everyone will eventually need.

Huo Da has a clear judgment on this: "The shift from reckless all-in enthusiasm to cautious rationality in embodied intelligence may arrive soon."

WeRide has already validated this judgment once. The autonomous driving industry has lived through the same script—when the wind comes, everyone floors the accelerator, money pours in, and stories get told. But soon, investors, users, and the industry begin to question: after raising so much money, what can you actually deliver? WeRide has crossed that questioning cycle, and so has the Jingshuo team. That is why they know better than anyone: data infrastructure must come first.

A Different Choice for Embodied AI Startups

For the vast majority of embodied AI startups, building proprietary data infrastructure means answering two sharp questions: is there enough time? Is the cost worth it? Most teams likely have neither the cost nor the time window.

Beyond capital, the more critical asset is know-how—knowing what the model is missing, what data is genuinely useful, and how to design data that unlocks the next capability leap. This kind of know-how can only be accumulated by running closed loops inside complete business scenarios and surviving enough real cases.

Jingshuo offers an "Infrastructure-as-a-Service" alternative: WorldEngine handles data, GENESIS-Robotics generates synthetic data, SkillForge outputs ready-to-fine-tune skill packages—three layers delivered as a bundle, and engineers can start training the moment they receive it.

Not every team needs to dig the foundation from scratch. Leveraging mature infrastructure and concentrating resources on models and scenarios is the more rational choice when reading the cost and R&D timeline ledgers.

A Hidden Through-Line

WeRide has lived through all three rounds of the autonomous driving culling: algorithm competition → compute competition → data competition. The through-line that runs across all of them is the continuous construction of data infrastructure—precisely what the Jingshuo team had been doing internally before going public and independent.

This through-line is now in the spotlight because it lands precisely on the most scarce resource in embodied intelligence. As the race between "brains" and "bodies" enters its next stage, someone is finally willing to systematically build the "foundation" of physical AI.

Embodied intelligence companies no longer need to accumulate physical-world experience from scratch, no longer need to spend years building data flywheels—they can directly call on Jingshuo's world model capabilities and purchase ready-to-use skill packages. This is the first Robotaxi stock, after commercial monetization has been validated, betting on the next decade.

Source: QbitAI, "With the same script as NVIDIA and CATL, embodied AI's first 'infrastructure provider' has emerged," 2026-07-16; QbitAI, "The first Robotaxi stock is going wild again," 2026-05-14.