Gasgoo Munich-Three months, three rounds, over 2 billion yuan — that is the fundraising scorecard Lightwheel AI posted for the first half of 2026.The latest announcement came on June 23: a 1 billion yuan strategic round. Investors included government funds like the Zhongguancun Science City Fund and the Sichuan Development Science and Technology Innovation Fund, alongside industrial capital from Giant Interactive, Yusys Technologies, and Boton Technology.At the same time, Lightwheel AI has spent the past two months adding names like PICO, Alibaba Cloud, Wuji Tech, and Boton Technology to its partnership roster.How does a three-year-old startup convince government funds, industrial giants, and financial institutions to place bets all at once?The answer lies in a broader shift: Physical AI is moving from a contest over hardware bodies and models to a battle over infrastructure. Lightwheel AI aims to be the one redefining the starting line."Four Aces: Data, Simulation, Evaluation, DeploymentLightwheel AI positions itself as a "Physical AI data and evaluation infrastructure service provider."Put simply: Lightwheel AI doesn’t build embodied intelligent robots. Instead, it provides the underlying support—"data, simulation, evaluation, and deployment feedback"—for every company that does.This positioning stems from the systematic judgment of founder and CEO Xie Chen regarding the data dilemma in Physical AI.Xie has deep roots in autonomous driving and Physical AI simulation. He previously worked at Cruise, where he led core simulation efforts, before joining NVIDIA to architect the company’s autonomous driving simulation stack.In 2023, Xie founded Lightwheel AI, focusing on the Physical AI infrastructure track to solve the core pain points of embodied intelligence R&D and deployment.In Xie’s view, unlike autonomous driving where the bottleneck is algorithm iteration, the ultimate ceiling for Physical AI and embodied intelligence is inevitably data. "The scale of data required for Physical AI is 1,000 times that of autonomous driving," Xie asserts.This gap stems from two main dimensions.First, the foundation of pre-training datasets. Large language models have free, open-source web data; autonomous driving has data loops from millions of production cars. Embodied intelligence, however, has no free, standardized, or general public pre-training dataset yet—a fundamental shortcoming for the industry.Second, the exponential difference in physical interaction complexity. Autonomous driving involves limited interactions between vehicles and road dynamics. Embodied intelligence must replicate fine-grained human physical operations across all scenarios, involving massive amounts of high-degree-of-freedom, high-precision force and posture interactions. The difficulty and data demands far exceed those of autonomous driving.A bigger challenge remains: Autonomous driving can rely on production cars to continuously feed back real-world road data. Embodied intelligence, however, is unlikely to see millions of physical units deployed in the short term. Consequently, 99.9% of training data for Physical AI cannot come from the hardware itself.In other words, data iteration for embodied intelligence must rely on human demonstrations and synthetic simulation to obtain training materials, all without a sufficient fleet of physical robots.From this, Xie derives a core conclusion: Simulation is the only viable path for large-scale evaluation in Physical AI, and the only way out for the industry.More specifically, Xie believes the industry will eventually require two billion-level generation systems: a billion-scale human data generator to support training, and a billion-scale simulation generator to support large-scale evaluation.Image Source: Lightwheel AIBased on this logic, Lightwheel AI has built a "Real2Sim2Real" (Real World → Simulation World → Real World) closed-loop system and polished four core products:EgoSuite (Data): Provides real-world experience. Focused on human behavioral data, it accumulates high-quality, scalable, cross-platform human operational experience—mainly observations, manipulations, error corrections, and long-horizon task experiences from the real world—offering robots scalable learning materials.RoboFinals (Evaluation): Validates model capabilities. Designed for industrial-scale evaluation, it uses standardized tasks, reproducible environments, and comparable metrics to determine what the robot model has learned, where its capabilities lie, and what its failure modes are, while identifying the next round of data requirements.RoboStack (Deployment): Feeds back deployment insights. Once robots enter factories, warehouses, farms, or logistics sites, they continuously encounter new task distributions, anomalies, failure samples, and on-site constraints. These feedbacks are brought back into the data, simulation, and evaluation systems, becoming the starting point for the next round of learning.SimFoundry: The Physical AI simulation infrastructure supporting the "Data-Evaluation-Deployment" loop. It solves a more fundamental problem: how to transform the real world at scale into a simulation world that robots can learn, train, and verify within.Image Source: Lightwheel AITo this end, Lightwheel AI developed a proprietary "Solver-Measure-Generate" full-stack simulation platform. It enables SimFoundry to convert real-world physical properties, scene distributions, and task experiences into executable, trainable, and evaluable simulation assets and scenes, supporting data generation, training verification, and evaluation iteration.Key to this is Lightwheel AI’s in-house high-precision GPU physics solver, which features differentiable, multi-physics, and multi-material unified solving capabilities. It supports high-precision real-time simulation of complex physical processes involving rigid bodies, soft bodies, fluids, and particles.This strategy of building from the ground up reflects Lightwheel AI’s attempt to establish a fundamental differentiation in "physical consistency."Currently, Lightwheel AI has established deep partnerships with overseas leaders like OpenAI, DeepMind, and Figure, as well as most top-tier domestic embodied intelligence startups, industrial giants, and automakers.According to previously released data, Lightwheel AI secured 550 million yuan in new orders just in the first quarter of this year.Notably, while the industry generally views data services as customized, non-standard offerings, Lightwheel AI has pioneered the "standardization" of data products. "One hour of our standardized data product can serve 10 clients across different industries simultaneously," Xie notes.Yang Haibo, co-founder and president, previously revealed that the resale rate of Lightwheel AI’s data for premium scenarios has already exceeded 10 times.This means Lightwheel AI is breaking through the "diseconomies of scale" bottleneck typical of traditional data services, converting long-tail scenario data into reusable products to dilute marginal costs.However, Xie clearly recognizes that building a complete Physical AI lifecycle education system cannot be done by a single company alone. This is precisely why Lightwheel AI is intensively connecting with various sectors of the industry.An Expanding EcosystemLooking at Lightwheel AI’s recent partnerships—spanning data collection hardware, cloud computing platforms, scenario deployment, and industry standards—this company, which builds no robots, is trying to make itself the indispensable player in the Physical AI infrastructure layer.Image Source: Lightwheel AISecuring the Data Entry: Locking in First-Hand Data Sources from the Physical World.Physical AI training begins with answering "what to learn." How humans operate, how objects bear force, how the environment reacts—this raw data is the starting point for all model capabilities. And acquiring data requires hardware.On June 18, Lightwheel AI announced a deep partnership with PICO. The two will collaborate on building infrastructure for human data collection in the Physical AI sector, jointly creating a next-generation general-purpose hardware solution for human data acquisition.Specifically, PICO will provide hardware R&D and product engineering capabilities, while Lightwheel AI brings high-quality human video data, synthetic simulation data, and industrial-grade simulation evaluation capabilities—a classic "hardware entry + data definition" combination.Almost simultaneously, Lightwheel AI reached a strategic partnership with Wuji Tech. Together, they will formulate next-generation human data collection standards, focusing on key capture areas like dexterous hand movements, tactile feedback, and force control signals, aiming to create standardized capabilities reusable across the industry.Vision and touch are the two most critical perception modalities for Physical AI. By binding key players in both fields, Lightwheel AI is, to some extent, fighting for the first gateway where data flows from the physical to the digital world, defining the "format" and "standards" of data collection.Building the Computing and Platform Base: Moving from "Workshops" to "Standardized Cloud Services."If the partnerships with PICO and Wuji Tech are about securing where data comes from, the collaborations with Alibaba Cloud and Moore Threads are about solving where data goes to be processed.On June 17, Lightwheel AI announced a deep cooperation with Alibaba Cloud to build cloud infrastructure for Physical AI. This includes creating a Physical AI simulation and evaluation cloud, a Physical AI continuous learning cloud, and an open egocentric data infrastructure. Through cloud and engineering capabilities, they aim to make originally scattered and complex workflows more accessible to developers and enterprises.Prior to this, Lightwheel AI partnered with Moore Threads in May. Leveraging Moore Threads’ full-featured GPUs and the Ku'ae intelligent computing cluster, combined with Lightwheel AI’s proprietary simulation platform, they are jointly developing a domestic, self-developed simulation synthetic data solution.These two partnerships point toward cloudification and localization, respectively.Cloudification allows Lightwheel AI to quickly reach more SME developers, turning simulation evaluation from a "luxury" for big tech into an inclusive tool for the industry. Localization embeds it within an autonomous, controllable tech stack, securing a strategic position at the policy level.Image Source: Lightwheel AIDeepening the Scenario Loop: Feeding Model Iteration with Real Industrial Data.Data and computing solve "where capabilities come from," but whether infrastructure is truly effective ultimately depends on whether it can run in real-world scenarios.Lightwheel AI is going deeper in this phase—not just signing agreements, but directly establishing joint ventures with industrial players.At the end of April, Lightwheel AI and New Hope formed a joint venture called "Xinguang World," covering logistics, farming, and retail scenarios, with the goal of accumulating high-value industrial data from real business needs.In mid-June, Lightwheel AI and Boton Technology jointly funded another venture, focusing on high-risk positions in industrial and mining settings to build a continuous learning loop from data collection to on-site feedback.After all, no matter how precise simulation data is, it requires real-world feedback for calibration. By deeply binding with industrial players, Lightwheel AI is building a "scenario data moat" that competitors will find hard to cross.Shaping Standards: From "Player" to "Referee."Beyond data entry, computing base, and scenario loops, Lightwheel AI is doing something even more fundamental: helping define the rules.In this field, Lightwheel AI has partnered with the National Robotics Testing and Assessment Center (Headquarters) to jointly build a "Real + Simulation" closed-loop testing infrastructure, pushing embodied intelligence toward more verifiable, replicable, and scalable industrial deployment.On June 15, Lightwheel AI further partnered with Shengshu Technology to jointly build high-quality world model data standards and a reproducible, high-quality world model evaluation system. They will explore conversion paths from model capabilities to scenario applications, driving the verification, feedback, and iteration of world models and embodied intelligence in real industrial settings.Even earlier, Lightwheel AI was invited to join the Newton Technical Steering Committee, working alongside giants like NVIDIA, DeepMind, Disney, and Toyota to formulate global underlying standards for physical simulation.From industry participant to co-author of the rules, if this transition is complete, Lightwheel AI will no longer be just a data service provider. It will become a new infrastructure interface for the Physical AI era—capable of supporting robots from R&D and training to evaluation and deployment, all running on its infrastructure.Lightwheel AI’s "Hidden Risks"On the surface, Lightwheel AI’s layout appears to be a stack of partnerships. Beneath it, however, lies a systematic philosophy on "how data drives the implementation of Physical AI."But whether this logic can truly close the loop depends on several variables.For instance, is simulation evaluation a genuine industry necessity, or a "manufactured necessity"?Lightwheel AI places RoboFinals at its strategic core, based on the premise that large-scale, standardized simulation evaluation is the industry's most urgent need. Yet, this judgment may not be a consensus across the sector.Wang Zhongyuan, director of the BAAI, points out that while testing in simulation is convenient, there is a clear gap with real-world scenarios. Switching to real-machine testing brings new issues, such as scene fidelity and hardware variances across devices, which affect the fairness and objectivity of evaluation.Robots in the physical world are highly sensitive to environmental details. "Even if a single screw is slightly offset, the device's success rate drops drastically. Sometimes it’s mistaken for a model error, but everything returns to normal once the part is reset," notes Xu Huazhe, founder of Pokot Robotics.Image Source: Lightwheel AIThen there is the crisis of trust regarding the physical "gap."On one hand, the physical world is full of details hard to model precisely—tiny changes in material friction coefficients, subtle differences in ambient light, accumulated tolerances in part assembly. These are often simplified or ignored in simulation.If simulation produces "physical hallucinations," the "knowledge" output by the entire "education system" could be wrong, leading robots to learn incorrect operational habits. This is the common Achilles' heel facing all simulation companies in the Physical AI industry.On the other hand, the value of a simulation evaluation platform essentially depends on "simulation-reality consistency." If the simulation environment cannot accurately reproduce the physical laws of the real world, high evaluation scores are meaningless.Furthermore, as the ecosystem of partnerships expands, how Lightwheel AI maintains focus on its core business while effectively integrating resources from various parties is a challenge that cannot be ignored.There are no ready-made answers to these challenges. The only certainty is that Lightwheel AI has already secured a strategic position during the chaotic early stages of Physical AI.As for whether this four-pronged layout of "Data + Simulation + Evaluation + Standards" will ultimately become the infrastructure of the Physical AI era or devolve into an elaborate narrative game, time will tell.