Lightwheel

Lightwheel's entire business rests on a cheaper alternative to finding out the hard way. The Santa Clara company, founded in 2023 by Steve Xie, builds what it calls the data infrastructure layer for Physical AI. In plainer terms: it builds the simulated worlds where robots go to school. A robot that fails ten thousand times inside Lightwheel's simulator costs a few GPU-hours. A robot that fails ten thousand times on a real assembly line costs a great deal more, and someone eventually loses their patience.

This is not a new idea so much as a well-timed one. Xie spent years running simulation for autonomous driving, where the logic was identical: let the car crash a million times in software before it drives on a road with actual pedestrians. When humanoid robots and general-purpose manipulators started looking commercially plausible, Xie asked the obvious question - why should this be any different? - and pointed the same playbook at a much larger field.

The pitch has a pleasing contrarian shape to it. Most of the money and attention in robotics goes to the robots themselves and to the "brains" - the models that decide what to do. Those are the glamorous parts. Lightwheel deliberately skipped them. It went for the layer underneath, the part everyone needs and nobody enjoys building: physics-accurate assets, human demonstration data, and honest ways to grade whether a robot actually learned anything. In a gold rush, the reliable business is selling shovels. Lightwheel decided to sell the training ground.

The word the company keeps repeating is "SimReady," and it is worth unpacking because it is where the actual work lives. A SimReady asset is not just a 3D chair that looks like a chair. It is a chair that weighs the right amount, slides with the right friction, and tips over the right way when a robot bumps it. That fidelity is the difference between a skill that transfers to the real world and one that quietly falls apart the moment it leaves the simulator - the notorious "sim-to-real gap" that has humbled a lot of robotics ambitions.

How well is this working? The clearest signal is not a demo video but an order book. In May 2026, Lightwheel reported roughly $100 million in first-quarter orders for its physical-AI infrastructure. Order numbers are not audited revenue, and a healthy skepticism is warranted, but they are a better tell than any keynote. Customers writing purchase orders is what a market waking up looks like.

The company sits inside an unusually blue-chip web of partners. Its most consequential relationship is with NVIDIA, which featured Lightwheel in an official case study and co-develops the Newton physics engine - a GPU-accelerated, OpenUSD-based engine for robot learning. Reported customers and collaborators read like a directory of the field: Google DeepMind, Figure, AgiBot, ByteDance, Geely, BYD, Samsung, and Analog Devices among them.

The Geely example is the one that best captures the whole promise. Lightwheel used its SimReady assets and NVIDIA Isaac Sim to generate synthetic training data for Unitree H1 humanoid robots. The resulting policy was fine-tuned on NVIDIA's GR00T foundation model and then deployed on physical H1 robots inside a live Geely production facility. A robot learned its job in a simulator and then walked onto a real factory floor and did it. That sentence, compressed, is the entire thesis of Physical AI.

None of this requires Lightwheel to be large. The company is roughly 40 people. In deep infrastructure, that is not a weakness so much as the point: you do not need headcount, you need to be load-bearing. When your assets and data sit underneath everyone else's robots, being small and indispensable beats being big and famous. Whether Lightwheel stays indispensable as the giants - NVIDIA very much included - build more of this in-house is the open question worth watching. For now, it is one of the clearer bets that the robot boom has finally produced a real, boring, profitable infrastructure company.

Steve Xie is Lightwheel's founder and CEO. He holds a Ph.D. from Columbia University and a B.S. in Physics from Peking University, and he spent years leading autonomous-driving simulation before turning the same discipline toward general-purpose robots.

That background is the whole origin story. Xie had already watched simulation-first development work in one of the hardest safety-critical domains there is. Lightwheel is the wager that the approach generalizes - that the next wave of robots will be raised in software worlds before they are trusted with real ones.