Preface:

The degree of interlocking among algorithms, hardware, and environment is the most telling progress bar for physical AI today. The intelligence level of large models, by contrast, is the least worrying variable.

AI's brain, the machine's body, and the on-site environment are all indispensable. Discussing implementation without any one of them is like building a house on sand.

Author | Fang Wensan

Image source | Internet

The Computing Skeleton Collects Rent: Selling Shovels, Maps, and Guide Certificates Together

The term physical AI was first brought into prominence by NVIDIA. After Jensen Huang elevated it to the centerpiece of his keynote, the entire puzzle was already laid out on the exhibition floor at CES 2026.

The Cosmos 3 world model is divided into three tiers—Super, Nano, and Edge—with parameters ranging from 64B all the way down to 4B. The lightest tier is small enough to fit into the Jetson Thor edge module for real-time inference.

Use Super for training and Edge for deployment; one set of models handles the entire process from simulation to the roadside.

The most stable business in a gold rush is selling shovels. NVIDIA's ambition goes one step further: selling shovels, maps, and guide certificates together.

Isaac GR00T for humanoid robots, Alpamayo for autonomous driving, Omniverse for digital twins. Every step physical AI developers take forward means paying a toll to this company.

No matter whether Unitree or Figure has the last laugh, the winner at the computing layer is already locked in.

Evidence of implementation at the other end of the map is also accumulating. More than 1 million robots are running in Amazon's global warehouses, handling about 75% of fulfillment volume.

At extreme scale, physical AI has long passed the pilot stage and become infrastructure like water, electricity, and coal. The first layer that has surfaced is this skeleton of computing power and tools.

Capital Votes with Its Feet: Money Leaves the Torso and Flows to the Brain and Hands

Once the skeleton is drawn, capital's actions are more honest than research reports. The main theme of the 2026 primary market for embodied intelligence is decentralization. The share of funding absorbed by complete-machine companies is declining, while financing activity for core components and data infrastructure is rising sharply.

In May alone, multiple companies in the joint module, sensor, and data collection tracks—including Tihu Robotics, Bluepoint Touch, and Lingyu Intelligence—completed new financing rounds.

To understand this migration, look at Unitree, a rare profitable sample in the track. In 2025, revenue was 1.708 billion yuan, non-recurring net profit was 600 million yuan, and gross margin was 60.27%.

Breaking down the revenue structure, scientific research and education use accounted for 73.6%, while industrial deployment accounted for only 9%.

In other words, the company best at selling humanoid robots is actually selling high-priced scientific research and development boards. It is still a full commercialization cycle away from being a productivity tool that replaces human labor.

Capital sensed the shift. Zhiyuan Robotics simply split the company into three parts: the dexterous hand business was spun off as Linjiedian, jointly led by BlueRun Ventures and Hillhouse Venture Capital; the data collection business was spun off as Mifeng Technology, led by Sequoia China; and the leasing business was spun off as Qingtian Lease.

In February 2026, dexterous hand company Lingxin Qiaoshou completed a 1.5 billion yuan Series B financing.

Soochow Securities estimates that dexterous hands could account for 20% to 30% of complete-machine cost in the long run, second only to the body actuation system.

The body is depreciating, while reusable hands and brains are appreciating. The primary and secondary markets have already calculated this account onto the same page.

Money at the Perception Layer First Lands in Factories

For machines to work in the physical world, the first step is seeing and hearing. The consumer side relies on glasses, earphones, and headsets to capture sound and light, while the industrial side relies on 3D cameras, LiDAR, and force sensors so that robotic arms can recognize parts and avoid obstacles on production lines.

On the consumer side, Goertek holds acoustic components and smart hardware complete-machine contract manufacturing. Its microphone and speaker shipments rank among the top globally, and it handles precision components and complete-machine assembly for AI glasses and MR headsets. Its subsidiary Goermicro's MEMS acoustic sensors are core devices for on-device AI voice interaction.

The audio main control chips made by Bestechnic and Bluetrum determine whether earphones can continuously pick up sound and understand commands at low power consumption.

The industrial side is another world. Orbbec is advancing on multiple 3D vision routes at once, with structured light, iToF, dToF, and LiDAR all running. Robots rely on its depth cameras for navigation and object recognition, and its shipments rank among the top in the industry.

Luster LightTech focuses on industrial machine vision. Whether parts have defects and whether grasping positions are accurate, the production line only accepts the answers given by its vision system.

Many people still estimate its value using the shipment rhythm of consumer electronics. Seeing no blockbuster yet in headsets and glasses, they conclude that the perception layer is not yet ready to deliver. The reality is that the first wave of certain revenue at the perception layer is in factories.

The Decisive Factor at the Scenario Layer Is Hidden in the Hands of Veteran Workers

Where physical AI makes money, the answer is more conservative than imagined. Factory workstations are fixed and repetitive, and data flows back continuously, while the home environment is far messier. The first wave of large-scale revenue will definitely come from B-end production lines and energy stations.

Two paths run within this closed loop. Inovance's foundation is servo systems, inverters, and PLCs. The mechanism models and production line data accumulated over more than 20 years of industrial control are being fed into industrial AI and digital twins.

But digital twins are only a supporting capability for it. The core remains hardware control and mechanism models. It lives on this, not on selling twin software.

iSoftStone takes a pure software platform route. Its TianShu digital twin simulation platform integrates AI engines, large models, and twin agents, providing simulation training for manufacturing and energy scenarios. This is its flagship platform product marketed externally.

The most counterintuitive thing about this layer is that the barrier is not how smart the model is, but how deeply process experience has been digitized. Large models make knowledge cheap, but the on-site fine-tuning feel in the hands of veteran workers cannot be learned by general large models.

Inovance precipitates control models from countless production lines into reusable assets, while iSoftStone turns its twin platform into a reusable tool. Different paths point to the same destination.

Factories Pay Only for Results, and Equipment Sales Give Way to Continuous Services

The record disclosed by Figure at BMW's Spartanburg plant provides a rare observation window. Within 11 months, Figure 02 cumulatively loaded more than 90,000 parts, operated for more than 1,250 hours, and participated in the production of more than 30,000 BMW X3 vehicles.

These data from the company itself still need to be understood under its disclosure caliber, but they reveal the correct unit of commercialization. Customers care about whether a process with clear boundaries can operate every day, not whether a robot can demonstrate a full set of tricks in front of the camera.

Delivery is therefore increasingly like a long-term service. GXO and Agility Robotics signed a multi-year agreement in 2024 to deploy Digit in logistics scenarios through RaaS.

China's 2026 real-scenario practical training special action also wrote pay-per-utility and operating leases into policy, and proposed forming more than 100 high-value scenarios and ten-thousand-unit-scale implementation capability.

Hardware revenue remains important, while recurring revenue must be supported by uptime, skill updates, operations and maintenance, and responsibility management.

This business has not deleted people from the ledger. Remote takeover, on-site maintenance, scenario modification, safety assessment, and insurance will all change the cost of each effective operation.

If a robot service provider reports only autonomous operating hours without disclosing takeover frequency and fault recovery time, it is like a restaurant reporting only table turnover rate without counting how many people were used in the kitchen.

Scenarios will also stratify according to economic logic. Industrial handling, loading and unloading, inspection, and hazardous environment operations have relatively clear processes, and labor costs or safety benefits are easy to calculate, making them suitable for forming skill packages first.

The home environment has many objects, trivial tasks, and little room for error, so general capabilities need longer refinement.

Humanoid form is not the ticket to every scenario. It has sufficient reason to appear only when tools, passages, and processes suitable for humans exist at the same time.

Motion Data Has Become the New Oil Field, and the Drilling Platform Is Still Laying Foundations

Large models grew up drinking internet corpus, but robots do not have this ration. The data they need looks very strange: a segment of camera frames must be paired with joint angles and gripper force at the same moment. The cost of collecting such motion data is outrageously high, so the industry has piled up a three-layer pyramid.

The top of the pyramid is real-machine teleoperation, with the highest fidelity and also the most painful price.

The body is simulation data, generated by world models, with marginal cost approaching zero.

The base is bare-hand video from the human first-person perspective, with the largest volume and the lowest price. Models rely on it to learn physical common sense.

Supply-side figures have already emerged. Zhiyuan's public dataset contains more than 1 million trajectories, covering 217 tasks and 5 types of scenarios, collected using 100 dual-arm humanoid robots.

Guanglun Intelligence's human behavior database covers more than 25,000 environmental nodes and 100,000 types of tasks, with 550 million yuan in new orders in the first quarter of 2026. Mifeng Technology has set a goal of reaching tens of millions of hours of data production capacity in 2026.

MIIT has also stepped in to promote the real-scenario practical training special action, gathering fragmented data scattered across factories into reusable industry assets.

The same robot has to be retuned when switched to another production line, which shows that data has not yet grown into an asset. Only when collection standards and data levels rise will embodied intelligence meet its own ImageNet moment.

Conclusion:

The physical AI industry may develop along multiple commercial paths, growing application systems with their own characteristics on top of shared models, computing power, and data infrastructure.

This divergence precisely reflects the direction of industry maturity. Industrial competition will gradually move away from pure capability demonstrations. The meaning of machine learning understanding the physical world must ultimately be proven in reliable action.

Some materials referenced: International Federation of Robotics (IFR): "Global Sales of Professional Service Robots Surge 24%", Physical Intelligence: "Frontier Models for Every Robot", Waymo official blog: "Accelerating our global growth: Waymo raises $16 billion investment round", IDC China: "2.5 million units shipped in half a year, up 432.1% year on year: global humanoid robot commercialization exploration continues to heat up"