Unitree Stock and Physical AI: What Investors Are Pricing

The trigger: a stronger software story

At the World Physical AI Conference in Hangzhou on October 22, Wang Xingxing said large models could break through in physical AI within one to three years. He described a one-to-two-year window if progress goes well, and two to three years otherwise. He also said OpenAI's GPT-6 Astra could direct robots with zero training.

That "zero training" claim should be treated as a reported statement by an executive, not as independent evidence that robots can handle a broad range of tasks without any training. A demo can show that a model can interpret an instruction or guide a robot through a specific setup. It doesn't automatically show the system will handle new rooms, different objects, and mistakes safely and consistently.

JPMorgan's October 26 upgrade added another visible data point for the bullish case. The bank moved Unitree from underweight to overweight and raised its price target to 995 yuan. It cited a prior concern that Unitree relied too much on hardware and lagged in software. After hearing from Wang and other executives and reviewing Unitree's humanoid lineup in Hangzhou, the bank revised its view.

Upgrades can shift sentiment, but they aren't the same thing as a product milestone. The question is what evidence supported the change - and whether later results actually back it up.

Why Unitree attracts attention

Unitree already sells robots, including the G1 humanoid, which the source report says costs about $16,000. A relatively lower price can make it easier for research teams and organizations to test humanoid systems. It can also help Unitree place more hardware into the real world, creating more chances to learn from day-to-day use.

The company went public on August 19 on the Shanghai Stock Exchange's STAR Market, a technology-focused board. That gives investors a way to participate in the robotics thesis, but it still doesn't eliminate the usual questions about valuation, future sales, margins, and execution.

So the share-price story has two parts: Unitree has products in market today, and investors expect AI progress to make those products more capable. The second part is the larger promise - and the less certain one.

The system is bigger than the model

A practical robot depends on multiple connected layers:

• Perception: Cameras and other sensors gather information about objects, people, and the robot's surroundings.
• Interpretation and planning: Software turns a goal (like moving a box) into a sequence of actions.
• Control: The robot uses motors and joints to execute those actions with the right precision.
• Feedback and recovery: Sensors report what happened, so the system can correct mistakes or stop.
• Safety and operations: The robot has to operate within limits set by the environment, users, and deployment requirements.

Large models may help with interpreting flexible instructions, recognizing objects, or choosing a high-level plan. But they don't remove the need for the other layers. A model that suggests a plausible action isn't enough if the robot can't execute it accurately - or if a small error turns into a safety problem.

This is why polished videos can mislead. A controlled clip can show performance in a narrow set of conditions. A real deployment has to deal with variation: different table heights, unfamiliar objects, a person walking into the robot's path, or sensors that return incomplete data.

Why chatbots do not automatically become robot operators

Language models are built for patterns in data and responding to instructions. Physical environments add constraints that a text conversation doesn't have: real-world actions have real-world consequences, observations can be noisy, and errors can damage property or hurt someone.

Timing also matters. Some decisions can be handled at a higher level, while precise movement and safety-related responses may require specialized software closer to the hardware. The system also needs to decide what each component should do, instead of expecting one model to make every choice.

There's also an evaluation problem. A robot needs enough exposure to real variation - beyond the examples used to develop it - to behave reliably in the field. Simulation can help, but the gap between simulated settings and real workplaces remains a practical hurdle. Testing has to show what the system can do, where it fails, and when it should request human help.

For businesses, the key question isn't whether a robot has a "brain." It's whether the full system can perform a specific job with acceptable reliability, oversight, and operating cost.

What "zero training" may - and may not - mean

The phrase can mean different things. It might mean the model can interpret a new instruction without being specifically trained on that exact command. Or it could refer to a setup where the robot already has the right sensors, control software, and task capabilities, so the "training" happens once and then the model is used without further adjustments.

Neither interpretation automatically means the robot can show up in an unfamiliar environment and perform any requested task. The system may still depend on extensive prior training, careful engineering, limited task scope, or human supervision. Without details on how the tests were run, it's hard to compare the claim with what customers actually need.

A more practical way to judge progress is to ask concrete questions:

• What exact task did the robot do, and how was success measured?
• Was the environment controlled, or does it resemble real usage?
• How many attempts worked, and how often did a human need to step in?
• If something goes wrong, can the system recover - or does it stop when conditions change?
• Can it operate at a practical speed and cost?

These questions are less exciting than a headline about "understanding instructions," but they get closer to whether the technology is moving from demo to deployment.

Where the model fits in the product

In the near term, the strongest role for large models may be as one component in a layered system. The model can interpret a goal or help choose a plan; specialized software still needs to handle movement, timing, and safety constraints. That division can make the system more useful without asking a general-purpose model to directly control every motor.

It also creates real product and business decisions. If a robot depends on external computing or cloud services, customers need to understand latency, uptime, data handling, and ongoing costs. If more processing happens on the robot itself, the manufacturer has to plan around hardware limits and update processes. Those are operational choices, not small implementation details.

For Unitree, the opportunity depends in part on whether its hardware can work smoothly with increasingly capable AI software. But the company still has to prove that this integration results in robots customers can run reliably - not just a stronger technical story.

The value of hardware - and its limits

Unitree's existing robot business gives it a head start that companies without shipped products often don't have. Hardware you can operate lets you test, gather feedback, and iterate. The reported G1 price of about $16,000 also sits below some rival humanoid offerings, which may make experimentation easier for others.

Still, hardware can become harder to differentiate if competitors offer machines that look similar on paper. Longer-term value likely depends on a mix of capability, reliability, software, service, and access to customers. The market concern JPMorgan raised about Unitree's software goes straight to that issue.

Competition is already showing up in different ways. Tesla has started producing Optimus robots at its Fremont factory and expects deliveries around the end of 2027, according to the source report. Figure AI partnered with OpenAI on its humanoid robots in 2024. Together, these examples suggest the race involves both robot makers and the companies building the AI systems that may run them.

For customers, competition can be helpful if it lowers costs and improves performance. For investors, it means Unitree's upside has to be weighed against the possibility that rivals develop comparable - or better - products.

Security, privacy and accountability

Robots collect information about the places where they operate. Cameras, microphones, and spatial sensors may capture details about homes, workplaces, people, and day-to-day processes. The source report notes that researchers have flagged security flaws involving audio, video, and spatial data being streamed to outside servers without users being notified.

That concern matters commercially as much as it does ethically. A company considering robots for a workplace or a household needs to know what data is collected, where it goes, who can access it, how long it's kept, and whether processing can happen locally. In the UK and Europe, deployment may also trigger data-protection obligations like the GDPR, depending on what's collected and how it's used.

A practical procurement test is to ask for clear answers about data flows, access controls, retention, security updates, and incident response before the robot enters a sensitive setting. If the vendor can't explain those basics, the product may not be ready for that use case - even if a demo looks strong.

Accountability is another open issue. When a robot makes a wrong choice, customers need clarity on how the system behaves, how a human can intervene, and who is responsible when something goes wrong. Safety and governance are part of product readiness, not something to tack on after deployment.

What to watch next

For Unitree and the broader physical AI market, a few indicators stand out. Do robots move from controlled tests into repeatable customer deployments? Do customers come back for more units? Can systems handle a broader set of tasks without needing a proportional increase in human oversight?

It's also worth watching how software is delivered and priced. If a robot requires expensive support, constant connectivity, or heavy customization, the initial purchase price may hide the real total cost. On the other hand, if a common platform can be adapted across tasks with limited engineering, that should improve economics for both vendors and customers.

Finally, pay attention to safety and data practices. Trust is part of what makes deployments work - especially when robots operate around workers or capture information inside homes. Companies that treat security and transparency as core requirements are more likely to win customers than those that focus only on demos.