Hat YuShu zwei Arbeitsausweise hingeworfen.

Jensen Huang's speech at the Taipei Pop Music Center has reached the grand finale. As a result, it was a robot that took the stage.

It's unclear since when people started calling robots "vegetative people". Maybe it's because they're not flexible enough. It seems reasonable to say so.

01

Take a look at how Jensen Huang introduced this robot. It's 1.8 meters tall, weighs 68 kilograms, and has 75 degrees of freedom throughout its body. He made a joke on stage, saying that this height and weight "are about the same as mine". It's quite interesting.

This robot is called Isaac GR00T. NVIDIA's official definition is a reference design, with three suppliers each in charge of one part.

The body comes from Unitree's H2 Plus, the hands from Sharpa's five - fingered dexterous hands in Singapore, and the brain is NVIDIA's own Jetson Thor chip, along with the full Isaac GR00T software stack.

I noticed a detail:

Yingzi said, the target users of this reference design are higher education institutions and university researchers. The first - batch customers include Stanford and ETH Zurich.

The supporting development platform and model code will soon be put on GitHub and Hugging Face. The full software stack is ready to use out of the box, and the preparation time for research teams has been shortened from days to hours.

In other words, NVIDIA is not just making one robot.

It's a turn - key project. The body, brain, data generation tools, training framework, and simulation environment are all packaged for you. You just need to plug in the power and start doing experiments.

I checked their data generation ability.

Yingzi said that with Cosmos 3 and Isaac GR00T Blueprint, 780,000 synthetic motion trajectories can be generated in 11 hours. What does 780,000 mean? It's equivalent to 6,500 hours of human demonstration data; it's almost like an engineer teaching a robot actions continuously for 9 months.

Then, this afternoon, the Shanghai Stock Exchange's listing review committee announced the results. Unitree passed the initial public offering review and meets the issuance conditions.

In 73 days, from acceptance to passing the review, it raised 4.202 billion yuan, with an overall valuation of 42 billion yuan. The first A - share humanoid robot stock is locked in. I really want to describe this as a double blessing.

But there's a detail worth noting.

In Jensen Huang's speech, Unitree's name appears in the body section. Sharpa appears in the hands section, while NVIDIA itself occupies the entire section of the brain, computing power, models, simulation, and data generation.

This afternoon, at the Shanghai review, Unitree got a valuation of 42 billion yuan. It's clearly written in the prospectus that the largest investment in the fundraising is for the embodied large model, which is the brain.

NVIDIA says you're my body, but on the same day, Unitree says it wants to build its own brain. What's going on?

02

I thought of a term, reference design. This term is quite neutral, like a technical document or a set of solutions for you to refer to.

This term has appeared many times in the tech circle, and every time it appears, the subsequent plot is similar.

The most representative time was in the mobile phone industry.

Around 2010, Qualcomm started doing something. It packaged the Snapdragon chip, baseband, Android system, driver layer, and hardware interfaces together to form a complete mobile phone reference design.

In the industry, it's called turnkey, which means a ready - to - use solution when translated.

What does it mean? If you're a mobile phone brand, you don't need to have your own chip design ability, system debugging ability, or maintain a hardware R & D team. With Qualcomm's solution, find an ODM factory, change the shell, and stick your logo on it, and a mobile phone is ready.

The first - generation Redmi was created in this way. Back then, Xiaomi had Wingtech do the OEM, using Qualcomm's solution. That year, Wingtech shipped 65.5 million units.

It sounds like a win - win situation. Qualcomm sold its chips, the brand saved on R & D, and the ODM factory got orders.

Then I checked what happened later.

Huaqin Technology, the largest mobile phone ODM company in China, had revenues of over 70 billion yuan in the first three quarters of 2024, with a net profit attributable to the parent company of 2 billion yuan. Longcheer Technology had revenues of 35 billion yuan and a net profit of less than 500 million yuan.

With 70 billion yuan in revenues and 2 billion yuan in profits, the net profit margin is less than 3%.

The gross profit margin of mobile phone OEM for these companies has been hovering between 5% and 11% for a long time. People in the industry call this hard - earned money. They're pressured by chip suppliers above, compared on price by brand owners below, and squeezed by competitors in the middle. The bigger they get, the thinner the profit margins.

Wingtech Technology, once the champion in ODM shipments, did something at the beginning of 2025. It sold its entire ODM business to Luxshare Precision and completely exited the mobile phone OEM business. After the sale, it fully shifted to semiconductors. Its semiconductor business has a gross profit margin of 37.47%, more than seven times that of mobile phone OEM.

You see, even the world's number one in making the "body" finally decided not to do it anymore.

What does this story have to do with today? I compared what Qualcomm did back then with what NVIDIA is doing today.

Qualcomm came up with the chip, Android, and the reference design, and everyone in the mobile phone industry used it. What was the result? The hardware became homogeneous, and profits slowly flowed from brand owners and manufacturers to chip suppliers and operating system providers.

Today, NVIDIA has come up with the Jetson Thor chip, the Isaac GR00T model, and the reference design. The model code is all open - source, the simulation framework is also open - source, and the data generation tools are packaged.

I checked NVIDIA's current list of partners. Unitree is using Jetson Thor, as are Zhipu, Galaxy Universal, and Ubtech. Even Figure AI, Boston Dynamics, Amazon, and Meta are using it.

Unitree is one of more than a dozen body suppliers.

The VP of NVIDIA's robotics department once said: "We don't produce robots, nor do we manufacture cars. We provide technical support for the entire industry through infrastructure computers and software."

Qualcomm said almost the same thing fifteen years ago.

When a company says, "We don't make end - products, we only provide platforms and tools," it's actually announcing one thing: "I'll set the rules."

The open - source of the GR00T model follows the same logic as Google's open - source of Android back then. The software is free for you so that you'll be dependent on my hardware. If you use my model and simulation platform, you'll have to run them on my chips.

My view is as follows:

The reference design is like a power - distribution agreement. Whoever issues the reference design defines how much the "brain" and the "body" are worth in this industry.

The mobile phone industry has already answered this question. Companies making the "body" have 70 billion yuan in revenues but a profit margin of less than 3%. Companies making the "brain" can earn tens of billions of US dollars in patent licensing fees each year. Now, coincidentally, the robotics industry has received the same agreement.

03

I looked through Unitree's prospectus. Out of the 4.2 billion yuan in fundraising, 2.022 billion yuan will be invested in the R & D of intelligent robot models, accounting for 48%, which is the largest investment among all projects. 1.11 billion yuan will be invested in the R & D of the main body, 445 million yuan in new products, and 624 million yuan in building a manufacturing base.

The area with the most investment is the brain. Unitree must be aware of this situation.

Wang Xingxing once said that the biggest mistake he made in the past ten years was underestimating the technological progress of AI. His original team focused on the main body and motion control, and only started to increase investment in the embodied large model in the past two years.

While supplying the body for NVIDIA's reference design, Unitree is investing 2 billion yuan to build its own brain. This is an independent war under the guise of cooperation.

I checked the details. NVIDIA's GR00T N1.5 has been successfully run on Unitree's G1 robot. Developers in the open - source community have directly deployed and demonstrated operation tasks on the G1 using the code. There is a complete deployment tutorial on GitHub.

That is to say, NVIDIA's "brain" has been installed in Unitree's "body", and it's public. Anyone can replicate this process.

So what is Unitree doing itself?

In September 2025, Unitree open - sourced its self - developed world model UnifoLM - WMA - 0. In January 2026, it released the visual - language - action model UnifoLM - VLA - 0.

On May 25th, the day when the listing review announcement was released, Unitree tested and released the WVLA2.0 embodied large model, enabling the G1 robot to independently complete the organization and classification of items in a meeting room in a complex environment with people walking around, without any remote control.

Two "brains" are running on the same "body". One is NVIDIA's, open - source and available to the world. The other is Unitree's, just starting out and still catching up. How can I describe this?

There's another role worth paying attention to.

I found a company called Zhongke Fifth Century, which was established in September 2024. Its core team comes from the Chinese Academy of Sciences and Tsinghua University. It has received three rounds of financing this year. Sequoia China led the Pre - A round, and the latest A round was invested by Futeng Capital and Shanghai Semiconductor Industry Investment.

It is the No. 001 supplier of the embodied operation brain for Unitree Technology.

The two parties have developed a software - hardware integrated solution for the power industry based on Unitree's G1 humanoid robot platform. Zhongke Fifth Century is also cooperating with Midea, and its robots have entered the production line of Midea's Foshan factory for actual operation.

Have you noticed the problem?

There are not just two "brains" running on Unitree's "body", but three. NVIDIA's GR00T, Unitree's self - developed UnifoLM, and Zhongke Fifth Century's FAM series models.

Why does a company making the "body" need to connect with three "brains" at the same time? Because it doesn't have its own yet.

Unitree's R & D expense ratio in 2025 was 8.53%, which is 145 million yuan. Its peer, Ubtech, had a ratio of 25%, which is 507 million yuan. Unitree is one of the companies with the lowest R & D investment ratio among the industry leaders.

This 2 billion yuan is for making up for lost ground. The problem is that there is a window period for making up.

NVIDIA's GR00T is open - source and has a fast iteration speed. It took less than three months from N1 to N1.5. As long as GR00T is good enough, more and more developers and customers will default to choosing it.

Just like after Android became popular, it's not impossible to develop your own mobile phone operating system, but it's getting more and more difficult.

What Unitree is doing now is equivalent to making money by installing Qualcomm chips in Android phones while secretly developing its own chips and operating system in the laboratory.

I believe that the state of two "brains" coexisting won't last long. There are only two possible outcomes. Either the self - developed brain catches up and NVIDIA's can be discarded, or it can't catch up and NVIDIA's becomes the only choice, and then Unitree will really only have the "body" left.

04

Speaking of this, there's a question that can't be avoided. Is there anyone who really doesn't use NVIDIA's "brain" and does everything on their own?

Yes, there is one. Tesla. And currently, it's the only one.

The chip used in Tesla's Optimus humanoid robot is Tesla's self - developed FSD chip, which is the same one used for autonomous driving in its cars.

The same training pipeline, data annotation system, and neural network architecture are directly transferred from the cars. Even the inference hardware is universal. It's currently running on HW4 and will be upgraded to AI5 in the next generation.

I checked the latest progress. At the Q1 earnings conference call this year, Elon Musk confirmed several time points.

Optimus V3 will be released in the middle of the year, and mass production will start at the Fremont factory from July to August. The predecessor of this production line was the production line for Model S and Model X. After it was shut down in May, it's being transformed into a dedicated line for Optimus, with a target annual production capacity of 1 million units.

1 million units. Unitree shipped 5,500 humanoid robots in 2025.

The difference is 180 times.

At the same time, Tesla's AI5 inference chip has completed tape - out, and its self - developed chip supply system has taken shape. This means that from training to inference, from the cloud to the robot's edge side, there is no NVIDIA product in the whole chain.

I think Tesla achieved this by relying on three cards.

The first is the data flywheel of FSD. Millions of Teslas are on the road every day, continuously transmitting real - world visual data back.

This data is used to train autonomous driving and also the perception and decision - making of the robot. The Optimus team doesn't need to collect robot data from scratch because the car data can be reused.

The second is the self - developed chip.

From Dojo to HW4 to AI5, Tesla has been developing its own computing architecture. Although there have been many setbacks with Dojo and AI5 has just completed tape - out, the direction has never changed. It doesn't want to hand over the underlying hardware of the "brain" to others.

The third is the super factory.

The manufacturing system that Tesla used to build millions of cars can be directly used to build robots. Supply chain management, quality control, and production capacity ramping up are not things that can be quickly bought with money.

Looking back at Unitree now, it doesn't have any of these three cards. Does this mean that Unitree will definitely become another Wingtech? Not necessarily.

Because Unitree has a card that Tesla doesn't have. The self - production rate of its core components exceeds 90%. It makes its own motors, reducers, and controllers.

The motion control algorithm for its quadruped robots was developed from scratch. The H1 humanoid robot was completed six months after the project was launched, with only three full - time participants. This shows that Unitree's "body" has technical value.

There's a key difference that many people overlook when comparing mobile phones and robots.

The physical form of mobile phones has become similar.

It's just a screen, a chip, a battery, and different shells. There's almost no room for hardware differentiation. So, once the chip supplier comes up with a reference design, all mobile phones look the same, and brand owners can only compete on marketing and price.

Robots are different. The ability to walk steadily, stand on one foot without falling when kicked, and unscrew a bottle cap with five fingers still varies greatly among different companies today.

This means that, at least at the current stage, making the "body" is not necessarily a dead end. There is still room for premium on the "body" itself and it hasn't been standardized away.

However, there are also new trends emerging in the industry. I noticed a judgment that the demand for embodied intelligent chips is shifting from purchasing standard products to custom - made dedicated SoCs.