Jensen Huang was giving a speech at the Taipei Music Center, and for the grand finale, a robot came on stage.
At some point, people started calling robots "plant people." Probably because they're not agile enough yet; that description seems fitting.
01
Listen to how Jensen Huang introduced this robot: 1.8 meters tall, 68 kilograms, with 75 degrees of freedom throughout its body. On stage, he joked that its height and weight were "about the same as mine." Quite amusing.
This robot is called Isaac GR00T. NVIDIA officially defines it as a reference design, with three suppliers each responsible for a part.
The body comes from Unitree's H2 Plus, the hands are the dexterous five-finger hands from Singapore's Sharpa, and the brain is NVIDIA's own Jetson Thor chip, along with the full Isaac GR00T software stack.
I noticed a detail:
Jensen said, the target users for this reference design are higher education institutions and university researchers; initial customers include Stanford and ETH Zurich.
The accompanying development platform and model code will soon be released on GitHub and Hugging Face; the full software stack is ready to use, reducing research teams' setup time from days to hours.
In other words, what NVIDIA has made is more than just a robot.
It's a turnkey solution; body, brain, data generation tools, training frameworks, simulation environments—all packaged together. You just need to plug it in to start experimenting.
I looked into their data generation capabilities.
Jensen mentioned that using Cosmos 3 and Isaac GR00T Blueprint, they can generate 780,000 synthetic motion trajectories in 11 hours. What does 780,000 mean? It's equivalent to 6,500 hours of human demonstration data; roughly like an engineer continuously teaching robot movements for 9 months.
Then, this afternoon, the Shanghai Stock Exchange Listing Committee announced its decision: Unitree's IPO application has passed the review, meeting the issuance conditions.
73 days from acceptance to approval, raising 4.202 billion RMB, with an overall valuation of 42 billion. The title of the first humanoid robot stock on the A-share market is secured. With one event after another, I'm tempted to call it a double blessing.
But one detail deserves attention.
In Jensen Huang's speech, Unitree's name appeared in the body section. Sharpa was in the hand section; NVIDIA itself occupies the entire section for brain, computing power, models, simulation, and data generation.
In this afternoon's review in Shanghai, Unitree secured a 42 billion valuation. The prospectus clearly states that the largest portion of the raised funds is allocated to the embodied large model. The brain.
NVIDIA says you are my body. On the same day, Unitree says I will build my own brain. What's going on?
02
I thought of a term: reference design. It's quite neutral, like a technical document, a set of plans for you to reference.
This term has appeared many times in the tech world, and each time, the subsequent plot has been similar.
The most representative instance was in the mobile phone industry.
Around 2010, Qualcomm started doing something. It packaged the Snapdragon chipset, baseband, Android system, driver layer, hardware interfaces, etc., into a complete mobile phone reference design.
In the industry, it's called "turnkey."
What does that mean? If you are a mobile phone brand, you don't need to have your own chip design capabilities, system debugging capabilities, or maintain a hardware R&D team. Take Qualcomm's solution, find an ODM manufacturer, modify the shell, slap on your own logo, and a phone is ready.
The first-generation Redmi came about this way. Back then, Xiaomi contracted Wingtech for production using Qualcomm's solution. That year, Wingtech shipped 65.5 million units.
It sounds like a win-win. Qualcomm sold chipsets, brands saved on R&D, ODM factories got orders.
Then I looked into what happened afterward.
Chinaxinfeng Technology, China's largest mobile phone ODM company, had revenue exceeding 70 billion RMB in the first three quarters of 2024, with a net profit attributable to the parent of 2 billion. Longcheer Technology, revenue of 35 billion RMB, net profit less than 500 million.
70 billion in revenue, 2 billion in profit, a net profit margin of less than 3%.
The gross margins for mobile phone ODM at these companies have long hovered between 5% and 11%. People in the industry call this "sweat money"; squeezed from above by chip suppliers, pressured on price from below by brands, and caught in the middle by competitors' undercutting. Getting bigger, getting thinner.
Wingtech, once the ODM shipment champion, did something in early 2025: it sold its entire ODM business to Luxshare Precision, completely exiting mobile phone manufacturing. After selling, it fully pivoted to semiconductors; its semiconductor business gross margin was 37.47%, more than seven times that of phone ODM.
See? Making the body to become the world's number one, but ultimately choosing to stop.
How does this story relate to today? I compared what Qualcomm did then with what NVIDIA is doing now.
Qualcomm released chipsets plus Android plus reference designs, and everyone in the mobile phone industry used them. The result? Hardware became identical across thousands of models, and profits slowly flowed from the hands of brands and manufacturers to the chip suppliers and operating system providers.
Today, NVIDIA is releasing the Jetson Thor chip plus the Isaac GR00T model plus a reference design. The model code is fully open-source, the simulation framework is open-source, and the data generation tools are packaged.
I looked at NVIDIA's current partner list. Unitree is using Jetson Thor; Zhi Yuan (Agibot), Galaxy Universal, Ubtech are using it. Even Figure AI, Boston Dynamics, Amazon, and Meta are using it.
Unitree is one of over a dozen body suppliers.
The VP of NVIDIA's robotics division once said, verbatim: "We don't produce robots, and we don't make cars; we support the entire industry with infrastructure computers and software."
Qualcomm said an almost identical version of this sentence 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 set the rules.
Open-sourcing the GR00T model follows the same logic as Google open-sourcing Android. Give you the software for free, so you become dependent on my hardware. Use my model, my simulation platform, and you'll need to run it on my chips.
My take is this:
A reference design is like a power-sharing agreement. Whoever issues the reference design defines how much the brain is worth and how much the body is worth in that industry.
The mobile phone industry already answered this once. Companies making bodies: 70 billion in revenue, profit margins below 3%. Companies making brains: collecting tens of billions of dollars in patent licensing fees annually. Now, coincidentally, the robotics industry has received the same agreement.
03
I looked through Unitree's prospectus. Of the 4.202 billion RMB raised, 2.022 billion is to be invested in intelligent robot model R&D, accounting for 48%, the largest single project; 1.11 billion in body R&D, 445 million in new products, 624 million in building a manufacturing base.
The biggest investment is in the brain. Unitree certainly knows the stakes.
Wang Xingxing once said something to the effect that the biggest mistake made in the past decade was underestimating the pace of AI progress. His team's focus had always been on the body and motion control, and only in the last two years did they start heavily investing in embodied large models.
Supplying the body for NVIDIA's reference design on one hand, investing 2 billion RMB to build their own brain on the other. This is a war for independence cloaked in cooperation.
I checked the details. NVIDIA's GR00T N1.5 is already running on Unitree's G1 robot; open-source community developers have taken the code and directly deployed and demonstrated operational tasks on the G1. There's a complete deployment tutorial on GitHub.
In other words, Jensen's brain is already installed in Unitree's body. And it's public; anyone can replicate this process.
So what is Unitree itself doing?
In September 2025, Unitree open-sourced its self-developed world model, UnifoLM-WMA-0. In January 2026, it released the Vision-Language-Action model, UnifoLM-VLA-0.
By May 25th, the day of the approval announcement, Unitree tested and released the WVLA2.0 embodied large model, enabling the G1 robot to independently complete tasks like organizing and categorizing items in a meeting room under complex conditions with human movement interference, without any remote control.
Two brains, running on the same body. One is NVIDIA's, open-source, available to the world. The other is Unitree's own, just starting, still playing catch-up. How should I describe this?
There's another player here worth watching.
I found a company called Zhongke Diwuji (ZKDWJ). Founded in September 2024, its core team comes from the Chinese Academy of Sciences and Tsinghua University. This year, it secured three rounds of financing; Sequoia China led the Pre-A round, and the latest A round was invested by Funteng Capital and Shanghai Semiconductor Industry Investment Fund.
It holds the position of Unitree's No. 1 embodied operation brain supplier.
Based on the Unitree G1 humanoid robot platform, the two parties developed an integrated hardware-software solution for the power industry. ZKDWJ is also collaborating with Midea; its robots have already entered the actual production lines of Midea's Foshan factory.
Do you see the issue?
The brains running on Unitree's body aren't just two; there are three. NVIDIA's GR00T, Unitree's own UnifoLM, and ZKDWJ's FAM series models.
Why does a company making bodies need to interface with three different brains simultaneously? Because it doesn't yet have its own definitive one.
Unitree's R&D expense ratio for 2025 was 8.53%, amounting to 145 million RMB. In comparison, competitor Ubtech's was 25%, or 507 million RMB. Unitree is among the industry leaders with the lowest R&D investment ratio.
This 2 billion RMB is money for remedial study. The problem is, there's a window period for catching up.
NVIDIA's GR00T is open-source and iterating fast. From N1 to N1.5 took less than three months. As long as GR00T is good enough, more and more developers and customers will default to choosing it.
Just like after Android proliferated, making your own mobile OS wasn't impossible, but it became increasingly difficult.
What Unitree is doing now is equivalent to shipping phones with Qualcomm chips and Android to make money, while secretly working on its own chips and operating system in the lab.
I believe the state of two brains coexisting won't last too long. The outcome is one of two possibilities. Either the self-developed brain catches up, and Jensen's can be phased out. Or it doesn't catch up, and NVIDIA's becomes the only choice, leaving Unitree truly just a body supplier.
04
This brings up an unavoidable question. Is there anyone who truly doesn't use NVIDIA's brain and handles everything themselves?
Yes, one company. Tesla. And currently, it's the only one.
The chip used in the Optimus humanoid robot is Tesla's self-developed FSD chip, the same one used for autonomous driving in its cars.
The same training pipeline, data labeling system, and neural network architecture were directly ported from the car. Even the inference hardware is shared; currently running on HW4, with the next generation upgrading to AI5.
I checked the latest progress. During this year's Q1 earnings call, Musk confirmed several timelines.
Optimus V3 is set for release mid-year, with mass production starting at the Fremont factory in July or August. This production line was formerly the Model S and Model X line. After being decommissioned in May, it is being converted into a dedicated Optimus line, targeting an annual capacity of 1 million units.
One million units. Unitree shipped 5,500 humanoid robots in all of 2025.
A difference of 180-fold.
Meanwhile, Tesla's AI5 inference chip has completed tape-out, forming its self-developed chip supply system. This means that from training to inference, from cloud to robot edge, there's nothing from NVIDIA in the entire chain.
I think Tesla managed this by playing three cards.
First card: The FSD data flywheel. Millions of Teslas on the road every day continuously send back real-world visual data.
This data trains autonomous driving and also trains the robot's perception and decision-making. The Optimus team doesn't need to collect robot data from scratch because car data can be reused.
Second card: Self-developed chips.
From Dojo to HW4 to AI5, Tesla has consistently built its own computing architecture. Although Dojo faced setbacks and AI5 is newly taped-out, the direction hasn't changed. It doesn't want to hand over the underlying hardware of the brain to others.
Third card: The Gigafactory.
The manufacturing system Tesla used to build millions of cars can be directly applied to building robots. Supply chain management, quality control, and production ramp-up—these aren't things you can quickly buy with money.
Looking back at Unitree now, it doesn't have any of these three cards. Does this mean Unitree is destined to become like Wingtech? Not necessarily.
Because Unitree has one card Tesla doesn't: over 90% self-developed and self-produced rate for core components. Motors, reducers, controllers—all made in-house.
Motion control algorithms for quadruped robots developed from scratch; the humanoid robot H1 prototype completed in six months with only three people fully involved. This shows Unitree's body has technical depth.
There's a crucial difference here that many overlook when drawing analogies between phones and robots.
The physical form of mobile phones has converged.
A screen, a chip, a battery, with different shells. There's almost no room for hardware differentiation. So when a chip supplier releases a reference design, all phones look the same, and brands can only compete on marketing and price.
Robots are different. Whether they can walk steadily, stand on one foot without falling when kicked, or use five fingers to twist open a bottle cap—these capabilities, even today, vary greatly between companies.
This means, at least at the current stage, making the body isn't necessarily a dead end; the body itself still has room for premium pricing; it hasn't yet been consumed by standardization.
However, new trends are emerging in the industry. I've noted one observation: the demand for embodied intelligence chips is shifting from purchasing standard products towards custom application-specific SoCs.
This means, in the future, each robot company might partner with a chip company to develop its own dedicated chip. If this trend holds, the lock-in effect of NVIDIA's reference design would weaken.
For now, this window is still open. Climbing through it leads to Tesla. Failing to climb through leads to Wingtech. The 2 billion RMB that Unitree is betting is on this very thing.
The window won't stay open forever. Each iteration of GR00T lowers it a bit more. From N1 to N1.5 took three months. The time left for Unitree might be two or three years.
Of course, don't be too pessimistic. Just some personal research observations.
This article is from the WeChat public account "王智远" (ID: Z201440), author: Wang Zhiyuan
