At the climax of Jensen Huang's speech at the Taipei Pop Music Center, a robot took the stage.
I don’t know when people started calling robots “vegetative states.” Maybe it’s because they aren’t flexible enough—saying that doesn’t seem entirely wrong.
01
Look at how Jensen Huang introduced this robot: 1.8 meters tall, weighing 68 kilograms, with 75 degrees of freedom; he made a joke on stage saying this height and weight are “about the same as mine.” Quite interesting.
This robot is called Isaac GR00T; NVIDIA officially defines it as a reference design, with three different suppliers each handling one component.
The body is from Unitree's H2 Plus, the hands are from Singapore's Sharpa dexterous five-fingered hands, the brain is NVIDIA's own Jetson Thor chip, combined with the full Isaac GR00T software stack.
I noticed a detail:
Eiko said the target users for this reference design are higher education institutions and university researchers; the first customers include Stanford and ETH Zurich.
The accompanying development platform and model code will be uploaded to GitHub and Hugging Face shortly; the full software stack is ready to use out of the box, reducing the research team’s setup time from days to hours.
In other words, NVIDIA does more than just one robot.
It’s a turnkey solution—everything you need—the hardware, software, data generation tools, training framework, and simulation environment—is all bundled together. Just plug it in and start experimenting.
I checked their data generation capabilities.
Eiko said that using Cosmos 3 and the 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—roughly the amount of time it would take a single engineer to continuously teach a robot motions for nine months.
Then, this afternoon, the Shanghai Stock Exchange's Listing Review Committee announced the results: Yushu Technology has passed its initial public offering review and meets the issuance requirements.
In 73 days, from acceptance to approval, it raised RMB 4.202 billion with an overall valuation of RMB 42 billion—the first humanoid robot company on the A-share market, officially secured. I’d describe both the lead-up and the outcome as a double celebration.
But there is one detail worth noting,
In Jensen Huang’s presentation, Unitree appears under the “body” section; Sharpa appears under the “hands” section; NVIDIA occupies the entire segment covering the brain, computing power, models, simulation, and data generation.
In the afternoon, during the review in Shanghai, Unitree received a valuation of 42 billion. The prospectus clearly states that the largest portion of the fundraising will be directed toward embodied large models—the brain.
NVIDIA said you are my body, and on the same day, Unitree said I’m building my own brain. What’s going on?
02
I came up with a term: reference design. It’s neutral, like a technical document or a set of solutions—feel free to refer to it.
This term has appeared many times in the tech world, and each time, the subsequent storyline is much the same.
The most representative instance was in the mobile phone industry.
Around 2010, Qualcomm began doing something: it bundled the Snapdragon chip, baseband, Android system, driver layer, and hardware interfaces into a complete smartphone reference design.
In the industry, it's called turnkey, which translates to "key-ready."
What does it mean? You’re a smartphone brand with no need to develop your own chips, tune the system, or maintain a hardware R&D team. Simply take Qualcomm’s solution, partner with an ODM factory, modify the casing, add your logo, and you’ve got a smartphone ready.
This is how the first Redmi came about. Back then, Xiaomi partnered with Wingtech for manufacturing, using Qualcomm’s solution; that year, Wingtech shipped 65.5 million units.
It sounds like a win-win situation: Qualcomm sold its chips, brands saved on R&D, and ODM factories secured orders.
Then I looked into what happened afterward.
Huaqin Technology, China's largest smartphone ODM company, generated revenue of over RMB 70 billion and a net profit attributable to parent shareholders of RMB 2 billion in the first three quarters of 2024. Longqi Technology reported revenue of RMB 35 billion and a net profit of less than RMB 500 million.
$70 billion in revenue, $2 billion in profit, with a net profit margin of less than 3%.
The gross profit margins of these companies' mobile manufacturing services have long hovered between 5% and 11%. Industry insiders call this hard-earned money—pressed from above by chipmakers, undercut from below by brand owners, and squeezed in the middle by competitors. The bigger they grow, the thinner their margins become.
Wingtech, formerly the top ODM shipper, sold its entire ODM business to Luxshare Precision at the beginning of 2025, completely exiting phone manufacturing; after the sale, it fully shifted to semiconductors, where its semiconductor business boasts a gross profit margin of 37.47%, more than seven times that of phone manufacturing.
Look, he reached the number one position in the world in bodybuilding, but ultimately chose to quit.
What does this story have to do with today? I compared what Qualcomm did back then with what NVIDIA is doing today.
Qualcomm released a chip, Android, and a reference design, and everyone in the smartphone industry adopted them. What happened? Hardware became homogenized, and profits gradually shifted from brands and manufacturers to chipmakers and operating system providers.
NVIDIA today released the Jetson Thor chip, the Isaac GR00T model, and a reference design. The model code is fully open-sourced, the simulation framework is open-sourced, and the data generation tools are bundled together.
I checked NVIDIA’s current list of partners—Unitree is using Jetson Thor, Agi and Galaxy General are using it, as well as UBTECH. Even Figure AI and Boston Dynamics are using it, along with Amazon and Meta.
Unitree is one of a dozen body suppliers.
The VP of NVIDIA’s robotics division said: “We don’t build robots or manufacture cars; we provide the foundational computing and software that support the entire industry.”
Fifteen years ago, Qualcomm said nearly the exact same thing.
When a company says, “We don’t make end products—we only provide platforms and tools,” it’s really declaring one thing: I set the rules.
The GR00T model is open-source, following the same logic as when Google open-sourced Android: the software is given to you for free, so you become dependent on my hardware. If you use my model and simulation platform, you must run it on my chip.
My view is this:
Referring to the design is like a power allocation agreement— whoever submits the reference design is defining how much the brain and the body are worth in this industry.
The mobile phone industry has already been there: companies making bodies generated $70 billion in revenue with profit margins under 3%; companies making brains collected tens of billions of dollars annually just from patent licensing. Now, coincidentally, the robotics industry has secured the same agreement.
03
I reviewed Unitree’s prospectus. Of the RMB 4.2 billion raised, RMB 2.022 billion will be allocated to intelligent robotics model development, accounting for 48%—the largest portion among all projects; RMB 1.11 billion will go toward body design research, RMB 445 million toward new product development, and RMB 624 million toward building a manufacturing base.
The most money is spent on the mind. Unitree certainly understands this game.
Wang Xingxing once said that the biggest mistake over the past decade was underestimating the technological advancements in AI; his team had long focused primarily on robotics and motion control, only beginning to heavily invest in embodied large models in the last two years.
While supplying bodies for NVIDIA’s reference designs, spending $2 billion to build your own brain. This is an independence war dressed in the guise of cooperation.
I checked the details—NVIDIA’s GR00T N1.5 has already been successfully run on Unitree’s G1 robot; developers from the open-source community have directly deployed and demonstrated operational tasks on the G1. A complete deployment guide is available on GitHub.
In other words, Eiko's mind has been uploaded into Yushu's body—and this process is public, allowing anyone to replicate it.
What is Utsuki doing himself?
In September 2025, Unitree open-sourced its proprietary world model, UnifoLM-WMA-0. In January 2026, it released the vision-language-action model, UnifoLM-VLA-0.
By May 25, the day the meeting announcement was released, Unitree publicly demonstrated its WVLA 2.0 embodied large model, enabling the G1 robot to independently organize and categorize items in a cluttered environment with people moving around, without any remote control.
Two sets of brains running on the same body—one is NVIDIA’s, open-source, available to everyone worldwide; the other is Unitree’s own, just starting out and still catching up. How am I supposed to describe that?
There is another role worth noting.
I found a company called Zhongke Fifth Epoch, established in September 2024, with a core team originating from the Chinese Academy of Sciences and Tsinghua University; this year, it has secured three rounds of funding, with Sequoia China leading the Pre-A round, and the latest A round led by Futeng Capital and Shanghai Semiconductor Industry Investment.
It is the official embodied operation brain supplier for Unitree Robotics, No. 001.
Both parties have developed an integrated software-hardware solution for the power industry based on Unitree's G1 humanoid robot platform; Fifth Generation of CAS is also collaborating with Midea, and its robots are already operating on Midea’s production lines in Foshan.
Have you found the issue?
The Unitree robot's body runs not just two, but three brains: NVIDIA's GR00T, Unitree's proprietary UnifoLM, and the FAM series models from Fifth Generation of China.
Why would a company that builds bodies need to connect to three sets of brains? Because it doesn’t have its own yet.
In 2025, Unitree's R&D expense ratio was 8.53%, amounting to RMB 145 million; in comparison, Ubtech's was 25%, or RMB 507 million. Unitree is among the companies with the lowest R&D investment ratios among industry leaders.
These 2 billion are for catching up. The issue is that there’s a window for catching up.
NVIDIA's GR00T is open-source and iterates rapidly. The transition from N1 to N1.5 took less than three months; as long as GR00T is practical, more and more developers and customers will default to using it.
Just like after Android expanded, creating your own mobile operating system isn't impossible—it's just becoming increasingly difficult.
What Unitree is doing now is equivalent to earning money by shipping Android phones with Qualcomm chips, while secretly developing its own chips and operating system in the lab.
I believe the state of having two brains coexisting won't last long—there are only two possible outcomes. Either our in-house brain catches up, making Yingzi’s unnecessary; or it doesn’t, making NVIDIA’s the only option, leaving Yushu with nothing but its body.
04
Speaking of which, there’s one question that can’t be avoided: Has anyone really managed to do everything on their own without using NVIDIA’s brain?
Yes, one: Tesla. And currently, only this one.
The chip used in Optimus, Tesla's humanoid robot, is Tesla's in-house developed FSD chip—the same one used in its vehicles for autonomous driving.
The same training pipeline, data annotation system, and neural network architecture were directly transferred from the vehicle. The inference hardware is also compatible, currently running on HW4, with the next generation upgrading to AI5.
I checked the latest updates, and during the first-quarter earnings call, Musk confirmed several timelines.
Optimus V3 will be launched mid-year, with mass production set to begin at the Fremont factory in July to August. This production line, previously used for Model S and Model X, is currently being reconfigured for Optimus after production ended in May, with an annual capacity target of 1 million units.
1 million units. Unitree shipped 5,500 humanoid robots in total for the full year of 2025.
180 times different.
Meanwhile, Tesla’s AI5 inference chip has completed tape-out, and its self-developed chip supply chain is now in place—meaning that, from training to inference, and from the cloud to robotic endpoints, the entire pipeline no longer relies on NVIDIA’s products.
I believe Tesla achieved this through three key strategies.
First, the FSD data flywheel. Millions of Teslas drive on roads every day, continuously sending back real-world visual data.
This data is used to train autonomous driving systems and also to enhance robots' perception and decision-making; the Optimus team doesn't need to collect robotic data from scratch, as vehicle data can be reused.
Second, proprietary chip.
From Dojo to HW4 to AI5, Tesla has consistently developed its own computing architecture. Although Dojo faced numerous challenges and AI5 has only just been taped out, the direction has remained unchanged—Tesla does not want to entrust the underlying hardware of its brain to others.
Third, the super factory.
The manufacturing system Tesla used to produce over a million vehicles can be directly applied to building robots; supply chain management, quality control, and production scaling are not things you can quickly buy with money.
Looking back at Yu Shu, none of these three cards were drawn. Does that mean Yu Shu must have become Wentai? Not necessarily.
Because Unitree holds a card that Tesla doesn’t—over 90% of its core components are自主研发 and self-produced, including the motors, reducers, and controllers.
The motion control algorithm for the quadruped robot was developed from scratch, and the humanoid robot H1 was completed within six months of project initiation, with only three full-time team members involved—this demonstrates that Unitree’s hardware technology is highly sophisticated.
There’s a key distinction that many people overlook when comparing phones to robots.
The physical form of smartphones has become standardized.
A screen, a chip, a battery—just different casings. There’s almost no room for hardware differentiation. So once chipmakers release reference designs, all phones look the same, and brands can only compete on marketing and price.
Robots differ significantly—some can walk steadily, balance on one foot even when kicked, and open bottle caps using five fingers. Today, the gap in these capabilities between different companies is substantial.
This means that, at least for now, physical delivery isn't necessarily a dead end; the physical asset still has room for premium and hasn't yet been fully commoditized.
However, new trends are emerging in the industry. I’ve observed a shift in demand for embodied intelligence chips—from purchasing off-the-shelf products toward customized dedicated SoCs.
This means that in the future, every robotics company may form joint ventures with chip companies to develop their own specialized chips; if this trend holds, the lock-in effect of Yingzi’s reference design will be weakened.
Right now, this window is still open; if you leap through, it’s Tesla. If you can’t, it’s Wentai. Yu Shu is betting 2 billion on this very thing.
The window won't stay open forever; with each iteration of GR00T, this window closes a little more. From N1 to N1.5, in three months, the time left for Unitree might be two or three years.
Of course, don't be too pessimistic—these are just some personal research opinions.
This article is from the WeChat official account "Wang Zhiyuan" (ID: Z201440), authored by Wang Zhiyuan.