Two events that until this morning sat in different sections of the financial press converged in a single news cycle on June 1. In Shanghai, the STAR Market listing committee convened to review Unitree Robotics’ IPO application — 4.2 billion yuan ($620 million) raise, target valuation 42 billion yuan ($6.2 billion), filed seventy-three days ago. In Santa Clara, Nvidia announced the Isaac GR00T Reference Humanoid Robot — the first open humanoid reference design built on Jetson Thor — and named its chassis vendor. The chassis vendor is Unitree.
PANews put it most cleanly: on the day Unitree went public, Nvidia unveiled a humanoid robot. The two sentences describe the same morning.
The reference design in numbers
The reference platform is the Unitree H2 Plus, an upgraded version of the H2 chassis Unitree has been quietly shipping to research customers since late 2025. Nearly six feet tall. 150 pounds. 31 degrees of freedom across the body. Onto each wrist, Nvidia and Unitree bolted a Sharpa Wave tactile five-finger hand at 22 DoF per side, bringing the system to 75 total degrees of freedom from shoulder to fingertip. The onboard compute is Nvidia’s Jetson Thor module — the same edge-AI accelerator Nvidia announced for robotics at GTC March — running the open Isaac GR00T software stack. The system will be available from Unitree in late 2026, and anyone with the catalog price can buy it.
Day-one research customers named in the press materials: Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego’s Advanced Robotics and Controls Laboratory. Four institutions whose graduate students have spent the last five years building bespoke humanoid platforms from off-the-shelf parts and proprietary toolchains. Those workflows are now on a clock.
What the reference design changes for everyone else
The single sentence that matters in Nvidia’s release is that the design is open. Not the Unitree chassis hardware itself — that remains a buy — but the integration layer: the Jetson Thor pinout, the GR00T model architecture, the Sharpa Wave hand API, the data-capture-to-deployment pipeline. The point of a reference design is that no PhD candidate, no postdoc, no startup founder needs to redo the integration work. Buy the H2 Plus, plug in the Sharpa hands, install GR00T, and you are at the starting line of the same race as everyone else.
For the humanoid robotics industry that has spent eighteen months arguing about whether the platform layer or the model layer is the bottleneck, Nvidia just answered the question with a stack. The platform layer is now a commodity reference. The model layer is now open. The startups whose differentiation was “we figured out how to make this hardware talk to that software” have a roadmap problem.
For Boston Dynamics, Figure, Apptronik, 1X, Agility, AGIBOT, and a long tail of better-capitalized humanoid vendors, the implication is more pointed. The reference design’s spec sheet — 75 DoF, ~6 feet, 150 lbs, Sharpa-grade fingers — is the new floor every academic lab in the world can buy off the shelf. Anything below it is hard to justify as a research platform. Anything above it has to explain the premium.
What the IPO day adds
The Shanghai listing review is the second-order event. Unitree is asking the STAR Market for 4.2 billion yuan to scale production and stake the claim of being the “first embodied intelligence stock” on China’s A-share market. The prospectus shows revenue growing from RMB 159 million in 2023 to RMB 1.699 billion in 2025 — a 226.78% CAGR — and the company shipped more humanoid units globally in 2025 than any other vendor.
The Q1 2026 detail that the AI Weekly tracker flagged is the wrinkle: revenue grew 68.49% year-over-year, and net profit after deductions fell 52.55%. RMB 38.3 million of incremental R&D spend, and significantly higher sales costs, ate the margin. That is the price of preparing a chassis to be the Nvidia reference platform while simultaneously scaling production for an IPO road show.
The Nvidia announcement reframes that margin compression in a useful way for Unitree’s prospectus reviewers. The R&D burn is the moat-builder; the partnership confers a Western research-market distribution channel that Boston Dynamics and Figure have spent years building. The same morning, Wang Xingxing’s company moves from “Chinese humanoid vendor with strong domestic shipments” to “Nvidia’s reference chassis partner, available globally through academic procurement.”
What it does to the rest of the humanoid market
Three immediate reads.
The first is on price. The Unitree H2’s previous list price was around $30,000-$35,000; the H2 Plus has not been priced publicly yet, but the academic-research positioning suggests it will land in the same band. Set against Foundation’s Phantom MK-1 at $150,000 and Tesla Optimus’s stated $20,000-$30,000 production target, the H2 Plus puts the reference price firmly in the “every robotics lab can budget for this” range. The under-$50K segment becomes the only segment that matters for research procurement.
The second is on software. Until today, the GR00T foundation model lived inside Nvidia’s developer cloud and could only be evaluated on either Nvidia’s own simulation stack or a small number of vendor partnerships. The Unitree H2 Plus reference makes GR00T a deployable platform a postdoc can run end-to-end. That is the single biggest distribution boost the foundation-model layer of humanoid robotics has gotten this year.
The third is on the China-US axis. Nvidia choosing a Chinese chassis vendor for its first open humanoid reference design is a strong statement about what the export-controlled compute layer can ship together with. The Jetson Thor module is not on the entity list. The Unitree H2 Plus is a commercial product available globally. The combined platform is being sold to US universities. For the policy framework around AI hardware exports, this is a moving target.
What to watch for in June
- H2 Plus pricing. Unitree has not published the H2 Plus list price yet. The number determines whether the reference design lands as the new floor for academic robotics or as a premium tier above the existing H2.
- Western-vendor responses. Boston Dynamics, Figure, Apptronik, and 1X all have research-procurement channels of their own. The first vendor to either match the GR00T-compatible reference (license GR00T on their own chassis) or counter it (publish their own open reference design) sets the second-mover template.
- The Shanghai listing decision. The STAR Market review on June 1 was a hearing. The formal approval or rejection lands within weeks. Approval converts Unitree into the first publicly traded humanoid pure-play. Rejection forces a re-file.
- Tariffs and export friction. A Chinese-chassis, US-compute reference design will draw attention from both Commerce and MIIT. The first formal letter about it sets the regulatory framing for the rest of the year.
Two press releases dropped at roughly the same hour this morning. One says Unitree’s IPO has just been reviewed in Shanghai. The other says Nvidia has just made Unitree’s chassis the world’s open reference humanoid. Those are not unrelated. They are the same event told from two ends of the supply chain. The humanoid platform layer is now a commodity reference, the embodied-intelligence model layer is now open, and the company holding both is, as of this morning, asking Shanghai for the public-market valuation to scale up.