DeepSeek has invested 140.8 million yuan (about $20.8 million) in Unitree Robotics’ Shanghai initial public offering, securing a 2.31 per cent stake and signing a mutual procurement agreement that ties the two Hangzhou-based companies’ roadmaps together. The investment, disclosed in a stock-exchange filing on August 6, 2026, comes with a 36-month lock-up period — meaning DeepSeek cannot sell its shares for three years.
This is not a passive bet. Under the agreement, Unitree will give DeepSeek preference when procuring model-training services and technical solutions. DeepSeek, in turn, will favour Unitree when buying robots or exploring embodied-AI applications. The two companies have also committed to jointly developing AI models for humanoid robots.
What the deal actually covers
The partnership targets what the filing calls a key bottleneck for humanoid developers: building a robot brain capable of understanding unfamiliar surroundings and reliably turning instructions into physical actions. Chinese manufacturers have proven they can build inexpensive machines that walk, run, and perform choreographed routines. Turning instructions into reliable physical action in an unfamiliar room is a different problem entirely, and it requires models trained on physical-world data that is far scarcer than text.
This is where the trade works. DeepSeek brings model-building strength — its models have performed well on coding, mathematics, and reasoning benchmarks at aggressive prices. But that strength sits in language models rather than an integrated multimodal system that processes vision alongside text. DeepSeek has published separate vision-language research, including its VL2 model, but has not folded it into its flagship commercial line. A steady supply of robots, and the physical-world data they generate, is the asset Unitree offers in return for priority access to DeepSeek’s training services.
According to Reuters, the investment lands DeepSeek 933,399 shares in the strategic placement, alongside other strategic investors including a Tencent investment vehicle. DeepSeek’s subscription cap was set at 155 million yuan.
The IPO that frames it
Unitree priced its Shanghai STAR Market offering at 150.80 yuan per share, selling 40.45 million new shares — 10 per cent of its enlarged capital — to raise about 6.1 billion yuan at a valuation near 61 billion yuan ($9.04 billion), CNBC reported. That makes Unitree the first mainland-listed maker of humanoid robots in China. The price-to-earnings ratio of 219.23 sits well above the industry average of 38.56, Cailian Press noted.
The prospectus shows a company growing fast off a small profit base. Revenue more than quadrupled to 1.7 billion yuan in 2025, with humanoid robots overtaking quadrupeds as the largest business at 867.8 million yuan in sales. But momentum cooled in early 2026: first-quarter revenue rose 68.5 per cent to 422.8 million yuan, while profit excluding one-off items fell 52.6 per cent to 40.3 million yuan as research and marketing spending climbed.
We covered the IPO’s broader significance in our earlier reporting on Unitree’s $620 million public listing.
The US sales question
One exposure stands out in the filing’s risk language. US sales accounted for 13.3 per cent of Unitree’s 2025 revenue, and the prospectus says tariffs, limits on government purchases, export controls, or the loss of existing approvals could hurt overseas growth and disrupt supplies of imported parts. The company says its current humanoid and four-legged robots hold US approvals, but future models could be barred from sale there.
The Trump administration imposed new restrictions on foreign-made humanoid and four-legged robots in late July, adding a geopolitical wrinkle to the listing. Unitree’s products have already drawn FCC scrutiny in the United States.
Why a model lab wants a robot company
The strategic logic for DeepSeek is straightforward. The AI model layer is becoming commoditised — DeepSeek itself has driven prices down dramatically, offering models at a fraction of what OpenAI or Anthropic charge. Owning a stake in a hardware company that generates physical-world training data is a hedge against that commoditisation. Robot companies are building data-collection facilities where human operators repeatedly guide humanoids through tasks like folding clothes and opening doors. That data — demonstrations of manipulating objects, recovering from mistakes, responding to environments that change mid-task — is exactly what vision-language-action models need.
The deal also reflects a pattern we are seeing across the industry: AI labs and robot manufacturers are vertically integrating. NVIDIA has built its own robot foundation models. Mitsubishi is converting a car engine plant into a humanoid factory. DeepSeek buying into Unitree is the software-model version of the same consolidation.
What to watch
Subscriptions for Unitree’s offering open August 10, with payment due August 12. The joint model-development work is where the two companies say the actual product of this deal will show up — but with a 36-month lock-up, DeepSeek is making a three-year commitment, not a quick trade.
The question worth asking is whether the partnership can close the gap between Chinese humanoid hardware, which is cheap and capable of impressive locomotion, and the model layer that turns those machines from dancing demos into useful workers. The hardware is ready. The brain is not. Whether DeepSeek can build it — and whether Unitree’s robots can generate enough data to train it — is what the next three years will answer.
