DeepSeek Backs Unitree IPO to Build Humanoid AI Models

Under a strategic agreement disclosed on August 6, 2026, AI laboratory DeepSeek is investing 140.8 million yuan (about $20.8 million) in Unitree Robotics' IPO on the Shanghai Stock Exchange's STAR Market. Trading under ticker 688836, Unitree priced its offering at 150.80 yuan per share, selling 40.45 million new shares to raise roughly 6.1 billion yuan at a valuation near 61 billion yuan ($9.04 billion). DeepSeek's investment secures 933,399 shares, or 2.31% of the strategic placement, subject to a 36-month lock-up. DeepSeek's subscription cap was set at 155 million yuan. Subscriptions for the IPO, which carries a price-to-earnings ratio of 219.23 compared to the 38.56 industry average, open on August 10, 2026.

The partnership addresses the physical-world data bottleneck that prevents humanoid robots from operating reliably in unfamiliar environments. While Unitree excels in mechanical engineering, DeepSeek brings massive computational expertise. DeepSeek's models perform strongly on coding, mathematics, and reasoning benchmarks, but its strength has remained in language rather than integrated multimodal systems. Under the pact, Unitree will prioritize DeepSeek for model-training services, while DeepSeek will favor Unitree hardware. This exchange allows DeepSeek to feed real-world physical data into its systems, helping transition its AI to multimodal systems. DeepSeek has previously explored this space with its VL2 vision-language model but has yet to integrate it commercially.

Unitree's revenue quadrupled to 1.7 billion yuan in 2025, with humanoid sales overtaking quadrupeds to reach 867.8 million yuan. Although first-quarter 2026 revenue rose 68.5% to 422.8 million yuan, increased spending caused profits excluding one-off items to drop 52.6% to 40.3 million yuan. Unitree projects first-half 2026 revenue between 1.052 billion and 1.128 billion yuan, up 35.6% to 45.4% year over year. However, geopolitical risks persist, as U.S. sales accounted for 13.3% of 2025 revenue, leaving the firm vulnerable to potential export controls.

For AI practitioners, this alliance signals a tightening integration between frontier foundation models and physical hardware. By securing a steady pipeline of physical interaction data, DeepSeek can train models to better translate abstract instructions into complex motor actions. This co-development model could accelerate the commercial viability of general-purpose humanoids, shifting the industry from pre-programmed demonstrations to truly adaptive, autonomous machines.