DeepSeek founder steps into embodied AI with second robot bet

  • The AI entrepreneur is backing both the “body” and “brain” of future robots as the industry races toward physical intelligence
  • The investments suggest embodied AI may arrive earlier than DeepSeek’s original AGI roadmap envisioned

DeepSeek founder Liang Wenfeng (梁文锋) is moving deeper into embodied AI, planning to lead a seed funding round for robotics AI startup Winverse AI with an investment of tens of millions of dollars, Bloomberg reported on October 13, citing people familiar with the matter.

The deal would value Winverse AI at about $400 million after the investment and mark Liang’s second investment in the embodied AI sector within two months.

From robot bodies to AI brains

In August, DeepSeek invested about 141 million yuan ($21 million) in Unitree’s strategic placement for its planned STAR Market listing.

Image generated by Tencent’s Yuanbao AI

The two companies also signed a strategic cooperation agreement covering artificial general intelligence (AGI), high-performance general-purpose robots and AI models.

While Unitree focuses on the robot’s physical body, Winverse AI is developing the “brain” — AI models designed specifically for robots.

The startup was founded by Hu Han (胡瀚), a former head of Tencent’s Hunyuan multimodal foundation model team, and focuses on robotics AI models.

Hu is also known for his work in computer vision, including the Swin Transformer paper, which won the Best Paper Award at the International Conference on Computer Vision 2021 (ICCV 2021).

AGI roadmap moves faster

Liang previously outlined a six-stage path toward AGI: language models, chain-of-thought reasoning, agents, continuous learning, self-improvement and embodied intelligence.

Embodied AI was originally viewed as the final stage. But Liang’s recent investments suggest the transition into the physical world may be happening sooner than expected.

One catalyst has been progress from OpenAI. In its RoboDojo evaluation released in September, an untuned GPT-6 Astra model achieved an average success rate of 22.48% across 42 robot manipulation tasks, outperforming 40 publicly available robot policies in the benchmark.

Unitree founder Wang Xingxing also said on October 12 that GPT-7 could potentially break through current bottlenecks in embodied intelligence.

Investment or direct entry?

According to ifeng.com, a news portal, Liang has recently held discussions with several embodied AI entrepreneurs and is considering whether DeepSeek should move further into the field. Possible options include building an internal robotics team or entering through investments.

If completed, the Winverse AI deal would become one of the largest startup investments made by DeepSeek since its founding.

The company is also reportedly preparing a new funding round targeting at least 100 billion yuan, with a valuation goal of about 500 billion yuan, as it moves toward a potential IPO in early 2027.

Header image credit: DeepSeek