DeepSeek's First Outbound Bet After Its Funding Pause Is a $400M Robot Brain — Built by Tencent's Ex-AI Chief

DeepSeek has agreed to lead a seed round in Winverse AI, a months-old, still-stealth Hangzhou robotics-AI startup founded by Hu Han, the former lead scientist on Tencent's vision-language and embodied foundation model teams. The deal values Winverse at roughly $400M post-money, with Meituan's venture arm Long-Z committing alongside, Bloomberg reported on Oct. 12, 2026. Round size is \"tens of millions of dollars,\" with investor roster still unsettled.

It is DeepSeek's first publicly disclosed outbound capital move since its July 2026 funding pause — when founder Liang Wenfeng suspended the lab's second external round after viral online posts of his investor calls — and the destination is robotics, not new model training. That is the strategically interesting part: a lab that built its brand on cheap-to-train frontier models is now deploying capital into the layer that runs models on physical bodies.

Three forces converged in October. First, DeepSeek just closed a $12B-plus round at a ~$74B pre-money valuation, with Tencent and CATL among the backers, and is publicly considering expanding it to as much as 100B yuan (~$15B). Second, Tencent Hunyuan's multimodal lead Hu Han resigned in July 2026 amid a broader personnel shakeup around the Hunyuan large-model system — and founded Winverse within months. Third, China's policy direction has shifted decisively: embodied AI was elevated into the country's 15th Five-Year Plan (2026–2030), and shipments of Chinese-made humanoids hit roughly 25,000 units in H1 2026, up 432% year over year per IDC, with Chinese makers holding ~95% of global volume.

DeepSeek's stated strategic intent, per Liang Wenfeng's recent public comments, is to evolve the lab \"from a language model brain into a brain that can command physical bodies.\" The Winverse lead is the first concrete dollar deployment behind that thesis — and it doubles as a talent acquisition. Winverse is still pre-product, with no public technical roadmap disclosed, so the bet is essentially a bet on Hu Han.

Hu Han is not a typical seed-stage hire. Tsinghua University PhD (2014), ex-Microsoft Research Asia, ex-Baidu Deep Learning Institute, then Distinguished Scientist at Tencent Hunyuan from January 2025, where he led the vision-language model (VLM) team, the embodied-intelligence foundation-model team, and the Hunyuan speech team. He is a co-author of the Swin Transformer paper, which won ICCV 2021 Best Paper, and was named to MIT Technology Review's \"35 Innovators Under 35\" (China) in 2022.

That track record is what makes the $400M post-money tag defensible for a stealth company with no announced product. It is also why Tencent reportedly remains a possible co-investor in DeepSeek's own round — the same employer Hu Han just left is now adjacent to both sides of the bet.

Winverse is being priced at roughly 1/35th of Skild AI on paper, despite sitting in a market where Chinese manufacturers shipped ~80% of the world's humanoid robots in 2025 and H1 2026. Two factors explain most of the gap: Winverse is at seed-stage with no product, while Skild closed a $1.4B Series C in January 2026 with a SoftBank-led consortium; and Chinese AI startups continue to be valued off domestic capital pools rather than the global lead-investor multiples Skild commands. The TARS Pre-A at $455M in June 2026 set the prior Chinese sector record — Winverse's $400M post-money is consistent with that range, not a premium.

Embodied AI stacks three layers. Winverse sits at the top — foundation models for perception, planning, and decision. Below it, the hardware bodies: Unitree (now a Shanghai-listed Yushu Technology at 688836 after a +460% IPO debut in August), XPeng's IRON humanoid which walked off an automated Guangzhou production line in September, and UBTech — together with AgiBot, the four shipped the bulk of China's ~25,000 H1 2026 humanoids. Below that, the silicon: Cambricon is targeting 500,000 domestic AI accelerators in 2026, and Hygon Information Technology launched its CPU 1000 series — purpose-built for robotics controllers and industrial edge — on Sept. 22, 2026.

DeepSeek and Meituan are also positioning on adjacent layers. Meituan's Long-Z has been one of the most active Chinese investors in embodied AI: a Series B in D-Robotics (March 2026), a $150M Pre-B in Even Realities (July 2026), a Series B1 in Digua Robotics, and a stake in Moon & AI — eight robotics deals between early 2024 and mid-2025 alone. The Winverse co-investor slot gives Meituan exposure to the brain layer that sits above all of those hardware bets. DeepSeek, for its part, is now positioned across body (Unitree), brain (Winverse), and infrastructure (its own model/data center build-out funded by the $12B round).

Short-term (days to quarters): the immediate watch-item is whether the Winverse round closes at the $400M post-money tag and whether the investor roster expands beyond DeepSeek and Meituan Long-Z — a Tencent participation, given its prior relationship with Hu Han and its $12B-round backing of DeepSeek, would tighten the cross-pollination story. The other near-term catalyst is DeepSeek's IPO, targeted for 2027 at a valuation that could climb past $100B if the round expands to 100B yuan; that listing window will price in the embodied-AI optionality from both Unitree and Winverse.

Long-term (1–3 years): the thesis holds if (a) Winverse ships a general-purpose robot foundation model that actually runs on shipping bodies (Unitree, XPeng, AgiBot) — the lab has not disclosed any technical route, so the first 12 months are binary; (b) Cambricon and Hygon deliver domestic silicon at scale that lets training and inference stay outside the NVIDIA orbit — the Sept. 22 Hygon 1000 launch is a start but a far cry from training-grade accelerators; and (c) Tencent Hunyuan's reorganization stabilizes, otherwise the supply of senior embodied-AI talent leaving that lab becomes the dominant channel for new Chinese brain startups and inflates founder premiums further.