Body First, Brain Second: DeepSeek Eyes an Investment in Winverse AI—Liang Wenfeng Moves Embodied AI Up the Schedule

Wednesday, October 14, 2026

Two pieces of news look especially interesting when read together.

One: DeepSeek plans to invest in Winverse AI, a robot startup founded barely months ago. If the financing lands, this company—founded by Hu Han, former head of Tencent’s Hunyuan multimodal LLM—is valued at roughly $400 million. Two: DeepSeek founder Liang Wenfeng has recently been in dense contact with the embodied-AI startup ecosystem, exploring “robots interacting with LLMs and the real world.”

Pull the timeline back a bit, and you’ll see this isn’t Liang’s first “move” in the embodied track. This August, DeepSeek participated in Unitree Robotics’ IPO strategic placement with roughly ¥141 million. From investing in Unitree (the “body”) to investing in Winverse (the “brain”), a clear path emerges: DeepSeek is extending, layer by layer, from LLMs into embodied intelligence. Behind it sits a reexamination of “can language models alone get to AGI” and the technical pressure brought by GPT-6.

In this article, I want to unpack DeepSeek’s move in full: What kind of company is Winverse AI? What is Liang betting on? How to understand the “from body to brain” investment logic? And—in this crowded embodied track—where are the upside and the worries of this deal?

Part One: Who Is Winverse AI?—A Company With “Nothing Yet” But a $400M Valuation

Start with the takeaway: the reason Winverse AI matters is not what it’s shipped so far, but the weight behind the phrase “betting on Hu Han.”

Winverse AI was founded in 2026, barely months old, and is an early-stage company that “hasn’t shipped anything yet.” What it does is develop robot AI foundation models—giving robots the ability to understand environments, make autonomous decisions, and execute tasks. In plain words: it’s building a “brain” for robots—not an integrated robot, not a component, but something closer to low-level infrastructure and model acceleration—a direction highly “in tune” with DeepSeek’s technical route.

And founder Hu Han is a recognized heavyweight in AI academia. He holds a PhD in automation from Tsinghua (bachelor’s, master’s, and PhD all at Tsinghua), spanned Baidu’s deep-learning institute, Microsoft Research Asia as a principal researcher, and joined Tencent in 2025 to head Hunyuan multimodal LLM—a resume cutting across academia and industry. His brightest marker: co-author of the milestone vision paper Swin Transformer—which won the ICCV 2021 Best Paper Award (Marr Prize) and whose architecture remains one of the most widely deployed visual backbone networks in industry.

It’s exactly this “star scientist + top-conference paper” narrative that puts a “has-nothing-yet” company at a $400M valuation. Liang’s move, more than valuing what Winverse ships today, is about pre-positioning a position of highest quality for the “build a robot brain” direction.


┌──────────────────────────────────────────────────────────────┐
│         Winverse AI: a company building robot "brains"         │
│  ┌────────────────────────────────────────────────────────┐  │
│  │ Founder Hu Han: Tsinghua automation PhD / Swin author   │  │
│  │        MSRA principal / Baidu / Tencent Hunyuan head    │  │
│  ├────────────────────────────────────────────────────────┤  │
│  │ Business: robot AI foundation models (embodied brain)   │  │
│  │       Infrastructure & model acceleration, no hardware  │  │
│  └────────────────────────────────────────────────────────┘  │
│  Months old, valued ~$400M (if DeepSeek funding closes)     │
└──────────────────────────────────────────────────────────────┘

Part Two: “From Body to Brain”—Why Does DeepSeek’s Embodied Path Go This Way?

To understand this investment, put it back into DeepSeek’s overall layout.

It starts in August. That month, DeepSeek participated in Unitree Robotics’ IPO strategic placement with roughly ¥141 million, subscribing to 933,400 shares, about 2.31% of the offering. Unitree founder Wang Xingxing addressed the relationship during the online roadshow: per the signed “strategic cooperation memorandum,” Unitree and DeepSeek will cooperate in three directions—collaborative R&D toward general AI, deep cooperation on high-performance general-purpose robots, and deep cooperation on AI foundation models. Back then, the industry already judged: “Model + Hardware” alliances between LLM companies and robot firms will become the mainstream embodied-AI paradigm.

And this time, from Unitree to Winverse, the path becomes clearer—eight words: body first, brain second.

Unitree is a leader in “hardware,” solving the robot’s “body”; Winverse does the embodied “brain,” solving the robot’s “intelligence.” One handles “moving,” the other handles “thinking clearly”—combined, they make complete embodied AI. DeepSeek, as a deep-pocketed, technology-faith-driven LLM company, clearly doesn’t want to be just “the cloud that provides LLMs.” It wants to extend its language-model capability into robot brains that can understand, act, and interact with the physical world.

More intriguingly, this is effectively an “advance” adjustment of Liang Wenfeng’s own AGI roadmap. Liang has publicly laid out a six-step progressive AGI roadmap: language model → chain-of-thought → agent → continual learning → self-iteration → embodied intelligence. Per that plan, embodied intelligence should be the last step. But recent investments in Unitree, leading Winverse, and contacting embodied founders all say Liang is moving the “embodied” step from the end of the roadmap to right now.

Why the advance? The likely answer sits in another variable: GPT-6.

Part Three: GPT-6’s “Disruptive Strike”—Why LLM Companies No Longer Stay Calm

A key variable behind DeepSeek’s embodied push now is OpenAI. This judgment gains support from the data.

Public info shows that in September’s RoboDojo tests, a GPT-6 Astra not specifically fine-tuned, directly serving as an action policy, achieved 22.48% average success across 42 robot-manipulation tasks and 2,100 trials. To be clear: that number beat 40 public robot policies in the evaluation. In other words: a general LLM that “was never robot-trained” already beats most dedicated robot policies out-of-the-box; once fine-tuned on robot data, its ability only grows stronger.

Adding to the industry’s pressure, at the October 12 industrial expo opening, Unitree founder Wang Xingxing declared: GPT-7 could break through the industry’s remaining embodied bottleneck.

It’s a sword hanging overhead. Imagine: if the next-generation GPT, without a separate complex embodied training pipeline, can directly drive robots to complete complex tasks, then the value of many “dedicated embodied models” gets massively diluted, and competitive focus shifts from “algorithms” to “who can get the most powerful general model fastest and turn it into a robot brain at low cost.”

This pressure is pushing DeepSeek and other LLM companies to reexamine a fundamental question: Can you really reach AGI just by scaling language models? Future general AI probably isn’t merely something that outputs text in a dialog box, but a system that can truly understand and interact with the physical world. To understand the physical world, embodied intelligence is very likely an unavoidable vehicle. That’s the deeper logic behind DeepSeek moving up its embodied move.


┌──────────────────────────────────────────────────────────────┐
│   GPT-6's embodied pressure is rewriting LLM-company strategy  │
│                                                               │
│   RoboDojo: unleaned GPT-6 Astra as action policy             │
│     42 tasks / 2100 trials / 22.48% avg success              │
│     → beats 40 public robot policies (disruptive strike)     │
│                                                               │
│   ↓ Industry reexamines:                                      │
│   "Can scaling language models alone reach AGI?"              │
│   Lean conclusion: to understand the physical world,          │
│   embodied intelligence is the unavoidable vehicle            │
│                                                               │
│   → DeepSeek moves "embodied" up in its AGI roadmap           │
└──────────────────────────────────────────────────────────────┘

Part Four: Where’s the Money, How Big’s the Bet?—DeepSeek’s Capital Hand

However grand the strategy, what sustains moves everywhere is ultimately “money.” And DeepSeek’s capital hand is no longer what it used to be.

Reports say DeepSeek will close its latest funding round soon, raised from ¥50 billion to roughly ¥100 billion. The recent cadence has been fierce: an April first round, scale over ¥50 billion, post-money valuation ~¥400 billion, setting a record for the largest single round in China’s LLM industry; a July new round with pre-money valuation rising to ~¥500 billion. Meanwhile, DeepSeek is gearing up for an IPO—it just hired Yan Wentao, former Hony Capital principal, as CFO, widely seen as personnel settling in ahead of a listing.

In other words, DeepSeek is using “massive funding + IPO preparation” as twin engines to sustain expansion across embodied AI and its broader AGI map. Investing in Unitree and leading Winverse are just two stones in a larger game.

And don’t watch Liang alone. Pan out, and the capital map of LLM and internet giants in robotics is already layered:

Alibaba has arguably the widest map, having invested in 26 robotics and related firms since 2023, covering Sudo Technology, Galaxy General, Itself Variable, and more.
Tencent placed bets on two star firms, Zhiyuan Robotics and Unitree.
Meituan’s map covers Dijia Robotics, StarMap AI, Galaxy General, and Unitree.

Zoom out to the whole track: IT Juzi event data shows domestic embodied-AI financing in H1 2026 reached ¥93.5 billion, five times the year earlier; 322 deals, up 137% year-on-year. CCSA reports H1 domestic humanoid shipments exceeded 40,000 units, about 97% of the global share. Under the heat, capital has long since flooded in.

Part Five: Upside and Worry—Two Sides of This Deal

Every investment has two sides. DeepSeek’s planned Winverse bet has both imagination and underlying concerns.

Start with upside. On one hand, if “model + hardware” really is the dominant embodied paradigm, then DeepSeek holding Winverse (brain) + Unitree (body) means it has built a complete “two-wheel” base on the embodied track—brain and body can co-evolve, something many single players can’t do. On the other hand, Winverse leans toward “low-level infrastructure and model acceleration,” in tune with DeepSeek’s tech. If DeepSeek can transplant its hard-won model-efficiency and training-acceleration strength into “robot-model acceleration,” that’s a compelling cost-down ability. In short, what DeepSeek is investing in isn’t just Hu Han the person, but the technical route where “general LLM + embodied brain + body” converge.

Then the worries. The most direct is valuation froth. Some investors say plainly: the frequent billions-scale financings in embodied AI today rest more on “growth-at-any-price (市梦率) faith” than commercial fundamentals. A “has-nothing-yet” Winverse at $400M is a microcosm of that. Second is that the technical route is far from unified—players differ wildly in data, model architecture, and product form; which route will actually work is too early to conclude. For a player like DeepSeek that both invests and considers doing its own R&D, deploying too early or too late can both miss.

Put the two sides together, and my judgment: DeepSeek’s step has the right strategic direction, but essentially it’s buying an option on a pivot that hasn’t fully arrived—“general models take over the physical world.” Whether the option was bought cheap or dear can only be answered when GPT-7 either “breaks the bottleneck” or not, and when embodied AI’s commercial loop truly lands.

Part Six: Skirting the “Big-Firm Path”—DeepSeek’s Dual Role

Why is DeepSeek seen as “the most special kind of player” on the embodied track? Because it plays two roles at once, and the tension between them is itself the point.

First, the “technology evangelist.” DeepSeek built strong influence via openness and “cost-efficient LLMs”; Liang Wenfeng is himself seen as an icon of “technology faith.” When such a technical company reaches into embodied AI, it brings not just money but a “model-efficiency-first” methodology. You can anticipate: if DeepSeek replicates its language-model playbook of “approaching the frontier at lower cost” onto “robot-model acceleration,” it’s very likely to pull the whole embodied industry’s cost-down path along with it.

Second, the “potential direct contender.” There are signs Liang doesn’t just want to “invest” in embodied AI—he may even consider doing robot LLMs himself. If DeepSeek truly flips from “investor” to “developer,” it stops cheering from shore and jumps into the same river as Unitree, Winverse, and countless embodied teams. A DeepSeek that “holds massive funding, wields top-tier model ability, and participates directly” would shake the existing competitive landscape far beyond any single investment.

But this “dual role” raises a question that must be faced: when the evangelist also becomes a contestant, where’s the trust boundary? On one hand, embodied startups need DeepSeek’s models and money; on the other, they must guard against it “acting as both referee and player.” This tension may confront every embodied team dealing with DeepSeek earlier and more realistically than valuation froth. Ultimately, DeepSeek’s final value on the embodied track depends not on how much it invests, but on whether it can use “cheaper, more reliable models” to raise the whole ecosystem’s water level—even if that means diving in itself.

Part Seven: Three Judgments for Observers and Practitioners

Finally, place this deal in the wider industrial context and offer three cool-headed judgments.

First, the “two-way embrace” between LLMs and embodied AI is accelerating, but don’t mistake “a move” for “landing.” DeepSeek’s moves on Unitree and Winverse are hallmark gestures of LLM companies extending downstream; embodied firms likewise actively connect with LLM capability. The fusion is a trend, but between “investment” and “product landing” lie a pile of hard bones: algorithms, data, engineering, safety. Seeing capital position itself, don’t rush to conclude “embodied AI is about to take off.”

Second, “brain” value will be increasingly revalued, but the “brain” needn’t belong only to general LLMs. Embodied AI needs a brain that understands the physical world and decides in real time—the general LLM is one card, but a dedicated/“embodied” brain optimized for robot motor control, safety boundaries, and real-time constraints may be equally key. Founders with backgrounds like Hu Han’s—“top-conference paper + big-firm exec”—are becoming a scarce resource capital fights over; whoever can fuse “general ability” with “physical-world real-time constraints” is more likely to capture the next wave of dividends.

Third, watch out for the harm of “市梦率” valuations to the industry. Embodied AI’s current high valuations essentially pre-price future business models. That’s not bad—it arms truly capable people with ammunition; but if valuations detach far from fundamentals, and an external variable like GPT-7 massively deflates the value of “dedicated embodied models,” the reckoning of froth will come fast. For practitioners, rather than chasing valuations, land the model in real scenes and guard the commercial floor with stable engineering.

From investing in Unitree’s “body” in August, to planning Winverse’s “brain” now, to Liang Wenfeng’s dense contact with embodied founders and thoughts of building his own—DeepSeek, in a string of moves, has pulled “embodied intelligence” from “the last step” of its AGI roadmap to “this very round.” It’s both a technical and a capital ambition, and under GPT-6, an unavoidable strategic position for a top LLM company. The embodied track is already crowded enough, and the real suspense has never been “who has more money,” but “whose brain can first truly understand and tame the physical world.” In this game, DeepSeek has taken its seat.