DeepSeek's First Robot-Brain Bet Is Priced at $400 Million. Skild AI Is Worth $14 Billion.

DeepSeek is set to lead the first funding round for Winverse AI, a robot-brain startup only a few months old, at a valuation of roughly $400 million after the investment, per a Bloomberg report carried by Investing.com (https://www.investing.com/news/stock-market-news/deepseek-to-lead-funding-round-for-robotics-ai-startup-winverse-at-400-mln--report-93CH-4944293). Skild AI, one of the two US companies Bloomberg uses as the benchmark for this layer, raised a Series C in January at a valuation of over $14 billion (https://www.skild.ai/blogs/series-c), about 35 times the Winverse figure (arithmetic). A robot brain, in the sense Skild and Bloomberg use it, is a foundation model that controls a robot: it takes in camera and sensor data and puts out the commands that move arms, legs and grippers. Bloomberg's report makes the distinction plainly. Making machines smart enough to act in the world "requires different AI models from those popularized by OpenAI's ChatGPT or DeepSeek," and "China leads the world in making the arms and legs of humanoid robots."

The same day the Winverse news broke, Unitree founder Wang Xingxing sat on the opening panel of the China International Industry Fair in Shanghai and, in Jiemian's account, said the biggest mistake he made over the past ten-plus years was not believing in AI enough (https://www.jiemian.com/article/15179916.html). He runs the company that sold roughly a third of the world's humanoids in 2025 (https://restofworld.org/2026/unitree-china-humanoid-robot-shanghai-ipo/). Our view: per dollar of revenue, the market now pays more for a robot's brain than for its body; China leads on bodies and has priced its brains at a fraction of US ones; and DeepSeek's two robot checks are small enough relative to its own fundraise to read as an option rather than a commitment. China sells the arms and legs. The market prices the head.

The price list, brains above bodies

The table sorts six companies by what they sell. "Brain" means the company's product is the model. "Body" means the product is hardware. Two Chinese firms sit in between because they build models and robots together. Prices are the most recent figures in the sources; revenue is whatever the company or a reporter has disclosed, and the fourth column is thin on purpose.

Add up the two US model-only companies at their last announced prices and you get $19.6 billion. Add up the two Chinese model-first companies, X Square and Winverse, and you get $3.2 billion, a ratio of about 6x (arithmetic). Use the $11 billion figure from Physical Intelligence's reported talks instead and the US pair rises to $25 billion (arithmetic). Galbot sits outside both sums because Caixin describes it as a humanoid firm, which puts it closer to Unitree's row than to Skild's.

What the US premium is buying: $100 million of recurring revenue and one company with no timeline

Skild's number is the one that makes the brain premium look less like a mood. In September the company said it crossed $100 million in annual recurring revenue ten months after its first commercial deployment (https://www.skild.ai/blogs/skild-crosses-100m-arr). It reports 60-plus paying customers, $50 million already recognized, and a mix in which about 90 percent of revenue comes from manipulation tasks. Mobility is 10 percent, and within that the AMR solutions that came with the Zebra and Fetch Robotics acquisition account for 4 percent. With NVIDIA and Foxconn it is deploying the Skild Brain on dual-arm manipulators assembling NVIDIA Blackwell systems. Its own line on the moment: "The era of demos is over; the era of deployment has begun."

Price that against the revenue and the brain premium is still steep. At the January valuation, Skild trades at about 140 times its September run rate (arithmetic). The company had raised more than $2 billion to date by January, per CEO Deepak Pathak, and the $14 billion mark came about seven months after a $4.5 billion round, so the valuation more than tripled in that window. The revenue grew too. In January, Skild said live revenue had gone from zero to about $30 million within a few months of 2025; by September it reported $100 million of ARR. Those are different measures, but the direction is not in doubt. That $100 million is 40 percent of what Unitree booked for all of 2025 (arithmetic), and Unitree ships metal.

Physical Intelligence is the other half of the US premium and it has disclosed nothing on revenue. TechCrunch, citing Bloomberg, reported in March that the company was in early talks to raise about $1 billion at a valuation exceeding $11 billion, which would double the $5.6 billion mark it set in a $600 million round four months earlier. It had about 80 employees in January. Co-founder Lachy Groom said there was no timeline for commercialization and added, "There's no limit to how much money we can really put to work." Bloomberg's October report still uses the $5.6 billion figure, so the table does too.

So the US pair is one company with a fast-growing but still small revenue line and one company with none. The market is paying for the model's generality and for the belief that the deployment curve Skild describes will keep going. That is a bet on software economics inside a hardware industry, and it has not been tested by a full year of anyone's audited results.

What China's body leader is worth, and why it is spending half its IPO on a brain

Unitree closed on October 13 at RMB 436.63, for a market cap of CNY 176.60 billion and a trailing P/E of 302.28, per Longbridge. The ifeng analysis of its filings puts the first-day intraday high at RMB 1,100 and the peak market cap at RMB 444.9 billion, so the stock sits about 61 percent below its first-day high. Even after that fall, the company is worth about 103 times its 2025 revenue (arithmetic). For a company that sells robots at a 56 percent gross margin, that is a brain multiple attached to a body business.

The operating numbers are mixed. First-half 2026 revenue was RMB 1.152 billion, up 48.5 percent year over year, but non-recurring-adjusted net profit fell about 19.3 percent to RMB 244 million and gross margin slid from 60.4 percent in 2025 to about 56 percent. Smart Analytics Global, cited by ifeng, puts first-half global humanoid shipments at about 19,100, with AgiBot at roughly 8,400 units, or about 44 percent, and Unitree second at about 5,900, or 31 percent. Unitree's own 2026 target is 10,000 to 20,000 units, which means the second half has to deliver between 4,100 and 14,100 (arithmetic). A range that wide is a guess about demand, not a production plan.

Where the IPO money goes is the more telling line. The IPO issued 40.45 million shares at RMB 150.80, roughly RMB 6.1 billion (arithmetic) (https://technode.com/2026/08/07/deepseek-takes-rmb141-million-strategic-placement-in-unitree-ipo/). The largest single project in the use of proceeds is intelligent robot model R&D at RMB 2.022 billion, nearly half the total. Ifeng notes that the market has questioned Unitree's "brain," and that some products are said to need remote-control assistance. The body leader is spending its public money trying to grow a head.

Wang's own account of that effort is the most useful thing in the Jiemian piece. After GPT-6's release, he said, Unitree ran it on task tests across different robots and the results were "not bad." If GPT-7 comes out next year it might "break the embodied-intelligence problem wide open," with a tipping point possibly within one to two years, though he was careful to call that a judgment about model progress and not a commercialization date. He described robot model development as still fairly "primitive": engineers read papers, write models, hand them to testers, and iterate by hand. First-person human data has value, in his telling, but "the last few millimeters" of contact still need real-robot data, which is the same tension we traced in the data-vendor market at https://roboterminal.app/posts/robot-data-sellers-vs-unitree. His architecture keeps real-time motion control on the robot and pushes complex decisions to the cloud where possible.

Our read of those statements: the brain Unitree needs may come from a general-purpose model lab as much as from its own RMB 2 billion program, and its founder is willing to say so in public. That is an unusual admission from a company priced at 103 times revenue. We think it is also an honest one.

DeepSeek's two robot positions, sized

DeepSeek now has one position in a body and one in a brain, and both are small against its own balance sheet. Start with the body. In the Unitree IPO, Hangzhou DeepSeek AI Basic Technology Research took 933,900 shares in the strategic placement for about RMB 141 million, with a 36-month lockup. That is 2.31 percent of the 40.45 million shares issued in the IPO (arithmetic) and about 0.23 percent of the 404.46 million shares outstanding (arithmetic). Quartz's description of a "2.31% strategic stake" in Unitree is loose; the figure is a share of the stock issued in the IPO, not of the company (https://qz.com/deepseek-winverse-ai-robotics-seed-round-101226). At the October 13 close the shares are worth about RMB 407.8 million, or about $60.8 million, roughly 2.9 times cost on paper (arithmetic), and it cannot be sold until about August 2029. TechNode notes that strategic investors were selected partly for potential cooperation in large models, embodied intelligence and robotics, so the shares came with a relationship.

Now the brain. Bloomberg describes the Winverse seed as "worth tens of millions of dollars," with Meituan's Long-Z Investments among the other committed investors. DeepSeek, per the same report, is close to securing at least 80 billion yuan, about $12 billion, in a round mainly meant for model training and data centers, ahead of a planned market debut in early 2027. Even at $99 million, the very top of "tens of millions," the Winverse check would be under 1 percent of that raise (arithmetic). Bloomberg still calls it one of DeepSeek's biggest startup bets since Liang Wenfeng founded the company in 2023.

Who DeepSeek is backing matters as much as how much. Winverse was created in 2026 by Hu Han, most recently lead scientist and head of Tencent's vision-language and embodied foundation model teams, with a Tsinghua PhD and prior stints at Microsoft Research Asia and Baidu. His departure coincided with a restructuring of Tencent's AI research team. Bloomberg places him alongside other researchers leaving big Chinese tech to start labs, including Justin Lin, formerly tech lead on Alibaba's Qwen, and Tian Keyu, a former ByteDance intern. Hu Han, DeepSeek and Long-Z did not respond to Bloomberg's requests for comment, per Quartz.

Our view: a $60 million paper stake locked for three years plus a seed check under 1 percent of a $12 billion raise is a watching brief. DeepSeek is buying a seat near the body leader and a seat next to a well-credentialed brain team, and it is spending almost nothing to hold both. If Wang is right that a GPT-7-class model breaks the embodied problem open within two years, one obvious candidate to supply that model in China is DeepSeek itself, and the Winverse stake is a cheap hedge on the possibility that a dedicated robot model wins instead.

Objection docket

The Winverse number is not a price yet. Bloomberg is explicit that the deal has not been finalized and that valuation and investor participation could change. Until the round closes, the 35x gap to Skild is a comparison between a closed Series C and a rumor.

Skild's revenue includes an acquired robot business. Mobility is 10 percent of revenue and 4 percent of the total comes from the AMR business acquired from Zebra and Fetch. We would still call the 140x multiple a brain multiple, since about 90 percent of the revenue is manipulation, but the pure-model share of the $100 million is a bit smaller than the headline.

Physical Intelligence is pre-revenue, so any multiple on it is undefined. Its inclusion in the $19.6 billion US total inflates the "brain" side with a company whose own co-founder says it has no commercialization timeline. The honest version of the US-China ratio is to call it somewhere around 6x and accept that most of the numerator is expectation.

Unitree's price is noisy. Only 30.09 million of 404.46 million shares circulate, a float of about 7.4 percent (arithmetic), and the stock has lost about 61 percent from its first-day high. A market cap built on that float can move a long way on little volume, so the 103x figure is a reading of a thin market, not a settled judgment.

Chinese capital is thinner, which cuts both ways. Bloomberg's own framing is that investment liquidity is thinner in China, where Unitree and its ecosystem constantly look for capital to fund their next phases of research. Lower prices for Chinese brains may reflect scarcer money rather than weaker teams, and Meituan, Alibaba, ByteDance and Xiaomi have each led a round in X Square at some stage. X Square's open WALL-OSS-0.5 release, which it says achieved over 80 percent autonomous completion on four of 17 real-robot tasks without post-training, is the kind of artifact we track at https://roboterminal.app/posts/open-robot-models-download-test.

The bodies are mostly not working yet either. IDC, cited by Jiemian, put first-half 2026 global humanoid shipments near 25,000, up 432.1 percent, with China over 19,000, and said research and education, performance and display, and government data-collection centers still took 69 percent of shipments. Rest of World reported that over 70 percent of Unitree's 5,500 humanoids in 2025 went to research and education. At the same fair where Wang spoke, Sharpa co-founder Xiang Shaoqing described plans to deploy 30,000 robots across restaurants, retail and hotels; Jiemian's reporters, visiting the Sharpa robot store earlier, watched it take about 6 minutes per ice cream against 2 to 3 minutes for a skilled worker. Both the brain multiples and the body multiples are priced on buyers who are, for now, mostly students.

Three numbers to check

1. Whether the Winverse round closes, and at what valuation. Bloomberg's $400 million is a target with terms not final. A closed round with a disclosed size sets the first real price on a DeepSeek-backed robot brain and tells us whether Chinese model-first companies are converging toward X Square's $2.8 billion or staying an order of magnitude below Skild.
2. Skild's next ARR disclosure and a confirmed Physical Intelligence round. Skild reported about $30 million of live revenue in January and $100 million of ARR in September; the next figure shows whether revenue is closing the gap to a 140x price or the price is running ahead of it. If Physical Intelligence announces a round at or above the $11 billion reported in March, the US brain total moves to $25 billion with still no revenue attached.
3. Unitree's second-half shipments against the 10,000 to 20,000 target, and what it reports spending on model R&D. The company needs 4,100 to 14,100 more units to land in its own range, and it has RMB 2.022 billion earmarked for a brain. Where in that range it lands, and whether the model spend shows up as fewer products needing remote assistance, is the clearest test of whether a body company can price itself like a head.

China ships the arms and legs and is about to find out what its own brains are worth. The first bid from the country's best-known model lab is roughly one thirty-fifth of the American one.