180: The Money Game of Embodied AI: Progress Is Hard to Measure, Revenue Is Being Pulled Forward, and the IPO Race

Summary

• The core unease behind the two reporters’ coverage of “The Money Game of Embodied AI” is that the industry has raised a great deal of money without spending a commensurate amount on technology, turning into a financing and IPO race over “who has deeper pockets.” The hardest evidence: sources say only 8 companies across the industry have compute clusters with more than 1,000 GPUs, and readers responded that even 8 was too many; one embodied-AI founder said no one in the industry had run a model even once on more than 100,000 hours of data. One highly valued company spent about RMB30M on R&D last year but had raised more than RMB5B in total—“you could put all that money in Yu’e Bao and make RMB50M a year,” enough to cover the R&D bill through investment income alone.
• Progress is impossible to observe, creating the conditions for a bubble as the industry swings between being overhyped and underappreciated. There is no accepted benchmark, and hardware inconsistency makes replication difficult. After visiting UBTECH 3 times, 李梓楠 still “couldn’t feel any obvious progress”; a company less than a year old took 70 seconds to have a robot pick up a bearing and place it in a box, while a courier “might not even need 2 seconds.” Based on 李梓楠’s tracking, Elon Musk’s Optimus Gen 3 may be about 18 months behind its original design-finalization schedule, and the requirements were lowered along the way: “I’m often late, but I usually get it done.” The question is whether everyone can afford to wait.
• Revenue is being pulled forward by the Hong Kong market’s 18C threshold: local-government data-collection centers were the industry’s largest customer last year, accounting in 徐煜萌’s rough estimate for at least 30-40% of the total market. The model: local governments and robot companies form joint ventures to build data-collection centers; 100 robots at RMB400-500K each generate roughly RMB50M of revenue, and “4 such centers are enough to meet the 18C revenue requirement.” But the tide is already turning: a central provincial capital spent RMB80M on robots, then saw its data center break up early this year; the robots were sold to nearby schools as teaching aids. As ego data and human data become more important, demand for real-robot data collection may decline.
• Manufacturing-related circular transactions are the second revenue channel: robot makers and component suppliers become one another’s suppliers and customers, while the latter also invest in the former, partly because the valuation multiple changes from “15x with a bit of auto exposure to 30x as a robot company.” Nearly 40% of one embodied-AI startup’s revenue comes from manufacturing; several hundred suppliers each buying a small number of robots can add up to significant volume. Changsheng Bearing sold only about RMB8M of components to Unitree and others, yet its stock may have risen 5x from the trough; one board secretary told an analyst directly, “I’m no longer an auto company. I’m at least a 30x company.”
• Actual sales are far below the narrative: after canvassing more than 80 buy-side firms one by one, a foreign analyst estimated that domestic humanoid-robot sales were about 8,000 units at the end of last year, with Unitree and AgiBot far ahead and everyone after third-place UBTECH still in the hundreds. Musk said Tesla would build 300 Optimus units in a single month by August, or 3,600 annualized—30x below the earlier “100,000 units in 2026” target. By contrast, AgiBot’s Deng offered guidance of 16,000 units based on last year’s 4,000-5,000 units of production, which is not especially aggressive. A secondary-market contact’s dangerous line is worth recording: “In a To B business, you can sell as many as you want.”
• The new primary-market play is to preassemble the syndicate through “table-setting” and Club Deals, lining up 6-8 T0 industrial investors from the outset—AgiBot, X Square, Leju, Galbot and others—alongside 10-15 T0 VCs including Sequoia, Hillhouse, Shunwei, Primavera and Source Code. Each round is broadly “3+3,” with the marquee roster providing cover for subsequent rounds and creating expectations of valuation increases. Hillhouse and Sequoia have co-invested in more deals than during the foundation-model cycle; the investor logic is: “If everyone makes money together, what’s there to reject?” One hot world-model company even required investors to commit to leading multiple consecutive rounds, a demand one contact considered unreasonable and declined.
• The next 2 tests are whether embodied-AI scaling works once leading companies approach the 10-million-hour data threshold—“there may be some answers this year”—and how Unitree trades during its first 6 months as a public company. Once the public market has fully priced it, early investors who pushed the stock higher may exit, leaving a value that is theoretically closer to intrinsic value. The key expectation gap is Wang Xingxing’s public guidance that the industry will see some applications land within 2-3 years: “lower than what industry investors are pricing in, but probably more credible.”

Deep dive

1. The Industry’s Biggest Red Flag: Financing Has Replaced Technology

• 李梓楠 relayed one interviewee’s analogy: “It’s like playing Texas hold’em to see whose pockets are deeper”—who has more chips, who commands the higher valuation, who gets to market first. “We’re not reporting on a technology vertical. We’re reporting on something driven by financing.” One source even asked him how many people in the industry were still seriously doing the work.
• The spending evidence is the most jarring. One interviewee said only 8 companies across the industry have compute clusters with more than 1,000 GPUs; after publication, some readers responded that even 8 was too many. In April this year, a secondary-market source relayed an embodied-AI founder’s claim that no one in the entire industry had run a model even once on more than 100,000 hours of data—at odds with the public claims of many companies.
• Data supply is just as thin. A data company that calls itself one of the industry’s top 2 produced a cumulative 50,000 hours of data in the first half of this year, with its largest signed order totaling 10,000 hours. 程曼祺 added an important counterpoint: many full-stack companies collect and consume their own data, and some even sell data externally, as AgiBot does. Adding up supplier volumes is not the same as measuring the industry’s total.

2. Progress Is Unobservable, So the Industry Swings Between Bubble and Undervaluation

• 李梓楠 pointed out that large language models have public benchmarks, open weights and broad accessibility, while embodied AI is To B, expensive and physical—ordinary people cannot interact with it. 程曼祺 said the contrast with large language models is especially stark. The same model can perform differently on different hardware because of consistency problems, making replication harder and more expensive. The result: the industry can be massively overvalued or undervalued, with ample room for storytelling.
• 李梓楠’s field experience was blunt: after 3 visits to UBTECH, he still could not see obvious progress. At a company founded less than a year earlier, a demonstration took 70 seconds for a robotic gripper to pick up a bearing and put it into a plastic box. When the editor-in-chief asked how many BYD workers it could replace, he had no answer—a courier turning around to put something down “might not even need 2 seconds.” His conclusion: “If I were an investor, I probably wouldn’t invest.”
• The wavering conviction also gets wrapped in polished language. When a founder changes technical direction, the public explanation is “I iterated and upgraded,” not “I switched to a different route.”

3. Even Musk Overestimated the Slope: Optimus Gen 3 Is 18 Months Late

• 李梓楠’s regular tracking produces a reverse-engineered timeline: under Musk’s original comments, Optimus Gen 3 should have finalized its design in Q1 of last year. Based on 李梓楠’s own follow-up, it may not have reached that point until the end of June this year—at least 18 months late—and, in his understanding, some requirements were lowered along the way. Tesla went from posting photos and videos weekly to going quiet for long stretches; his judgment is that Musk himself must have overestimated the slope of technical progress. Musk’s defense remains on the record: “I’m often late, but I usually get it done.”

4. Valuation Logic Has Turned Upside Down: Too Low a Price Is Seen as Proof You’re Weak

• 徐煜萌 heard a baffling version of the market’s logic: an academically strong team with a few additional hires can quickly reach a RMB500M or RMB1B valuation, and “if you quote too low a price now, everyone will think you’re not actually a good company.” RMB500M has become the market’s basic floor.
• 李梓楠 challenged it on the spot: “I genuinely don’t know what kind of company is worth RMB500M in its first round.” His impression is that the starting point is at least RMB2B; a company valued at $1B from day one should have founders and a team with some standing in the industry. The approach described by one contact was to ask mechanical and automation professors at a particular university, as well as auto executives at large companies, one by one: do you want to start a company?

5. R&D Spending Comps: LatePost, Mixue Bingcheng and Yu’e Bao

• The newsroom’s self-deprecating comparison has become an industry footnote: one highly valued robot company spent about RMB30M on R&D last year despite raising more than RMB5B—“LatePost spends more than that on R&D,” and “you could put all the money in Yu’e Bao and make RMB50M a year.” The interest on the financing alone could cover the R&D bill. Unitree’s publicly disclosed R&D spending last year was RMB145M, compared with Mixue Bingcheng’s RMB105M. Based on nonpublic information, 李梓楠 estimates AgiBot’s R&D spending was probably higher, perhaps around RMB500M, which may also reflect its roughly 1,300 employees.
• 李梓楠 believes some Chinese embodied-AI companies lack the ability to make technical bets—to decide which direction deserves the wager. 程曼祺’s core question is similar: how should R&D spending map to technical progress? AgiBot’s parallel bets across multiple directions are a resource-intensive strategy with its own logic, but a classic technology-driven company should validate a strong technical view step by step, moving from demo to product prototype to something usable.
• 李梓楠 asked whether people without technical judgment who spend money indiscriminately are not “just dying faster.” 徐煜萌’s response: they could simply raise less money, because capital has a cost wherever it sits; these financings also generally include repurchase agreements, so companies need to consider whether they can cover the obligation.

6. Spring Festival Gala Economics: If Everyone Wants on the Gala, That’s the Problem

• Investors’ criticism of low R&D spending is rooted in comparison. Robot companies have been happy to spend on advertising: Unitree, Songyan Dynamics and Galbot each appeared on the Spring Festival Gala, at a cost of roughly at least RMB60M per company. Against the RMB30M R&D budget cited above, one Gala appearance would fund 2 years of R&D. The first company to advertise on the Gala may benefit; everyone doing it may not. 徐煜萌’s contact put it succinctly: “There’s nothing wrong with going on the Spring Festival Gala. The problem is when everyone wants to.”

7. The IPO Race and 18C: Revenue Has to Be Pulled Forward

• Once a company raises serious money, it has to produce an answer; the immediate competition is the IPO. Apart from the already-listed Unitree, investment banks are confirmed to be working with at least 10 companies, and more are considering listings. Hong Kong’s 18C regime was first proposed in October 2022 and took effect in March 2023, requiring revenue above HK$250M. With robots averaging RMB200-300K, reaching that revenue level is “not especially difficult,” but the queue is long, so the goal becomes ever higher revenue and prettier growth numbers.
• That creates the paradox: technical maturity and real-world adoption remain low, but companies must manufacture large-scale revenue. The next sections lay out the industry’s various solutions to that contradiction.

8. Last Year’s Biggest Customer: Local-Government Data-Collection Centers

• The model was “really complete”: local governments contributed cash or land and formed joint ventures with robot companies to establish data-collection companies; third-party operators hired college students or young people to collect data and perform teleoperation. One robot paired with 2 people, so 100 robots created 200 jobs. 李梓楠’s reaction: “For a local government, this industry is really too good”—no pollution, concrete employment, predictable returns, and far more manageable construction spending than building an auto plant.
• The numbers explain the appeal. Robots sold to data-collection centers cost roughly RMB400-500K each; a typical center bought 100, generating about RMB50M, potentially plus construction costs. 程曼祺 did the math on the spot: “Put together 4 such centers and the company has enough revenue for 18C.” By the end of last year, public information identified more than 80 centers, including in Zigong and Panzhihua. 徐煜萌’s rough estimate is that the business represented at least 30-40% of the overall market. Robot companies would typically also agree to prioritize purchases of the data produced by a center, or to buy it within a specified period, giving the local investment some protection.

9. The Data-Collection Retreat: An RMB80M Investment Unwinds, and Robots Become Teaching Aids

• The rollout did not match the design. Utilization was a problem: if the data could not be sold, the center simply did not operate. A central Chinese provincial capital bought RMB80M worth of robots at the beginning of this year, but the data-collection center “effectively broke up”; the robots were later sold to nearby schools as teaching aids. 徐煜萌’s assessment of the loss: “Financially, it may be a loss. But from a learning perspective, it may not be… You now understand the depth and limits of the business.”
• The deeper pressure is a shift in technical direction. Companies may increasingly rely on more ego data and human data, reducing the need for real-robot collection. Some robot companies themselves believe the business may not be sustainable and are proactively reducing its share of revenue. 徐煜萌 compared the centers with the earlier data-center buildout: companies bought a lot of V100s, then 2 years later switched to H-series, B-series or newer products, leaving the previous generation’s investment potentially unrecoverable.
• The lack of convergence in technical routes is both a risk and the reason new companies continued to appear from 2023 through 2025. In large language models, the startup window had largely closed by the second half of 2023.

10. The Research Market: Little Revenue, but a Possible Ticket Into Infrastructure

• Unitree’s more clearly identifiable revenue comes from research institutions, schools and reception or tour-guide applications—the least disputed part of the industry because the demand is real and the value is clear. 徐煜萌 believes the research market is a viable place to start.
• 李梓楠’s response to the claim that the research market is too small: if the people at the technological frontier use your hardware, that proves the hardware is likely to support the most advanced technology in the future, potentially making you one of the infrastructure providers for the field. Unitree should not be judged only by how much revenue research customers generate.

11. Manufacturing Circular Trades: Supplier, Customer and Shareholder to One Another

• The mechanism is straightforward: a robot company sells robots to a manufacturing-components company; the components company sells parts back to the robot maker and may also invest in it. Nearly 40% of one embodied-AI startup’s revenue comes from manufacturing. A robot company has to connect with at least several hundred suppliers, so small purchases across many suppliers can generate significant volume: if 100 suppliers each buy 20 robots, that is 2,000 units. But 20 units “really can’t do much”; they are generally bought by R&D staff for disassembly and testing or placed on production lines for trials.
• The manufacturing companies’ math is simple. Rather than spend RMB1B building a plant, they can spend RMB500M investing in a robot company and enter a new industry more naturally. The change in public-market multiples is another incentive: “With a bit of auto exposure, the market cap multiple may be relatively low; as a robot company, the valuation can be different”—perhaps 15x for auto and 30x for robotics. Changsheng Bearing is an extreme example: robot companies including Unitree bought only about RMB8M of its components, yet its stock may have risen 5x from the trough. 李梓楠 personally heard one board secretary tell an analyst: “I’m no longer an auto company… I’m at least a 30x company.”
• 李梓楠 believes Nvidia, OpenAI and other circular transactions can also be understood through similar logic, as long as the transactions ultimately create real value and close the loop. Better to redirect underutilized auto capacity toward robot components than leave factories idle and workers on furlough.

12. Actual Sales: An 8,000-Unit Industry and a 30x Miss

• The robot industry does not have automotive sales data accurate to the individual unit every week. A leading foreign analyst worked through company announcements, then contacted more than 80 buyers one by one, piecing together an estimate of roughly 8,000 humanoid robots actually sold domestically by the end of last year. Unitree and AgiBot were far ahead; after third-place UBTECH, companies were still selling at the hundreds level. These are actual sales, not a word game around channel “shipments.” On this basis, Chinese robot companies were already “unambiguously ahead globally” on the production side; Tesla may have produced only a few hundred units last year.
• 李梓楠 considers Musk the single biggest person responsible for the industry’s excessive expectations. The earlier claim of 100,000 units in 2026 made suppliers reach for their calculators. By August, according to 李梓楠, Musk was producing 300 units a month, or 3,600 annualized—30x below 100,000, “a super-large miss.” AgiBot’s Deng, by comparison, guided to 16,000 units on a base of 4,000-5,000 produced last year, “not especially aggressive growth.” 李梓楠 also believes Deng’s credibility record is better than Musk’s.
• A secondary-market contact offered the “terrifying line”: “In a To B business, you can sell as many as you want—as long as 2 companies agree.” 徐煜萌 said that is indeed true from first principles; 李梓楠 called the statement “pretty dangerous.”

13. The Apple Chain Becomes the Tesla Chain Becomes the Robot Chain: The Second Generation’s New Business

• 徐煜萌 shared an elevator anecdote. At one Apple-supply-chain company, “the boss’s son-in-law now runs the robotics business.” At another auto-parts company, the owner’s son had just graduated from Carnegie Mellon with a degree in robotics, and his father gave him money to run the robotics business. At least among large auto-parts and manufacturing companies in Zhejiang, handing the robotics business to the second generation is “quite common.” The progression is clear: part of the 3C Apple supply chain became the Tesla supply chain, and over the past 3-4 years some Tesla-chain companies have become robot suppliers. Some Tesla suppliers in Ningbo, Jiaxing and Jiangsu began making robot components as early as 2022.
• Both speakers consider this “one of the more real parts” of the industry. The upstream supply chain is being pulled forward by downstream demand, and Chinese manufacturing is genuinely strong in assembly and precision machining—even if end demand, from consumers and industrial or commercial users, has not yet reached the point of meaningful robot adoption.
• There are limits to how large a company can be for this ecosystem to matter. No one would call CATL a robotics company just because it bought 10,000 robots; investors would still focus on its core business. The better fit is a mid-sized supply-chain company with roughly RMB30B of market value: if the robotics business doubles its valuation multiple, market cap could rise from RMB30B to RMB60B. Sanhua Intelligent Control is the reference case: before its partnership with Optimus, it may have been worth just over RMB100B; today it is a RMB200B company.

14. AgiBot’s Distributor Tiers: The VAP System and “Self-Purchases Count”

• AgiBot grades distributors by annual sales: RMB20M qualifies as a VAP, RMB10M as gold, RMB5M as silver, and RMB2M as the lowest certification tier. Higher tiers receive more inventory and support. At the time, AgiBot’s highest-tier distributor partners included Joyson Electronics, Ningbo Huaxiang and Wolong Electric Drive—all stocks that performed well in the robotics segment last year. Distributors also proactively helped develop use cases, such as putting robots in battery factories to move batteries.
• The key detail: buying RMB10M of robots for yourself still counts as helping sell RMB10M. “It doesn’t care where these things ultimately go. You created that much sales for it,” and RMB20M qualifies as a VAP. The win-win loop has another layer: a components company can invest in the robot maker while buying its products. “I spent RMB50M buying robots, but the investment I made in it may have gained RMB50M, so the return came back from somewhere else.”
• Asked what he would do as an investor, 李梓楠 answered: “I would trade the swings.” The industry has no clear earnings anchor, so the cycle moves with expectations as they are continually revised. Some companies now have robotics revenue, but whether that revenue is sustainable remains to be seen.

15. Syndication, Club Deals and Multi-Round Lead Commitments: The Primary-Market Reshuffle

• 徐煜萌 was “genuinely shocked” when he heard how the syndicates were assembled. An FA or active institution lines up the roster from the first round: 6-8 T0 industrial investors, including AgiBot, X Square, Leju and Galbot, plus 10-15 T0 VCs, including Sequoia, Hillhouse, Shunwei, Primavera and Source Code. A hot project typically has roughly “3+3” in each round. Industrial investors signal relatively concrete orders; VCs signal technical validation. The marquee syndicate provides cover for future rounds and creates expectations of rising valuations.
• Club Deals are more common than during the foundation-model era. Public information shows Hillhouse and Sequoia co-investing in the same rounds of 它石, 无界, 铭感智能 and 摩感科技; they are also both shareholders in 星尘智能 and AgiBot. Asked why competitive boundaries seemed to have disappeared, the answer was: “If everyone makes money together, what’s there to reject?”
• One highly valued world-model company even required investors to commit to leading multiple consecutive rounds. 徐煜萌’s contact considered the demand unreasonable and ultimately passed. Some T0 institutions selectively reject Club Deals; what they value most in those cases is still “the founder’s resolve.”

16. The Bubble’s Fifth Step: Validating Scaling and Unitree’s Six Months

• The article’s closing suspense was brought into the podcast: the first 4 steps are questioning the bubble, understanding the bubble, embracing the bubble and enjoying the bubble. What is the fifth? After publication, 徐煜萌 felt some people were taking schadenfreude too far—“there’s no need for that.” Secondary investors are doing the work because entry prices are already high, asking highly specific questions: how much will R&D spending be this year and next, where will revenue come from, what orders are on hand, and how likely is repeat purchase?
• 李梓楠 argued that if a company could answer those questions completely, it might not be a typical embodied-AI company—perhaps it would even “be fake.” 徐煜萌 pushed back: not necessarily. Some companies carrying the embodied-AI label have relatively mature businesses, while a true embodied-AI company is fundamentally exploring technology and developing large models, making these questions intrinsically difficult to answer.
• The 2 issues that really matter are, first, technology. Some companies, including AgiBot, have said that 10 million hours of data may be enough to validate whether a model scales. Leading companies may not be far from that target, so “some of the technical questions that were unclear before may have answers this year.” Second is Unitree’s price performance during its first 6 months after listing. Once the public market has fully priced the company, early investors who pushed the price higher may gradually exit, leaving a value that is theoretically closer to intrinsic value. “Everyone wants to know whether it really reaches RMB300B in 6 months, or ends up below where it is now.”
• Wang Xingxing is the reference point for expectation management. He has said publicly that the industry will see some applications land within 2-3 years—“lower than what industry investors are pricing in, but probably more credible.” Investors fall into 2 camps: those making money from the theme and looking to exit quickly, and those who believe in embodied AI’s long-term value and periodically check whether founding teams are seriously building the technology. The latter are less anxious about the IPO timeline. Ultimately, whether a company can list—and whether anyone will want to buy it—still depends on whether its products and technology can demonstrate clearer value.