Manchi | 晚点聊 LateTalk Summary

晚点聊 LateTalk | Episode 180 Guests: Zinan (Tech Reporter, LatePost), Yumeng (Platform & Investment Reporter, LatePost) Host: Manchi

1. Key Themes

The Embodied Intelligence Industry Is a Fundraising Race, Not a Technology Race

The dominant activity in China's humanoid robot sector is not R&D competition but capital competition — who has the deepest pockets, the highest valuation, and the earliest IPO. The underlying technology progress is slow and opaque, creating a vacuum that fundraising narratives fill.

"It feels like we're not reporting on a technology track — we're reporting on something driven by fundraising. What people talk about more is not technology, but fundraising." — Zinan [00:02:59.500]

Industry Progress Is Fundamentally Unobservable, Creating Massive Speculative Space

Unlike large language models — which have public benchmarks, open weights, and accessible products — humanoid robots have no widely accepted benchmarks, are expensive physical objects, and are B2B products. This opacity cuts both ways: the sector can be dramatically overestimated or underestimated with no mechanism to correct.

"The progress of this industry is very difficult to observe. This means it could be greatly overestimated, or greatly underestimated, or just oscillating between those two states. The space for storytelling — or blowing bubbles — is very large." — Zinan [00:06:55.060]

Spending Is Shockingly Low Relative to Capital Raised

Despite raising billions, companies are barely spending on the two most capital-intensive needs: compute and training data. The industry has at most eight companies with clusters above 100 GPUs; one founder told a secondary market contact that no company has ever used more than 100,000 hours of data to train a model.

"A leading company's R&D budget last year was about 30 million RMB. To date, it has raised over 5 billion RMB. If you just put that money in a money market fund, you'd earn 50 million a year in interest — more than the R&D spend." — Zinan [00:13:59.680]

"One company said it had cumulatively produced 50,000 hours of data in the first half of this year, and claimed to be a top-two player. Its largest single order was 10,000 hours." — Zinan [00:04:57.400]

Data Collection Centers as Government-Subsidized Revenue Laundering

The single largest revenue source for the industry last year was local government-backed data collection centers — a circular structure where robot companies sell robots to these centers, which produce data that robot companies then contractually agree to buy back. This creates apparent revenue, employment, and government buy-in, but the underlying economics are fragile.

"Last year, the biggest customer for the whole industry was data collection centers set up by local governments everywhere. The robot company sells robots to the data collection center, and the data collection center sells data back to the robot company — so they are each other's supplier and customer." — Zinan [00:19:51.900]

"There are roughly over 80 such data collection centers publicly identifiable as of last year." — Zinan [00:23:11.160]

"I'd estimate this segment accounts for at least 30–40% of total revenue across humanoid robot companies." — Zinan [00:24:06.640]

The Circular Economy Between Robot Startups and Manufacturing Supply Chain Companies

A significant portion of revenue — up to 40% for some companies — comes from mutual purchasing arrangements with manufacturing supply chain companies. Robot companies buy components from manufacturers; manufacturers buy robots from robot companies; manufacturers then become investors in robot companies. The motive is partly operational and largely financial: changing a manufacturer's stock market label from "auto parts" to "robotics" can double or triple its valuation multiple.

"Robot companies sell robots to manufacturing component companies, and component companies sell their components back to robot companies — so they are each other's supplier and customer. It's a circular relationship." — Zinan [00:29:48.860]

"If you're labeled as an auto company, your market cap multiple might be 15x. But if you're a robotics company, your multiple might be 30x. For a company that wants to raise money or build reserves, getting into the robotics business is very cost-effective from every angle." — Yumeng [00:30:46.380]

The IPO Race Is Driven by Bull Market Window Anxiety, Not Readiness

The urgency to IPO is not primarily about technology milestones — it's about a perceived closing window. The current bull market has lasted over 20 months from September 2024, which is already long by historical standards. At least 10 companies have investment banks actively working on their listings.

"We've now had over 20 months of bull market since September 2024. By A-share history, this is already a relatively long bull run. People feel: if I don't catch this window, and the market turns down when I try to list, my valuation gets a haircut — or I might not be able to list at all." — Zinan [00:49:11.160]

Established Tech Giants (Xiaomi, Xiaopeng, Tesla) Are Not Selling Robots — and That Tells You Everything

The large players with strong cash flows and no IPO urgency are building robots without selling them. This is the most rational behavior given technological immaturity — but it is only available to companies that don't need external validation through revenue.

"Have you noticed that companies like Xiaopeng and Xiaomi are not actually selling their robots? They're not rushing to generate revenue from robots. They already have strong cash flows. They also don't have an urgent need to IPO — they're already listed." — Zinan [00:47:42.700]

Elon Musk's Overcommitments Have Structurally Distorted the Entire Sector's Expectations

Musk's public forecasts — 100,000 Optimus units in 2026, third-generation design locked in Q1 2024 — pulled forward expectations across the entire global supply chain. Actual production as of August 2025 is around 300 units per month, annualizing to ~3,600 — a 28x miss against the 100,000 target. This expectation gap has caused systematic underperformance across supply chain companies.

"Elon's responsibility cannot be escaped — he pulled expectations too high. He said 100,000 units in 2026, and everyone started calculating: if you sell parts to Tesla, multiply your unit price by 100,000, and that's your revenue uplift. Most of those supply chain companies are now reporting earnings misses." — Yumeng [00:45:47.040]

"300 units a month, annualized is 3,600. The target was 100,000. That's a 30x miss." — Yumeng [00:46:14.040]

Wang Xingxing's (Unitree) Conservative Public Guidance Is the Most Reliable Signal

Against a backdrop of rampant hype, Wang Xingxing (CEO of Unitree) has publicly guided for 2–3 year application timelines, which is more conservative than the market is pricing. His 16,000-unit guidance for the year, while modest, is considered credible because it represents a reasonable step-up from last year's ~4,000–5,000 units and his track record is better than Musk's.

"I find Wang Xingxing's public statements the most worth referencing. When Unitree listed, I reviewed his public comments — he was actually managing expectations downward. He said meaningful application deployment might take 2–3 years. That's more conservative than what investors are pricing in." — Yumeng [00:56:59.260]

"Wang Xingxing's promise fulfillment rate is better than Musk's. And 16,000 units is not a wildly aggressive number given they had production of around 4,000–5,000 last year." — Yumeng [00:46:43.600]

2. Contrarian Perspectives

The Circular Revenue Model May Not Be Fraud — It Might Be Rational Industrial Policy in Disguise

Most observers view the government data center model and mutual purchasing arrangements as signs of artificial revenue inflation. But there is a counterargument: these structures are efficiently mobilizing idle industrial capacity (auto parts factories with declining orders), creating legitimate employment, generating real data, and giving both sides optionality. The key question is not whether the structure is circular but whether it produces durable value.

"This is actually mobilizing resources. If an auto parts company's factory is sitting idle because auto sales are down, and it converts to making robot components — that idle capacity is being used. I think as long as these transactions ultimately have real value, it's fine." — Yumeng [00:34:04.180]

Not Spending R&D Money May Actually Be Rational When You Don't Know What to Invest In

The conventional criticism is that companies are under-investing in R&D relative to their fundraising. But there's a contrarian argument: if you don't have a clear "big bet" on the right technical direction, spending heavily is not virtuous — it's just burning money faster. The real problem is the absence of a technical thesis, not the absence of spending.

"The core issue is that some Chinese embodied intelligence companies lack the judgment to decide what direction to bet on. When you don't have that judgment, you can't really invest in R&D properly. Of course, one approach is to try many directions at once — that's a resource-based playbook. But a truly tech-driven company has a long-horizon technical conviction, and builds experiments around that conviction." — Zinan [00:15:27.780]

The "Science and Research" Market for Humanoid Robots Has a Strategic Value That Everyone Underweights

Most analysts dismiss the science/education revenue as too small to matter. The contrarian view: if the world's most technically advanced researchers choose your hardware, you are positioned to become the default infrastructure of the field — analogous to how certain compute platforms became industry standards.

"The science and research market's real value is often overlooked. If the most technically advanced people in the entire industry are using your hardware, that proves your hardware has the best chance of being the most compatible with the most advanced technology in the future. You could become a technical infrastructure provider for the field." — Manchi [00:25:02.900]

Founders Who Appear to "Iterate" Are Often Just Pivoting to Survive — and the Market Rewards This Linguistic Trick

In normal tech markets, a pivot is a red flag. In this sector, the opacity of progress allows founders to rebrand full strategy changes as "iteration" — and investors, unable to verify technical progress, accept the framing.

"This founder had a very subtle way of describing it to outsiders. He wouldn't say 'we switched to a different track.' He would say 'we iterated.'" — Zinan [00:07:54.320]

Club Deals Among Competing VCs Are More Common in Robotics Than in Any Previous Tech Wave — and Signal Risk, Not Strength

When direct competitors (Gaorong, Honshan/Source Code) co-invest in the same rounds at high frequency, it is typically read as a validation signal. The contrarian reading: it reflects that no one has sufficient conviction to lead alone, risk is being socialized, and the deals are structured around financial engineering rather than fundamental conviction.

"In the LLM era, Gaorong and Honshan investing together in the same round was quite rare. In embodied intelligence, it happens far more often. I asked these investors why the competitive relationship seems to have disappeared and why they're willing to collaborate. The answer I got was: 'If we can all make money, why would we object?'" — Yumeng [00:53:33.380]

3. Companies Identified

Unitree Robotics (宇树科技) Leading Chinese humanoid robot company; already gone public. The market benchmark everyone watches. CEO Wang Xingxing has a reputation for under-promising and over-delivering relative to peers.

"The second and third place were Zhiyuan and UBTECH. After that, sales drop to around the 100-unit level. These are real sales figures, not 'shipment' figures." — Zinan [00:35:31.780]

Zhiyuan Robotics (智元) Second-largest humanoid robot company by real sales volume. Notable for building the most systematic sales/channel ecosystem in the industry, including a tiered VAP partner program.

"Zhiyuan has a tiered sales architecture. If you help Zhiyuan achieve 20 million RMB in sales, you're a VAP partner. 10 million is Gold, 5 million is Silver, 2 million is Certified. The higher your tier, the more allocation and support you receive." — Zinan [00:41:23.480]

UBTECH Robotics (优必选) Third-ranked humanoid robot company by real sales. Already publicly listed.

"Third place is UBTECH, and after that it drops to the hundred-unit level." — Zinan [00:35:31.780]

Leju Robotics (乐聚) Humanoid robot company; now classified as a T0 "strategic investor" in the next generation of robotics startups, having transitioned from being a startup itself.

"T0 strategic investors in the current wave include Zhiyuan, Leju, and Galaxy General, among others." — Yumeng [00:51:08.500]

Galaxy General (银河通用) Humanoid robot company operating as both a robotics developer and now a strategic investor in the next wave.

"T0 strategic investors include Zhiyuan, Leju, Galaxy General, and others." — Yumeng [00:51:08.500]

Sanhua Intelligent Controls (三花智控) A-share listed manufacturer, best example of a supply chain company successfully repositioning as a robotics company. Market cap roughly doubled from ~100 billion to ~200 billion RMB after announcing Optimus partnership.

"Sanhua Intelligent Controls is probably the largest listed company in the robotics space on the A-share market, with a market cap around 200 billion. Before its Optimus collaboration, it was probably just over 100 billion." — Yumeng [00:40:53.600]

Changsheng Bearing (长盛轴承) Bearing manufacturer whose stock rose approximately 5x after Unitree's Spring Festival Gala appearance, despite only contributing about 8 million RMB of robot-related parts revenue to Unitree.

"After Unitree appeared on the Spring Festival Gala, people dug into Changsheng Bearing's financials and found that Unitree and other robot companies had only bought about 8 million RMB worth of parts from them. But Changsheng's stock rose about 5x from its trough. The contribution to actual revenue was modest, but the market cap impact was enormous." — Yumeng [00:32:39.680]

Tesla / Optimus Benchmark for global humanoid progress. Running far behind Musk's own public commitments. ~300 units/month as of August 2025, annualizing to ~3,600 against a stated goal of 100,000 in 2026.

"As of August, they've produced 300 units. Annualized, that's 3,600. The target is 100,000. That's a 30x miss." — Yumeng [00:46:14.040]

Meituan (美团 / referred to as Eagle Data — likely "鹰眼数据" or similar) Data company that has raised significant funding in the robotics data collection space.

"A lot of data companies have raised a lot of money this year — Eagle data companies, glove collection companies — their valuations have grown very quickly." — Zinan [00:04:28.220]

BAAI (Beijing Academy of Artificial Intelligence / 智源) Research institution with a humanoid intelligence data business; also sells embodied intelligence data externally. R&D spending estimated at approximately 500 million RMB last year with roughly 1,300 employees.

"BAAI's R&D spending should be considerably more — probably around 500 million last year. They currently have around 1,300 people." — Zinan [00:14:57.760]

Meixue Bingcheng (蜜雪冰城) Bubble tea chain used as a benchmark comparison: its R&D spending last year was 105 million RMB, higher than Unitree's 145 million — cited to illustrate how low humanoid robot R&D spending is.

"We looked up Meixue Bingcheng's R&D spend — it was 105 million RMB last year. Unitree's was 145 million." — Zinan [00:14:28.880]

Junsheng Electronics (均胜电子) Auto parts manufacturer and top-tier VAP partner of Zhiyuan; also supplies components to multiple robot companies.

"Some of the top-tier VAP partners include Junsheng Electronics, Ningbo Huaxiang, and Wolong Electric." — Zinan [00:42:21.980]

Ningbo Huaxiang (宁波华翔) Auto parts manufacturer and top VAP partner of Zhiyuan.

"Top VAP partners include Junsheng Electronics, Ningbo Huaxiang, and Wolong Electric — these were some of the best-performing robotics-related stocks last year." — Zinan [00:42:21.980]

Wolong Electric (卧龙电驱) Electrical manufacturer and top VAP partner of Zhiyuan.

"Wolong Electric is among the highest-tier VAP partners of Zhiyuan." — Zinan [00:42:21.980]

Honshan Capital / Source Code Capital (红山) Top-tier VC; early investor in Unitree from the seed stage (~2019), before any industry consensus. Representative of genuine early-conviction investing.

"Honshan invested in Unitree very early — seed stage, around 2019. At that point there was absolutely no consensus. They just saw Wang Xingxing holding a robot dog he'd built himself, and that was enough." — Yumeng [01:00:21.960]

Gaorong Capital (高榕资本) Top-tier VC; increasingly co-investing alongside Source Code Capital in robotics deals, a notable shift from their historically competitive relationship.

"Gaorong and Honshan investing in the same round was rare in the LLM era. In embodied intelligence, it happens much more." — Yumeng [00:53:03.660]

Shunwei Capital (顺为资本) T0 VC in the humanoid robot space.

"T0 VCs include Honshan, Gaorong, and also Shunwei, Chunhua, Yuanma, and others." — Yumeng [00:51:37.440]

Chunhua Capital (春华资本) T0 VC in humanoid robotics.

"T0 VCs include... Chunhua, Yuanma, and others." — Yumeng [00:51:37.440]

Yuanma (圆马) T0 VC in humanoid robotics.

"T0 VCs include... Yuanma." — Yumeng [00:51:37.440]

TASI / Tashi (踏实) Humanoid robot company co-invested by Gaorong and Honshan in the same round.

"Looking at public information, both Gaorong and Honshan appear in the same round in companies including Tashi, Wujie, Liqing Intelligence, and Mogan Technology." — Yumeng [00:53:03.660]

Wujie Robotics (无界) Humanoid robot company with co-investment from Gaorong and Honshan.

Liqing Intelligence (理清智能) Humanoid robot company with co-investment from Gaorong and Honshan.

Mogan Technology (魔感科技) Humanoid robot company with co-investment from Gaorong and Honshan.

Xindong Jiyuan / Heartward (心动机缘) Humanoid robot company with both Gaorong and Honshan as shareholders.

4. People Identified

Wang Xingxing (王兴星) — CEO, Unitree Robotics The most credible public voice on realistic timelines in the industry. Has consistently under-promised relative to peers. Guided 2–3 years for meaningful application deployment, and 16,000 units for the current year — both considered conservative but credible.

"Wang Xingxing is actually managing expectations downward. He said meaningful application deployment might take 2–3 years. That's more conservative than what investors are pricing in — but probably more reliable." — Yumeng [00:56:59.260]

Elon Musk — CEO, Tesla/SpaceX Cited as the single figure most responsible for inflating the global humanoid robot bubble through serial public overcommitments on Optimus timelines and production volumes.

"His responsibility cannot be escaped. He pulled expectations way too high. He was very confident for a while, then went quiet after realizing the technology wasn't progressing as fast as he thought." — Zinan [00:10:36.980]

Deng Qingfeng (邓清风) — CEO, Zhiyuan Robotics (implied as "邓总") Architect of Zhiyuan's tiered distributor ecosystem and the 16,000-unit guidance. Credited with a better public commitment track record than Musk. The hosts suggest inviting him to explain his channel design thinking.

"We should ask Deng how he came up with this system design. I think this is probably a common approach in industries with high gross margins and complex distribution." — Manchi [00:43:43.400]

5. Operating Insights

Design Your Channel Ecosystem With Financial Incentives That Align Partners' Self-Interest With Your Revenue Growth

Zhiyuan's tiered VAP program is the most sophisticated commercial ecosystem in the sector. By setting clear revenue thresholds (2M/5M/10M/20M RMB) for partner tiers, Zhiyuan converts suppliers, investors, and distributors into active sales agents who proactively develop new application scenarios — without Zhiyuan having to identify them centrally.

"Zhiyuan's top-tier VAP partners include Junsheng Electronics, Ningbo Huaxiang, and Wolong Electric. These partners actively help Zhiyuan find new application scenarios — for example, one suggested using robots to move batteries in battery factories, a use case Zhiyuan's own team hadn't considered." — Zinan [00:42:50.600]

When Evaluating Any Company in an Opaque Sector, Ask the Operational Questions Secondary Investors Are Now Using

As the robotics sector matures, sophisticated secondary market investors are developing a filtering protocol that cuts through narrative: R&D budget this year and next, expected revenue and its sources, current order backlog, and repeat purchase rates. Companies that cannot answer these questions in detail are still in "narrative stage" regardless of their valuation label.

"Secondary investors are explicitly asking: how much do you plan to invest in R&D this year and next? What is your revenue expectation and where does it come from? What's your current order backlog? What's the repurchase probability in this sector? If you can answer these questions, you're a different kind of company." — Yumeng [00:55:30.440]

Spring Festival Gala Spend Is a Useful Diagnostic for Capital Allocation Discipline

Several robotics companies spent ≥60 million RMB each to appear on CCTV's Spring Festival Gala — double the annual R&D budget of at least one company in the sector. The first mover gained real attention; every subsequent company gained less. This is a transferable lesson: any marketing format where the marginal return collapses as more players adopt it should be treated as a signaling game, not a business investment.

"The first company to appear on the Spring Festival Gala probably got real advertising value. But when everyone appears on it together, the effect disappears. Going on the Gala is fine; everyone going on the Gala is a problem." — Yumeng [00:16:55.580]

6. Overlooked Insights

The Fastest Path to a Valuation Re-Rating Is Not Building a Robot Company — It's Buying Robots as a Mid-Cap Manufacturer

The most actionable and underappreciated insight in this episode is the valuation arbitrage available to mid-sized listed manufacturers (market cap ~3–10 billion RMB) who buy robots from startups. They receive components at cost, book the robot purchases as business development, get re-rated from an "auto parts" multiple (15x) to a "robotics" multiple (30x), and simultaneously participate in the startup's valuation appreciation as an investor. The strategy is self-funding: the stock price appreciation may recoup the capital spent on robot purchases.

"If you spend 50 million buying robots, and your investment in that company has appreciated by 50 million because of the sales volume you helped them achieve and the IPO timeline you pulled forward — the money comes back to you from another direction." — Yumeng [00:44:19.200]

This dynamic is most potent for companies with market caps around 300 billion RMB — large enough to credibly enter the space, small enough that the re-rating is meaningful. Sanhua Intelligent Controls (which doubled from ~100B to ~200B) is the clearest proof case.

The "100,000-Hour Data" Scaling Threshold May Be the Sector's Hidden Inflection Point — and It May Be Reached This Year

Almost no one in public discussion is tracking this specific metric, but it is the actual technical gate that matters. Industry insiders believe that once a company accumulates ~10 million hours of data, they can finally test whether embodied intelligence models follow the same scaling laws as LLMs. The most advanced companies are reportedly approaching this threshold in 2025 — meaning this year may produce the first real evidence of whether the entire technical thesis is valid.

"Companies including Zhiyuan have said: if we get to 10 million hours of data, we might be able to verify whether this model actually follows scaling laws. The leading companies are not far from that target now. So some of the technical questions that have been unclear may actually get answers this year." — Zinan [00:58:54.240]

If scaling laws do hold for embodied intelligence, the implications for valuation would be enormous and rapid. If they don't, the correction would be equally severe. This is the single most important binary outcome to monitor in the next 12 months — and it is almost entirely absent from public investor discussion.