Manchi | 晚点聊 LateTalk Summary

1. Key Themes

The "Boring vs. Suffering" Framework for Career Motivation

Yao Song articulates a fundamental life philosophy that explains why high-achieving people leave stable, prestigious paths. He describes visiting Silicon Valley peers who were earning well but comparing how many U.S. states they'd visited, and feeling that this life — however comfortable — was intolerable.

"There was a teacher who summarized it perfectly — he said life has two states: boring and suffering. It's just a question of which you can tolerate less. For me, I cannot tolerate boring." [00:18:07]

This framework helps explain the behavior of an entire cohort of elite Chinese tech founders who left academic or corporate tracks earlier than expected.

The Compounding Value of Information-Gathering Between Ventures

Rather than immediately launching a second company after selling Deephi (深鉴科技), Yao spent three years systematically mapping every adjacent frontier — laser lidar, brain-computer interfaces, controllable fusion, quantum computing, commercial rockets — by joining Coatue (经纬) as a venture partner. He explicitly frames this not as becoming an investor, but as structured intelligence collection.

"David [Zhang Ying] saw it very clearly. He said: Yao Song, you don't need the money — you need enough information. So at Coatue China, any meeting you want to attend, you can just push open the door — whether it's a partners meeting, investment committee, or any project review." [01:05:00]

The "Good, Multiple, Fast, Economical" Framework Reordered for Market Fit

In analyzing the commercial space launch market, Yao developed a specific prioritization framework — taking the classic Chinese industrial slogan "好多快省" (good, multiple, fast, economical) and deliberately reordering and reweighting it based on what a satellite internet constellation customer actually needs.

"Good represents high success rate. If you have a 50–60% launch success rate, no satellite internet constellation plan would use you. Second is 'multiple' — satellite internet requires launching a full orbital plane of 12–18 satellites at minimum, needing around 5–6 tons of payload capacity. Third is 'fast' — to capture the market. So we lead with good, multiple, fast, and place economical last." [01:17:57]

VLA Has Hit a Ceiling — The Next Wave is Reinforcement Learning + World Models

Yao identified the architectural limits of Vision-Language-Action models (VLA) independently before this became a mainstream view, based on reading papers showing that adding more training data was no longer improving — and in some cases was degrading — model performance.

"I rapidly realized that VLA had basically run its course. First, there were already papers showing that adding more training data to a VLA model didn't lead to the expected gradual improvement — performance actually started to decline. The marginal gain was shrinking. Second, I saw some companies already starting to shift from VLA to reinforcement learning." [01:38:09]

Physical AI Requires System-Level Integration, Not Just Model Innovation

Yao's most sustained argument is that "small geniuses" building world models in isolation cannot solve physical AI — the problem requires simultaneous mastery of data, compute infrastructure, algorithm efficiency, hardware, and deployment. He traces this directly to a conversation with Momenta's founder a decade earlier.

"Cao Xudong said: 50% is data, 30% is compute, 20% is human experience and inspiration — and he put experience before inspiration. Going back ten years later and looking at the leading model companies in China, none of them are solo young professors or small geniuses. DeepSeek with 100 talented people is important — but still not sufficient as a system. You must solve data, compute, and talent pool simultaneously." [01:48:42]

Latency and Real-Time Constraints Make Physical AI Fundamentally Different from LLMs

Yao articulates a technical point that has major implications for investment theses in the physical AI space: the gap between a training-time model and a deployment-time model in robotics could be a million-fold in parameters, versus a few-fold for language models.

"Physical AI demands real-time performance. You must achieve 30 frames per second — without that, the system is meaningless. You need very low latency, very fast speed. And in many cases, you can't compute in the cloud — it must run on an edge chip. This is a fundamentally different constraint from large language models." [02:19:03]

International Architecture Must Be Built In From Day One

Yao argues that most Chinese physical AI companies cannot internationalize because their shareholder structure and resources were set before the opportunity was recognized — and once fixed, they cannot be undone in the current geopolitical environment.

"This is related to equity structure. Everyone understands we are in a period of historic transformation. Once your shareholder structure is set, you may not be able to offset certain factors. Most first-time founders are exploring the map — they think first about algorithm, hardware, demo, and only afterward about where to sell it." [01:41:29]

The Anti-"All-In" Philosophy: Keep Strategic Optionality

Yao explicitly rejects the conventional startup wisdom of "burning your boats" — and grounds it in a mathematical argument about ruin probability, not just personal philosophy.

"The gambler always fails — even if the odds are 50-50, there's a moment when your capital hits zero. Once it hits zero, you have no capital left to recover. I never encourage mortgaging your house. You should put your time, energy, attention, and resources all-in — but never seal off your retreat." [01:31:05]

The Three-Stage Market Adoption Model Applied to Physical AI

Drawing explicitly from the LLM experience, Yao maps a three-stage adoption arc: (1) technical proof for enthusiasts, (2) mainstream awareness moment (analogous to DeepSeek's Spring Festival launch), and (3) verified commercial value (analogous to Anthropic achieving $40B ARR). He wants Striding AI to participate in all three stages of physical AI.

"In the LLM domain, when did the second stage happen? That was last January when DeepSeek open-sourced. Around Chinese New Year, roughly 60–70 million Chinese people experienced it — writing poems, Spring Festival couplets — and felt how useful it was. That was when the industry truly took root." [02:10:11]

Embodied Intelligence Will Surpass the Automotive Industry Within 15 Years

Yao makes a confident prediction about the economic scale of physical AI, which he sees as encompassing every major sector — manufacturing, services, agriculture, retail, automotive.

"The embodied intelligence industry — or the robotics industry — in a 15–20 year cycle may surpass the total output value of the entire automotive industry. I'm even more optimistic — I think in a 10–15 year cycle it will surpass automotive." [02:32:44]

2. Contrarian Perspectives

The "Small Genius" Model Cannot Win in Physical AI

Against the prevailing narrative that a brilliant PhD student or young professor founding a company around a single breakthrough algorithm (world model, VLA, etc.) is the winning archetype, Yao argues this is structurally insufficient — and cites specific paper evidence for VLA's plateau.

"Being frank: one or two young teachers or one or two small geniuses cannot solve these problems. VLA has basically run its course. If you want to consistently produce industry-leading algorithms, you need a very large talent resource pool, data infrastructure, and compute simultaneously." [01:47:43]

Humanoid Robots Are the Wrong Form Factor for Near-Term Deployment

Against the dominant narrative that humanoid robots are the correct end-state form factor, Yao argues that wheeled dual-arm robots are the correct near-term commercial choice — and that the number of degrees of freedom a robot should have is determined by algorithmic control capability, not hardware capability.

"The number of degrees of freedom a robot should have is not determined by its hardware — it's determined by your algorithm's control capability. Autonomous driving has only forward/backward/left/right and we've been working on it for ten years without fully solving L4. Embodied intelligence is far more complex than autonomous driving by two to three orders of magnitude." [02:04:53]

Demo Videos Are Systematically Misleading — The Industry Has a Widespread Deception Problem

Yao makes a pointed and specific accusation about the current state of the physical AI demo landscape, with concrete technical details about how to identify misleading videos.

"A large number of companies are using teleoperated videos instead of truly autonomous operation as demos. Every video I see now, I look for: first, is there a small label saying 5X or 10X speed? Because most VLA and world model videos are sped up 5, 10, even 20 times. Second, is it labeled 'fully autonomous'? I specifically told our brand team: every demo video we release must label these things." [02:35:57]

Strategic Retreat (Retaining Optionality) Is More Important Than Strategic Purity

Against the typical VC-backed startup mentality of "pursue the biggest outcome at all costs," Yao argues for maintaining non-financial retreat options — specifically, he declined a concurrent investment from Nvidia, Intel, and Xilinx simultaneously because having all three as shareholders would have created strategic paralysis around an acquisition.

"Pursuing success and preventing failure are two completely different paths. When you're pursuing success, you take risks — your plan may require continuous fundraising to sustain. But when you're preventing failure, you become paralyzed — every dollar invested must return two, you shrink the team to extend your runway. The rhythm becomes severely distorted." [00:32:05]

Language Cannot Model the Physical World — "Language Is the World" Is Wrong

Yao directly challenges the emerging thesis that large language models can serve as sufficient world models for physical AI systems.

"Language is inherently a digitized thing. The physical world is analog. Please describe in language the exact angle at which I'm tilting this cup right now. Please describe in language exactly how many meters apart we are. You cannot give a precise answer. The real world is an analog world — all its signals are continuous, not discrete." [02:18:20]

3. Companies Identified

Deephi Technology (深鉴科技) AI accelerator chip startup co-founded by Yao Song with Professor Wang Yu and Han Song from Tsinghua University. Why mentioned: Sold to Xilinx for over $300M within ~2.5 years of founding; regarded as one of the most successful early hard-tech exits in China's AI chip wave.

"We got a very attractive offer. In the end we accepted it. From my personal perspective, at that time I was maybe 25–26 and came from a salaried family — the number was definitely attractive." [00:52:40]

Orient Space (东方空间) Commercial rocket company founded by Yao Song in 2021. Why mentioned: Successfully launched Gravity-1 (引力一号), China's first privately livestreamed rocket launch; developed China's largest solid-fuel commercial rocket; signed contracts worth over 10 billion RMB.

"The moment of our first successful launch — Gravity-1 in early 2024 — that was definitely my happiest moment." [01:20:55]

Striding AI (振行创新) Yao Song's current physical AI startup. Why mentioned: Founded with strategic shareholders Charoen Pokphand Group (正大集团) and Wingtech Technology (华勤技术) to build system-level physical intelligence with day-one international deployment capability; 70 employees, ~$100M Series A completed.

"We have two major shareholders besides myself — Thailand's CP Group and China's Wingtech Technology. CP had revenues of $110 billion last year. This gives us a global commercial network from day one." [01:34:18]

DeepSeek Chinese AI lab known for extremely efficient large language model development. Why mentioned: Cited as proof that sustained model leadership requires hundreds of talented people plus massive compute (tens of thousands of GPUs), not individual geniuses — and that open-source release created the "mainstream awareness moment" for LLMs in China.

"DeepSeek's achievement is not simply because they have 100–200 graduates from Tsinghua and top universities. It's because they have tens of thousands of GPUs. Without that compute, none of it is possible." [01:52:34]

Momenta Autonomous driving company. Why mentioned: Founder Cao Xudong's decade-old insight — 50% data, 30% compute, 20% human expertise — is cited as a founding principle of Yao's physical AI thesis.

"Cao Xudong's answer: 50% is data, 30% is compute, 20% is human experience and inspiration — and he placed experience before inspiration." [01:48:42]

Hesai Technology (禾赛科技) LiDAR manufacturer. Why mentioned: Li Yifan's insight that engineering execution — not technology path choice — determines LiDAR market outcomes was a pivotal data point that led Yao to pass on investing in the LiDAR sector.

"Li Yifan's conclusion: the technology path doesn't matter — what matters is engineering capability. Going from prototype to product to small batch to mass delivery, replacing hand assembly with automation — that entire pipeline is hard. Hesai had a head start on that path." [01:03:22]

Anthropic US AI lab. Why mentioned: Cited as the evidence that LLMs have entered verified commercial value stage — achieving $40B ARR and reported quarterly operating profitability.

"In Q1 of this year, Anthropic reached a $40 billion ARR. And some reports suggest Anthropic achieved quarterly operating profitability." [02:10:41]

Coatue (经纬中国) Venture capital firm. Why mentioned: Zhang Ying (David) gave Yao full access to all meetings as a venture partner — creating the information infrastructure that led to his identification of the commercial space opportunity.

"David said: Yao Song, you don't need money — you need enough information. Any meeting at Coatue China, you can just push open the door." [01:05:00]

Zhiyuan Robotics (智元机器人) Chinese humanoid robot company. Why mentioned: Cited as an example of a company that has moved aggressively into international business development, beginning in the second half of last year.

"For example, Zhiyuan has been very actively building out international operations since the second half of last year." [01:46:16]

Cambricon (寒武纪) AI chip company. Why mentioned: Used as a comparison case — Yao notes that without large model demand driving compute needs, the original GPU companies' growth trajectory would have been unclear; Cambricon's IPO was cited as a catalyst for domestic GPU startups.

"If there had been no large model driving extreme compute demand, these companies' business growth would have been unclear. In 2019–2020, one direct cause of domestic GPU creation was Cambricon's IPO." [01:07:15]

Horizon Robotics (地平线) Automotive AI chip company. Why mentioned: Used as a case study that what looks like a bubble at 30x revenue can prove accurate if the market arrives — Horizon shipped 100,000 chips by 2020 and IPO'd in Hong Kong in 2024 at over HK$150B peak valuation.

"In 2018, Horizon had roughly 100M RMB revenue, valued at 30-something billion USD — everyone thought the bubble was enormous. But by 2020, Yu Kai posted a Moments celebrating shipping 100,000 chips, and the industry began its rapid positive trajectory." [02:28:25]

CP Group (正大集团 / Charoen Pokphand) Thai conglomerate. Why mentioned: Largest external shareholder of Striding AI; provides 16,000 directly operated convenience stores (711 in Thailand/Myanmar), 2,700 supermarkets, and 40,000+ full-time employees as a proprietary closed-loop data source — reducing data acquisition costs to below 1% of market price.

"CP Group has 40+ million full-time employees. So this is not a resource you can simply buy with money." [01:51:00]

Wingtech Technology (华勤技术) World's largest smartphone and laptop ODM. Why mentioned: Co-investor in Striding AI; provides industrial manufacturing (3C electronics) deployment scenarios and proprietary process data.

"Wingtech covers smartphones, laptops, watches, earbuds, glasses, printers, power banks, robot vacuums — every electronics category's manufacturing." [01:34:45]

Shanghai Commercial Bank / Hailiang Home (海澜之家) Chinese apparel retailer. Why mentioned: Sponsored the name of the Gravity-1 rocket as "Hailiang Home" — the first commercial rocket naming sponsorship in China, organized by Yao.

"We called it the 'Hailiang Home rocket' — people found it very interesting. The chairman of Hailiang Home, Zhou Licheng, is also a Tsinghua alumni." [01:21:55]

4. People Identified

Wang Yu (汪玉) Professor, Tsinghua University EE Department; co-founder of Deephi. Why mentioned: Mentor who persuaded Yao's parents, helped structure the founding team, and served as the "commissar" to Yao's "commander." Identified the AI accelerator direction and remains a research collaborator at Striding AI.

"My relationship with Professor Wang was like a commander and commissar. He set the major strategic direction; I commanded day-to-day operations. If my decisions caused problems, he stepped in. If a client needed more authority than a 22-year-old could provide, he appeared." [00:44:01]

Han Song (韩松) MIT associate professor (tenured 2024); Chief Scientist of Deephi. Why mentioned: Pioneer of neural network sparsification and compression (Deep Compression, SqueezeNet); his work was complementary to Deephi's hardware acceleration — natural fit for the founding team.

"Han Song was a PhD student at the time — maybe second or third year. He is now MIT's tenured associate professor. His sparsification and quantization work was naturally paired with hardware acceleration." [00:27:43]

Zhang Ying (张颖 / David Zhang) Managing Partner, Coatue China (经纬中国). Why mentioned: Structured Yao's venture partner role in a uniquely intelligent way — giving him information access rather than a financial role — demonstrating exceptional understanding of what a serial entrepreneur actually needs between ventures.

"I think Zhang Ying is an incredibly sharp person. He gave me unrestricted access to all Coatue meetings — he was already preparing the foundation for my next venture." [01:05:21]

Li Yifan (李一帆) Co-founder, Hesai Technology (禾赛科技). Why mentioned: Provided the key insight that in the LiDAR industry, engineering execution over 3–5 years matters more than technology path selection — a principle Yao generalizes broadly to hard tech sectors.

"Li Yifan's judgment: if you're still developing solid-state LiDAR, Hesai and Robosense will still be ahead, because the engineering path takes 3–5 years to complete." [01:03:52]

Deng Feng (邓锋) Founder and Managing Partner, Northern Light Venture Capital (北极光创投). Why mentioned: Gave Yao the blunt, transformative feedback that shifted his pitch approach — from technology-first narration to market-first narration.

"Deng Feng was very direct. He said: this pitch deck, even after many revisions, is still a pure technologist's framing. Investors want to hear: there is a trillion-yuan market, it needs this product, and my technology happens to be able to deliver it with these proven metrics." [00:35:28]

Xie Guomin (谢国民) Honorary Chairman, CP Group (正大集团); age 86. Why mentioned: Despite being 86, sat through a 3-hour technology briefing without drinking water, without checking his phone, without leaving for the bathroom — and then insisted on personally scanning WeChat QR codes of all the young entrepreneurs present. Agreed to be Striding AI's anchor strategic partner.

"This old man — at that age — still maintains relentless curiosity about technology and new things, and shows boundless respect for people. He said: no, you are the teachers here, don't move — I'll scan your WeChat. And he bent down, one by one, to scan each person's QR code." [01:43:53]

Cao Xudong (曹旭东) Founder, Momenta. Why mentioned: His decade-old insight on the composition of AI algorithmic advantage (50% data, 30% compute, 20% human expertise) has proven prescient and became foundational to Yao's physical AI thesis.

"I asked Cao Xudong: what does building an excellent algorithm mostly depend on, and where will competitive advantage ultimately emerge? His answer: 50% data, 30% compute, 20% experience and inspiration." [01:48:42]

Han Yanjun (韩燕俊) NYU professor; Tsinghua EE Class of 2011. Why mentioned: The archetypal "genius" of Yao's cohort — could ace every exam in 30 minutes, so exceptional that professors designed final exam questions specifically to challenge him; now a professor at NYU.

"The professor who wrote the final exam designed the last question specifically to stump him — just him. Because he was so extraordinary that he could casually ace every exam in 30 minutes." [00:07:57]

Wang He (王赫) Robotics researcher/entrepreneur; one year ahead of Yao at Tsinghua. Why mentioned: Was chairman of EE Science Association one year before Yao; later invited by Xie Guomin to brief him on embodied intelligence; represents the "typical" path of PhD → professor → entrepreneur that Yao accelerated through.

"Wang He was president of the EE Science Association — one year ahead of me. He followed the typical path, and eventually also started a company." [00:23:24]

Yang Zhilin (杨植麟) Co-founder, Moonshot AI (月之暗面/Kimi). Why mentioned: Cited as an example of a founder who solved the talent pool problem through personal technical reputation, pulling in Tsinghua CS, Yao Ban, and EE classmates from the start.

"Yang Zhilin solved the talent pool problem through his technical reputation and industry influence — pulling together Tsinghua CS, Yao Ban, and EE classmates from the very beginning, then using their networks to attract the youngest talents through competitions." [01:56:20]

Gao Jiyang (高纪扬) Robotics entrepreneur; Stanford UGVR program alumnus. Why mentioned: Co-participant in Stanford undergraduate research program with Yao; later said to Yao that embodied intelligence will be "king of all industries."

"Gao Jiyang told me: the embodied intelligence industry, in the long run, will definitely become the king of all industries." [02:33:13]

Xu Huazhe (许华哲) Robotics researcher; co-founded a new company (一点). Why mentioned: Cited as taking the opposite strategic view to Yao on vertical integration — arguing that companies should hand more to the ecosystem rather than doing everything themselves; CFound has invested in his company.

"Xu Huazhe's view is somewhat the opposite of mine — he believes previous embodied intelligence companies made the mistake of trying to do everything themselves, when some things should be handed to the ecosystem." [02:02:31]

Liang Wenfeng (梁文峰) Founder, DeepSeek. Why mentioned: Cited as an example of using personal technical credibility to attract exceptional talent — the talent-building model that Yao benchmarks against.

"For DeepSeek, very good compensation combined with Liang Wenfeng's personal technical credibility attracted many exceptional people." [01:56:49]

Gao Feng (高峰) Current algorithm lead at Striding AI. Why mentioned: Multiple professors independently evaluated him as someone whose PhD output equals ten people — the kind of individual that cannot be attracted purely through money or partnerships.

"Our current algorithm lead Gao Feng was evaluated by multiple professors as someone whose PhD work equals ten people. This kind of person cannot simply be bought with money or brought in through partner relationships." [01:57:16]

Zhou Erjin (周二进) Co-founder, Yuanli Lingji (原理灵机). Why mentioned: Was the algorithm lead on the original Deephi prototype project as undergraduates; later became founding team member of Yuanli Lingji.

"The algorithm lead of our original project was Zhou Erjin — now a co-founder of Yuanli Lingji." [00:24:49]

Tang Wenbin (唐文斌) Founder, Yuanli Lingji (原理灵机). Why mentioned: Former president of Tsinghua CS Science Association; long-time acquaintance of Yao; his company is building toward physical AI from the software side.

"Tang Wenbin was president of the CS Science Association — I've known him for many years." [00:24:49]

Zhao Hao (赵浩) Former Chief Scientist at Guanglun Intelligence; Chief Scientist at Jizhi Robot. Why mentioned: Was responsible for Tiangong Factory at Tsinghua; illustrates how the Tsinghua science innovation ecosystem produced a disproportionate number of key figures across the robotics industry.

"When I was there, the head of Tiangong Factory was Zhao Hao — EE Class of 2009. He was later Chief Scientist at Guanglun Intelligence, then joined Jizhi Robot as Chief Scientist." [00:14:16]

Tang Jie (唐杰) Professor, Tsinghua; founder Zhipu AI (智谱). Why mentioned: Cited as the talent-pool model where a professor's stream of increasingly exceptional PhD students becomes a continuously replenishing innovation pipeline.

"For Tang Jie at Zhipu, he naturally has many students — and the caliber and intelligence of his students keeps rising over time. They can continuously contribute new ideas to the company." [01:56:49]

5. Operating Insights

The "Sunk Cost Is Always Sunk" Decision Rule — Applied at Operational Speed

Yao killed Deephi's drone business one month after its public launch at CES, after a full year of development, because the market structure (70% DJI dominance, DJI refusing to partner with small companies) made the addressable opportunity permanently small. The speed of recognition and execution, not the emotional difficulty, is the key.

"You must believe that sunk costs are sunk costs. No matter how much you've invested, it's already a sunk cost. The drone market problem was structural: 70% was DJI, and DJI wouldn't work with small companies. So even the second-largest company in China couldn't generate meaningful volume — and all other companies combined wouldn't add up to much either." [00:47:52]

Pre-Map Your Exit Scenarios Before You Take the First Check

Immediately after receiving his first term sheet (March 2016), Yao began systematically war-gaming every possible exit scenario — IPO on A-share/US/HK, acquisition by Nvidia, Intel, Xilinx, or domestic players — and used this analysis to turn down a multi-party strategic investment that would have superficially boosted the brand but structurally impaired any acquisition.

"After getting the term sheet in March 2016, I started reverse-engineering how this company would end. If we went for IPO, what would each market require? If acquisition — what would Nvidia want to buy, what would Intel want to buy, what would Xilinx want to buy, what would Alibaba want to buy? You must be the world's best at something that is a critical link in their strategic direction. That's when they'll acquire you." [00:40:42]

CEOs Always Encounter Only the Problems Where Both Options Are Bad

Yao articulates a structural insight about why the CEO role is inherently uncomfortable — and why that discomfort is proof the role is being done correctly.

"The CEO always encounters the most difficult, most unpleasant problems. Why? Because when a decision has one good and one bad option, nobody brings it to you — everyone knows to choose the good one. When both options are good, nobody brings it to you either. Only when both options lead to bad or worse outcomes does everyone come to you and say: what do we do?" [00:49:18]

Build the Shortest Possible Pipeline from Research to Commercial Value

Drawing on regrets expressed by former Sensetime and Megvii employees, Yao identifies that the core organizational failure of the previous generation of Chinese AI companies was allowing the research-to-deployment pipeline to grow too long — leading to value decay as research leadership rotated to competitors.

"People who worked at Sensetime or Megvii — their one regret: the pipeline from research to commercial value in their department was too long. They didn't shorten it. And even the most leading companies cannot guarantee their algorithms are always ahead. When technology can no longer be the moat, you should find another moat." [01:58:14]

Fund the Talent Pool, Not Just the Core Team

Drawing from the history of deep learning breakthroughs (AlexNet → VGGNet → GoogleNet → ResNet → FasterRCNN), Yao argues that no individual or small team can maintain sustained algorithmic leadership across consecutive breakthroughs. The organizational implication is to build a large, diverse talent pool where different sub-teams can each lead for a cycle.

"No person can continuously produce world-changing algorithms and always stand at the frontier in a technical field. So you need a large resource pool, with different teams each leading for a period. If you can absorb enough diverse teams, you can maintain sustained industry leadership." [01:54:53]

6. Overlooked Insights

Proprietary Closed-Loop Data at 1% of Market Cost is the Real Physical AI Moat

Buried in the discussion of Striding AI's CP Group partnership is a specific, quantified claim that deserves much more attention: Yao states that by working with CP's 40,000+ employees across 16,000+ stores, he can generate training data at less than 1% of market procurement cost (market rate: 300–600 RMB per hour). This is not merely a cost advantage — it is a structural barrier that prevents any capital-rich competitor from replicating the data asset simply by spending more money, because the data is proprietary operational process data that CP would never sell to a third party.

"I can reduce my data costs to below 1% of the normal market procurement price through this approach. Normal market rate for first-person perspective video data is 300–600 RMB per hour. If you need a million hours of data — which is now considered the minimum — that's 300M to 600M RMB in data costs alone. CP Group has 40+ million full-time employees. This is not a resource you can simply buy with money." [01:50:09]

The non-obvious implication: the "small genius + big capital" model cannot close this gap. A competitor could raise $500M and still not get access to 16,000 retail stores' operational data, because the incumbent operators will not sell process data that reveals their competitive advantage. Proprietary deployment relationships — not model quality — may be the durable moat in physical AI.

The 6G Thesis: Satellite Internet as Cellular Base Station Replacement Was a Non-Consensus Insight That Proved Correct

Yao reveals that his decision to enter commercial space was driven primarily by a non-obvious thesis about telecommunications architecture — specifically, that 6G should route through low-earth-orbit satellites rather than ground base stations, because 5G already saturated speed improvements and the remaining gains are in coverage breadth. He arrived at this from his EE communications background, not from the rocket industry. This thesis led him to anticipate China's national satellite constellation program two to three years before it became public, and to design Orient Space's Gravity-1 specifically around the payload requirements of constellation deployment.

"I'm from an EE communications background. My basic thought: if 5G transitions to 6G, perhaps it's not purely a speed improvement anymore — we should pursue coverage breadth. One base station covering a radius of hundreds of kilometers rather than hundreds of meters — that means going to space. So satellite internet became the large market-driven customer for the rocket industry." [01:16:30]

The overlooked significance: this is a replicable framework — identify a communications generational transition, map the infrastructure requirements, and find the enabling hardware layer that is currently undersupplied. The same logic could be applied to identifying the next non-obvious hardware bet in the physical AI era.