T... | 张小珺Jùn｜商业访谈录 (Zhang Xiaojun Business Interviews) Summary

1. Key Themes

The Strategic Pivot: From AGI Dreams to Daily Active Applications

The entire AI industry has undergone a fundamental strategic shift from pursuing AGI (Artificial General Intelligence) to focusing on practical daily-use applications. Neil Shen observes: "Everyone is becoming more realistic. If you look at the three major model companies, they barely mention AGI anymore. They've dropped a lot of capabilities and are now focused more on application-side features like Posts, browsing, and even recently launched group chats. This was unimaginable before—how could a company pursuing AGI launch a group chat feature?" 00:01:59

This transformation represents OpenAI's strategic evolution from weekly active users (WAU) to daily active users (DAU). Shen explains: "Whether it's search or ChatGPT dialogue, these are weekly active scenarios. Weekly active scenarios make it very difficult to defend against major players' attacks. But daily active is different—the moat becomes deeper as usage frequency increases." 00:02:32 OpenAI's introduction of group chat functionality signals an even larger ambition: potentially rebuilding social relationships and becoming "a new era's social software." 00:03:17

China's Infrastructure Advantage: The Real Competition is Data Centers and Power

While many discuss an AI bubble, Shen sees the real competitive battlefield shifting to physical infrastructure—specifically data centers and electricity. He firmly states: "I think we won't see a bubble for at least three years. When everyone is talking about bubbles, the bubble definitely hasn't arrived yet... GPU overcapacity is impossible. There simply aren't enough cards. Even five, six, seven, or eight-year-old cards are still being used—as long as there are cards, they can be used." 00:09:49

China has a significant advantage in this infrastructure race. Shen notes: "Future competition isn't AI competition—it's data center competition, it's electricity competition. China has great opportunities in this regard. China is now building over a dozen nuclear power stations at once, including many offshore wind and solar power stations." 00:10:17 He contrasts this with America's challenges: "For the US, the key question is whether data centers and power can keep up. For Chinese entrepreneurs, this is an excellent opportunity for the next few years." 00:43:58

The 15-Degree Rule: Finding Profitable Niches Away from Crowded Markets

Shen has developed a disciplined investment strategy he calls "staying 15 degrees away from where companies are most concentrated." This means avoiding the most hyped sectors—foundation models and humanoid robots—and instead focusing on vertical applications and "dirty work" that major players won't touch. He explains: "General-purpose agents might be one street away from major companies, but vertical agents need to be three streets away—or do private deployments, like taking Qianwen's 32B small model for private deployment in vertical scenarios. These are opportunities because big companies don't want to do this work." 00:25:34

This strategy has led to investments in unexpected areas like boat-cleaning robots and massage robots with conversational AI. When asked about a boat-cleaning robot investment, Shen reveals: "We met for just over ten minutes and decided... Understanding seawater and corrosion is extremely difficult. Being able to see clearly in seawater and clean off all the barnacles—the barriers are very, very high." 00:29:22 The key insight: "When you offset by 15 degrees from where major companies are concentrated, you suddenly discover many companies with very high cost-effectiveness." 00:23:52

2. Contrarian Perspectives

Consensus Investment is a Trap—Differentiation is Survival

Shen offers a strikingly contrarian view on the current VC landscape, criticizing the herd mentality that has consumed the industry. He observes: "At last week's VC conference in Hong Kong, many LPs complained that this cycle's GP consensus is too concentrated. Every project has the same dozen GPs investing together, so everyone's ownership percentage is very small. How can you make money? GPs can't make big money, and LPs are very unhappy... And investing in different VCs seems about the same." 00:38:57

His critique goes deeper, questioning whether current consensus will translate to commercialization: "In the mobile internet era, even though there was consensus, it could eventually be proven through commercialization... Today I think some consensus is really far from commercialization—it may be very difficult for commercialization to catch up. This is a big risk." 00:39:56 When pressed on the implications, he's blunt: "If commercialization never materializes, then consensus is just a bubble... I don't know [what happens to those who invested]. That's why we keep saying you need to think clearly about where your commercialization is." 00:40:14

Foundation Model Companies Face Worse Economics Than Solar Panel Manufacturers

In a devastating comparison, Shen argues that today's foundation model companies face even more challenging unit economics than the solar panel manufacturers of the cleantech bubble: "Today's foundation model companies and those companies in the solar cycle—I think they're very similar, actually worse than solar. At least in solar, commercialization wasn't a problem. It was just that everyone's technology caught up quickly, so margins were very low. Today, even achieving commercialization may not be easy." 00:39:47

This perspective directly contradicts the trillion-dollar valuations being discussed for companies like OpenAI. Shen questions the fundamental divergence between US and Chinese AI valuations: "There's a huge difference between American AI and Chinese AI—American AI valuations are 100 times China's. One side must be wrong. Either their revenue is unsustainable, or China's revenue is underestimated." 00:12:36

Most AI Startups Won't Reach $1B Valuation—And That's Okay

Contrary to the unicorn-hunting mentality prevalent in VC, Shen sets dramatically lower expectations for AI startup outcomes: "If you're hoping to find a thousand-billion-dollar unicorn now, it's almost impossible. But if you go down a level to find a tens-of-billions or hundred-billion-dollar opportunity, there might still be some possibility... One billion dollars still has possibility." 00:29:15

This realism extends to his investment strategy: "You must look for opportunities that big companies can't see on day one. Only then is it possible. But you have to grow slowly—it might take three to five years. Three years later, they may not understand it and think it's a small opportunity. Only then do you have a chance." 00:28:28 This patient, lower-ambition approach contradicts the swing-for-the-fences mentality that dominated previous tech cycles.

The Real AI Moat Isn't Technology—It's "Dirty Work" Big Companies Won't Do

Shen repeatedly emphasizes that sustainable AI businesses won't be built on technological advantages but on execution in unglamorous verticals: "Any successful company from the mobile internet era—whether Didi, Meituan in China, or Uber and Airbnb in the US—they all did half-offline dirty work that big companies didn't want to do. TikTok, Kuaishou, and Xiaohongshu in China all existed because big companies kept making mistakes, giving them space." 00:27:32

He sees no such mistakes happening in the AI era: "In this AI wave, are there big companies continuously making mistakes? No. You can't see it in China or the US. No one dares to be careless. Not just first-tier big companies—even second and third-tier companies are all heavily involved." 00:27:47 This means survival requires doing work others actively avoid, not outcompeting them on obvious opportunities.

Token Consumption Growth Proves AI is Real—Not Hype

While many discuss an AI bubble, Shen points to a concrete metric that proves otherwise: "You can see from token consumption—China's token consumption is growing faster than the US. Many of our small companies now consume several hundred GPUs daily. OpenAI said surpassing ten thousand GPUs is a major milestone. In China, companies consuming one GPU daily are approaching ten thousand, even exceeding ten thousand—there are many. One million DAU means several hundred GPUs, and this volume is very easy to achieve. Many startups can do it." 00:10:45

This explosive growth in actual usage contradicts bubble narratives: "The core question is whether this stuff, whether on or off balance sheet, is actually being used and whether there's enough. This is the key issue. So this doesn't affect [my optimism]—it's just an accounting operation. But today cards really aren't enough, computing power really isn't enough. This is today's most core point." 00:13:10

3. Companies Identified

OpenAI

Description: Leading AI foundation model company, creator of ChatGPT
Why Mentioned: Strategic evolution and competitive positioning
Key Quotes:

"I think it's very good. I think its strategy is still very clear. This is the most natural strategic direction for model companies to take as application-side competitors—the strategic direction hasn't changed, it's still very clear." 00:02:24
"The big test is whether it can hold and defend its position as a super entrance on the C-side. This is a very important test. But looking at the third Otamin [o1], it proves its strategic choices and several financial decisions were all very well set." 00:13:41
"At $300 billion valuation, OpenAI is definitely at the ceiling globally. Many haven't even reached a fraction of that. But I think its future big test is whether it can hold its ground on the C-side super entrance." 00:13:35

DeepSeek (字节跳动/ByteDance)

Description: Chinese AI company, creator of influential open-source models
Why Mentioned: Leading domestic AI applications and infrastructure
Key Quotes:

"I think ByteDance and Qianwen, or Alibaba's ByteDance, are probably leading. Alibaba just started pushing hard on C-side recently, starting a bit late, but I think they're still in the front row." 00:03:29
"All open-source ecosystems today are definitely still priced based on DeepSeek. Recently I discovered several open-source models domestically, and their pricing may still be based on DeepSeek. This is still very valuable." 00:17:11
"If it weren't for DeepSeek, Chinese open-source wouldn't be so resolute today. Without DeepSeek, human AI might very likely be controlled by a few private company AI models. That would be dangerous for all humanity." 00:16:37

Google

Description: Technology giant with search and AI capabilities
Why Mentioned: Competitive position in AI race
Key Quotes:

"Google has always had very deep accumulation. I never thought Google was behind. Google's product and model have always been quite good, I think." 00:04:44
"Google itself is very decisive in Gemini. Its own search has also embraced it very decisively." 00:04:57
"It's definitely a competitive relationship. The key is who is more decisive in cutting over, who goes deeper in daily active scenarios. ChatGPT is at least ahead now. This group chat move is particularly beautiful." 00:05:11

Alibaba/Qianwen (阿里/千问)

Description: Chinese tech giant and its AI model division
Why Mentioned: Strong position in enterprise AI and open-source
Key Quotes:

"Alibaba has model capability. And their C-side app is also quite good. But they're just a bit late. But I think it's still worth looking forward to." 00:03:11
"Whether ByteDance or Alibaba's tokens are given so cheaply. Professional companies don't have foundation model ecosystems, so it's hard to compete." 00:19:48
"For many entrepreneurs, whether it's thousands or millions, they're now taking Qianwen's open-source small models for private deployment to do more important applications. All very successful, revenue explosions are very fast." 00:20:21

AAP Ban (AI Toy Company)

Description: AI-powered companion toy startup
Why Mentioned: Portfolio company showing exceptional growth
Key Quotes:

"AAP Ban is also doing very well. We chatted for about ten to twenty minutes and decided. This year's growth is also very good, super exceeding expectations. In less than half a year, it should at least be the top three in shipments now." 00:21:11
"Consumption metrics including token consumption metrics are also very impressive, very impressive. Also super exceeding my expectations." 00:21:21
"Even though it looks like such a small toy, making the user experience good is really not easy. And getting Wi-Fi connectivity stable is also not easy." 00:21:29

4. People Identified

Sam Altman

Description: CEO of OpenAI
Why Mentioned: Strategic leadership and decision-making
Key Quotes:

"What he says and what he does are two different things. He might say AGI, but in fact everything they're doing is on the application side. So it quite fits your viewpoint. I think he's also a very pragmatic person. An investor-backed company." 00:14:01
"I think he's a very pragmatic person—a company backed by investors. After all, you need to account to investors, to employees, to everyone." 00:14:04

Ilya Sutskever

Description: Co-founder of OpenAI, AI researcher
Why Mentioned: Comments on scaling law limitations
Key Quotes:

"You see Ilya recently said that scaling laws have been the main route these past few years—just add compute power and data and you can improve. Now we've entered a reasoning period where compute and data are about to run out, so we can't just rely on scaling models—we need to find new recipes." 00:08:36
"People have been saying this all along. At this point in time, everyone really feels that moving forward is very difficult. So everyone thinks this statement is correct. When Yang [likely Yann LeCun] said it before, when everyone was still moving forward smoothly, they said Yang was old and conservative, trying to discourage young people." 00:08:48

Zhou Hongyi (周鸿祎)

Description: Chinese internet entrepreneur and tech commentator
Why Mentioned: Comments on AI bubble
Key Quotes:

"I think what Alibaba said in its earnings call was quite good. I think we won't see a bubble for at least three years." 00:09:43

5. Operating Insights

Ownership Percentage Control is Critical in Concentrated Markets

Shen emphasizes maintaining meaningful ownership despite competitive rounds: "For a VC to make money, you definitely need to have enough ownership percentage in the companies you hit. Only then can you make money." 00:38:37 In an environment where "every project has the same dozen GPs investing together," maintaining 2-3% ownership even in early rounds becomes strategically important for fund returns.

The "Three Streets Away" Rule for Defensible Positioning

Vertical AI agents need to be "three streets away from big companies" through either: (1) private deployment models, (2) using small open-source models like Qianwen's 32B, or (3) focusing on niche verticals requiring deep domain expertise. 00:25:28 This creates natural moats because "big companies don't want to do this vertical scenario work. First, you have to do verticals. Second, you have to do private deployment. Third, you have to do sales. Big companies generally find startup companies to help them implement." 00:25:59

Token Consumption as a North Star Metric

For AI companies, daily token consumption has become the key indicator of real traction: "Now for many companies we're looking at, they basically consume several hundred GPUs per day, close to one thousand. One million DAU means several hundred GPUs. This volume is actually very easy to achieve—many startups can do it." 00:10:56 Shen tracks this more closely than revenue because it indicates genuine AI integration and user engagement.

Meeting Time as a Decision Filter

Shen has systematized rapid decision-making through strict time management: "Everyone knows I do ten-minute meetings. Whether investing or not, it's ten minutes. So I just do ten minutes. If I feel there's not much interest, ten minutes ends, and no one feels it's inappropriate because everyone knows my meetings are ten minutes." 00:32:41 Extending to 30 minutes signals serious interest: "If I meet for thirty minutes, I'm already quite interested. Basically means I'm going to invest." 00:32:57

Early Revenue Matters More Than ARR Scale

Unlike traditional B2B SaaS metrics, Shen focuses on revenue growth velocity over absolute numbers: "All are growing over ten times. But currently at small scale—large scale companies might grow three times annually." 00:30:38 He cites portfolio companies growing from "a few million last year, over ten million this year, twenty to thirty million next year" as examples of promising trajectories. 00:30:49 The key is sustainable growth rates, not hitting arbitrary ARR milestones.

6. Overlooked Insights

The Invisible Token Economy is Reshaping B2B Software Economics

Buried in the discussion about major models offering cheap or free tokens to startups is a profound insight about value transfer in the AI ecosystem. Shen notes: "Domestic foundation models give very cheap token pricing. Daily consumption of several hundred GPUs doesn't cost much, because most tokens are input-based. Input tokens are relatively cheap, and input volumes are smaller... If big companies think this scenario has such value, tokens continue to be discounted or given almost free." 00:11:16

This reveals a hidden subsidy model where foundation model companies are essentially paying startups to build applications on their platforms—the opposite of traditional cloud economics. The implications are enormous: application companies can scale usage dramatically without proportional cost increases, but they're also building dependencies on platforms that could change pricing at any time. The "two mao per DAU per day" 00:11:34 unit economics mentioned casually represents a structural advantage that could disappear overnight.

China's Open-Source Ecosystem Advantage Comes from Desperation, Not Strength

Shen makes a passing but significant observation: "If Llama or ChatGPT can't move forward, China's open-source ecosystem has great potential for further explosive development... We see many applications now taking, for example, Qianwen's open-source small models for fine-tuning to do more vertical applications. All very successful." 00:20:13

This suggests China's strength in open-source AI isn't primarily about technical capability—it's about necessity. Being cut off from the latest frontier models forces Chinese developers to extract maximum value from open-source alternatives through fine-tuning and vertical specialization. This constraint is breeding a generation of entrepreneurs who are exceptionally good at doing "more with less"—a competitive advantage that could persist even if access barriers fall. The throwaway mention that "recently I discovered several domestic open-source models, and their pricing may still be based on DeepSeek" 00:17:13 hints at an entire pricing ecosystem built around Chinese open-source models that Western observers likely underappreciate.

Conclusion: Neil Shen's 2025 perspective reveals an investor who has consciously zigged while others zagged—avoiding consensus investments in foundation models and humanoid robots while finding extraordinary returns in overlooked verticals like boat-cleaning robots and AI companion toys. His core thesis—that infrastructure advantages (data centers, electricity) and execution in "dirty work" will matter more than model capabilities—represents a fundamental bet against the AI bubble narrative while simultaneously rejecting the winner-take-all assumptions driving trillion-dollar valuations. Most strikingly, his prediction that "in ten years, China's AI will definitely lead the US because the US can't build data centers and power fast enough" 00:46:03 suggests he sees the current technology race as ultimately subordinate to physical infrastructure deployment speed—a dimension where China's state capacity provides overwhelming advantage.

The Third Serialization of Zhu Xiaohu's Realistic Story: The Feast and Bubble of Artificial Intelligence