Unity Ventures' Xiao Wang: AI Won't Repeat the Dot-Com Bubble, the Key Is Telling Real Innovation from Empty Hype
Where Do Entrepreneurs' Opportunities Lie in the AI Era?
Right now, AI is quietly reshaping the tech industry landscape, with entrepreneurs rushing in from the model layer to the application layer. But capital is concentrating at the top. In 2024, global AI funding surpassed $100 billion, with deals over $100 million accounting for 69% of that total — a clear sign of winner-take-most dynamics.
The market is also growing more intense. From large language models and Embodied Artificial Intelligence to AI applications, new companies are springing up like bamboo shoots after rain in nearly every track and scenario.
For entrepreneurs, the test is no longer just innovation and operational efficiency, but commercial wisdom — from data flow to scenario application: Is the product positioning clear? In the commercialization process, how do you build data loops and scenario moats to maintain core competitiveness? And amid this investment wave, how do you actually find a path to survival?
If these critical questions lack clear answers, and AI is simply treated as a "windfall opportunity," what it brings may not be dividends, but a pipe dream.
Recently, on the Huxiu · AI Paradox program, CEIBS Professor of Finance, Associate Dean, EMBA Program Director, and Director of the CEIBS Center for Enterprise and Capital Markets Huang Sheng, joined Xiao Wang, founder of Unity Ventures and CEIBS EMBA alumnus, to discuss AI opportunities from an investment perspective and insights into future AI applications. Chen Xiaoyan, host of the Xiaoyao Talks Business account and CEIBS EMBA alumnus, moderated the discussion, covering the following questions with the two guests:
- Does NVIDIA's $4 trillion market cap signal bubble risk?
- Will AI become the driving force of the fourth industrial revolution?
- Can the AI era escape "winner-take-all" dynamics, and what opportunities exist in vertical domains?
- What kind of AI companies do investors favor, and where do entrepreneurs' opportunities lie?
- In which fields will the next phenomenon-level AI application explode?
Below is a transcript of the conversation.
01
Where exactly are startup opportunities in this wave?
Chen Xiaoyan: NVIDIA's market cap has broken $4 trillion. How do you view this phenomenon?
Huang Sheng: Capital market valuations reflect high expectations for AI's future — discounting the free cash flows of a company's entire lifecycle to the present. AI-related companies currently have price-to-book ratios of 40-50x, similar to Apple. This is market sentiment, it's an expectation. People are "very high" on AI's future.
Wang: NVIDIA's chips have redefined the boundaries of global computability. As large models and the application layer expand, those boundaries keep widening, and the imagination space has been infinitely opened up. Human genes, manufacturing data, daily life, product purchases — all can be intelligently processed through GPU + large models. The market recognizes this underlying capability, and the stock market is willing to "bet" on the future, so NVIDIA has been assigned enormous valuation.
What's worth noting is that it symbolizes the opening of structural opportunities across the AI industry as a whole. Looking at AI's evolutionary trajectory, if NVIDIA can reach $4-5 trillion, model companies and application companies will see similar opportunities in the future.
Chen Xiaoyan: Why is AI seen as the core of the fourth industrial revolution, rather than technologies like IoT or big data?
Huang Sheng: AI replicates human decision-making. Its deep learning approach is like a "black box" — a mirror of human thinking. Other technologies have scientific logic, but they're somewhat distant from humans; artificial intelligence may be closer to us.
Wang: Big data, blockchain — these technologies didn't fundamentally shake society and technology. To become a fourth industrial revolution, it has to be foundational, a transformative force that broadly impacts human life. Past technologies followed a "human + tool" model. AI is completely different — it is productivity itself, capable of completing tasks independently without humans wielding tools alongside it.
Now, AI's intelligence is becoming increasingly general-purpose, able to penetrate nearly all human activities: production, consumption, scientific research. For example, new drug development used to take 10 years and $100 million; AI might get it done in 1-2 years, with doubled efficiency and dramatically reduced costs.
Moreover, tasks like writing presentations, programming, and making short videos — previously dependent on intensive intellectual collaboration — can now be done better through AI. The societal impact is disruptive, liberating human energy to focus on more creative work.
It's still too early to say AI is 100% the fourth industrial revolution, but it absolutely has that potential. Large AI models are very likely the starting point of this revolution, fundamentally transforming social structure and productivity.
Chen Xiaoyan: In 2024, global AI funding surpassed $100 billion, with deals over $100 million accounting for 69%. Head effects are pronounced. For domestic small and medium-sized AI startups, does this mean fewer opportunities?
Wang: Opportunities in AI's foundation models versus application layer are completely different. Foundation models show clear head effects — they require massive compute, top engineers, and enormous capital. Small and medium companies can barely squeeze in; it's basically a game for big players. But the application layer is entirely another matter.
Small companies can fine-tune models or train small models for specific scenarios, with low resource consumption and low barriers. Because application scenarios are diverse, the commercial and social value is enormous: any domain with data, from lifestyle services to industry decision-making, can use AI to improve efficiency and create value. The potential in these scenarios is far from exhausted, and the underlying value may be orders of magnitude higher than foundation models.
So if small companies plunge headlong into the foundation model "arms race," it's genuinely a waste of resources. Startups should focus on the application layer, digging deep into practical value for human life and society — the return potential is substantial.
Chen Xiaoyan: Benchmarking against the internet industry, what stage is AI development at? Do ordinary people still have opportunities?
Huang Sheng: AI right now resembles the Mobile Internet stage from over a decade ago. Back then, 4G-to-5G networks, mobile terminals — this infrastructure was mature. What was missing was applications, so Douyin and SaaS companies rose rapidly.
Now AI is similar: NVIDIA's compute, tech giants' large models have laid the infrastructure foundation. Domestic players are catching up; compute and models are basically in place. What follows is application-side explosion — whoever finds AI-suitable scenarios and provides high-value services can grow quickly.
In my view, AI won't be as "winner-take-all" as Mobile Internet, because vertical domains like healthcare and education have data moats, making it possible for multiple great companies to emerge simultaneously.
Wang: In my view, if we compare to internet development stages, it came in two waves: PC internet (around 2000 to 2010-2011) and Mobile Internet (2011-2012 to present). These two waves spawned different companies, but the core business model was similar — aggregating massive users and data through刚需 tools or media platforms, ultimately monetizing through advertising. The core was network effects generated by "connection."
Large AI models are completely different from the internet. They don't rely on "connecting" people to sell ads. AI's business model leans more toward direct service delivery for fees. For example, AI directly gives you programming code, edited videos, matched talent — delivering final results.
This direct-service model generates revenue from day one, from the first user. AI's profit model is naturally front-loaded. Therefore, the industry shape is more likely to be a hundred flowers blooming, with vertical domains each having their giants.
Benchmarking the time window, AI now resembles the early internet around 2000. Foundation model capabilities are already quite mature, but this is just the beginning. As model capabilities improve, the application layer will cover more job types and scenarios. Just as Mobile Internet exploded due to the carrier upgrade from PC internet, the AI application layer now stands on the eve of its own explosion.
What's worth noting is that AI's普及 speed is very fast. WeChat took over a year to reach 100 million users; ChatGPT took two months, DeepSeek App took two weeks. Correspondingly, AI application takeoff speed will be much faster than the internet's.
Chen Xiaoyan: What AI application tools do you normally use? Which domains do you lean toward?
Huang Sheng: I've tried AI tools domestic and international. I mainly lean toward research assistance, treating AI as an intelligent search engine — more convenient than Google or Baidu, capable of dialogue and continuous updates. For example, doing literature review: AI can quickly organize massive research and extract conclusions. When writing English articles, AI helps me polish the language; as a non-native speaker, AI's expression is more elegant while academic rigor remains unchanged.
Wang: The core of human work is acquiring and processing information, making judgments. AI helps tremendously with information gathering and organization. For example, looking up technical terms or market trends — AI can quickly梳理 knowledge, more efficient than traditional search. But judgment and推演 still depend on oneself; AI mainly accelerates preliminary information processing.
Truly life-oriented AI applications are still rare. Current large models focus on foundational capabilities; the application layer is still in development. Going forward, I look forward to more interactive, personalized applications in entertainment, education, and healthcare. For example, recommending content based on preferences or generating videos. Currently in education there are some AI companion products for young children, but intelligence is limited; learning applications are still being explored. Medical queries have potential, with rich knowledge reserves.
02
Will the AI industry repeat the internet bubble?
Chen Xiaoyan: How do you view investment bubbles in AI? Will it repeat the internet's mistakes?
Huang Sheng: The AI industry may have certain bubbles, but these differ from internet bubbles and may be necessary "bubbles" that drive technological progress. Bubbles can be divided into two types: asset bubbles and technology bubbles. Asset bubbles, like first-tier city real estate in China, have extreme price-to-income ratios or rent payback periods — decades of annual salary or a century of rent to break even, driven by excess liquidity: too much money, too few assets, prices pushed up. Such bubbles easily burst due to demand decline (like population reduction) or liquidity retreat, causing enormous losses — negative bubbles.
AI's "technology bubble" is different. Despite high valuations and large investments, technology investment is dynamic, able to push boundaries and expand human cognition and productivity. For example, AI technological progress can dramatically improve efficiency; as long as AI technology serves good, this bubble can create value and drive innovation — a positive force needed for industry development. Without this "bubble's" optimism and capital inflow, technological progress would be hindered.
Take NVIDIA as an example. Despite profit support, its price-to-book ratio reaches 47x, P/E over 50x, far exceeding Cisco's 16x during the internet bubble, and higher than Google, Microsoft, and Amazon's 8-12x. In 2024, global AI funding exceeded $100 billion; this capital frenzy also evokes memories of internet bubble-era狂热. But technology bubbles should be allowed to exist to some degree; innovation requires venture capital support and tolerance for failure.
Wang: We're in the primary market — essentially paying for the future, investing in dreams, not simply looking at "bubbles." AI industry valuation premiums reflect large-scale structural opportunities, representing超前 investment and imagination space for the future.
I'm more inclined to view AI's high valuations as "dreams" rather than negative "bubbles." Bubbles imply 100% certain rupture; dreams, even if only 1-2% are realized, can create enormous value. VC exists to pay for dreams backed by technology and social value, where successful projects can drive social progress and bring huge returns.
Our task is to filter out pure storytellers, find projects that can realize big dreams and create universal value, and support them to go further. AI's high valuations aren't a repeat of the internet bubble, but technology-driven dream investment — the key is rationally distinguishing true innovation from empty stories.
Chen Xiaoyan: AI investment bubbles are like beer foam. A good glass of beer, 20-30% foam is normal, adding flavor — but the glass must contain beer (technological substance). If the glass is empty and hot money pours in with foam at 80%, that's "black shop"劣质酒, worthless. Appropriate foam is good for companies with technological dreams. So how do you distinguish true innovation from empty stories? When investing in an AI project, what do you value most?
Wang: When evaluating an AI project, the core is whether it has found a genuinely valuable application scenario. A good AI project must focus on a specific domain, and this scenario must form a closed loop where data can continuously self-reinforce. For example, initially the system might only complete 70% of work, but as data accumulates and algorithms optimize, it gradually improves to 90% or higher. The project needs not only solid technical capability but also deep industry understanding to truly get the "data flywheel" spinning.
If we reject an AI project, there are usually two reasons: one, the entrepreneur's story is too small, lacking imagination; the other, more common, is the story is too big but the team completely lacks corresponding capability to support it. There must be a reasonable connection between vision and capability — even if the path isn't 100% clear, the logic must at least be self-consistent.
For example, we invested in a company using AI to transform the headhunting industry. Traditional headhunting work is essentially bilateral communication — understanding both company needs and candidate evaluation — a process actually well-suited for AI optimization. This company's two founders: one is a veteran headhunter with industry resources and client base; the other is a technical expert from Microsoft Research Asia, skilled in large model applications. This combination ensures both deep understanding of recruitment scenarios and capability to build AI systems.
More importantly, every match generates new data, making the system smarter with use. In the past, headhunter phone call information wasn't retained; AI can structure this interaction data, creating entirely new competitive moats. From a market size perspective, China's recruitment market is a hundred-billion-level market; AI can first serve high-end positions, then gradually penetrate. Such a project fully matches our definition of a good AI project — clear scenario, closable data loop, matched team.
03
Current investment logic in AI and Embodied Artificial Intelligence markets
Chen Xiaoyan: How do you view China's chip industry? What changes are occurring in the China-US tech competition landscape?
Wang: China's chip industry has indeed made remarkable progress. Currently, except for a very few highest-end products, domestic players have established relatively complete industrial chain capabilities. This mainly stems from support by the massive domestic market demand, plus changes in the external environment that have given local enterprises more market opportunities. But in reality, breakthroughs in high-end chips require longer-term technology accumulation. At the same time, we should understand that China's relatively weak chip industry foundation is related to its later start — the United States and other countries have over 70 years of technology accumulation in semiconductors.
However, although the United States sanctioned us, it was precisely because supply cutoff forced self-reliance that Chinese chips achieved breakthroughs in a short time. From DSP, CPU to GPU, Chinese manufacturers gradually achieved self-developed breakthroughs. Now Huawei's large chips can also be mass-produced — this is a milestone. We can say that the offense-defense phase in chips between China and the US has temporarily concluded, ushering in a "halftime break." Going forward, the competition focus is shifting toward AI model alignment and application落地.
Chen Xiaoyan: Where lie AI application opportunities? How do C-end and B-end development differ?
Wang: C-end is moving faster. Personally, I believe the most potential still lies in entertainment — interactive video, short drama generation, virtual chat companions, gaming assistants, short video creation tools. These scenarios have large user bases and strong demand, just like Douyin and Kuaishou in the Mobile Internet era.
B-end has good AI application opportunities in many vertical domains: end-to-end delivery, high commercial value scenarios, professional scenarios where model intelligence capabilities overflow — all have opportunities for AI to directly deliver results.
In the investment field, one concern is that as large model capabilities extend, companies without deep data flow capabilities and scenario definition capabilities may easily get swallowed.
Overall, on one hand you need to address large markets; on the other, application companies need the capability to first form relatively good data moats and experience moats themselves. Good product definition and scenario definition are very important.
Huang Sheng: I agree. B-end requires deeper vertical penetration. Customer service, sales lead generation, education, music, law, healthcare — these industries require professional know-how and data flywheel support. Although starting slower, once you go deep, you can form very strong moats.
Chen Xiaoyan: Compared to the United States, what gaps and advantages does China have in large models and AI applications?
Wang: I believe large models themselves are naturally cross-lingual and cross-regional, so the China-US gap isn't in the model itself but in regulatory environment, cultural atmosphere, and team execution capability. How big the technology capability gap actually is — I don't think it's necessarily that large.
In product operation capability, the Mobile Internet step already proved China's significant advantage in this regard. Given China's engineering capability, product definition capability, user operation capability, and supply chain capability, including hardware, I believe Chinese teams can more or less establish clear advantages globally.
This generation of AI startups differs from early internet companies in that they are "born global." Our entrepreneurs also have very strong execution capability. Plus our inherent billion-level user market enables faster product iteration speed, naturally possessing globalization potential.
Chen Xiaoyan: Will large model manufacturers directly swallow application-layer startups?
Huang Sheng: Looking at the United States' experience over the past two decades, among VC/PE-backed startups, about 90% ultimately exit through acquisition, with only 10% achieving independent IPO. Typically, companies with disruptive innovation but unclear profit prospects mostly head toward IPO; while application-type,补强-type technology enterprises are more easily acquired by large companies. This paradigm has been basically stable since 2000; US domestic IPO counts have declined year by year, mostly not exceeding 200 per year, far below the盛况 of the 1980s-90s internet bubble era.
By comparison, China's situation differs. Due to cultural and market structure differences, Chinese entrepreneurs generally harbor an "IPO faith," tending to view listing as a milestone in enterprise development, preferring "ringing the bell" over selling companies. Especially GPU, compute and other hard tech enterprises, as well as rapidly revenue-growing application companies, mostly prioritize IPO, achieving more efficient financing and international visibility through A-shares and Hong Kong stocks.
Meanwhile, China lacks the mature mid-sized tech startup ecosystem seen in the United States; large companies usually choose self-development over acquisition, further weakening the possibility of M&A exits. Overall, Chinese AI startups will more likely take IPO as their primary path rather than being acquired.
Chen Xiaoyan: In what domains do you think the next phenomenon-level application will appear?
Wang: I think it's very likely to appear in to-C, generative content consumption and social platforms — not necessarily in forms like intelligent agents.
AI as a whole is still in a very early development stage. Chinese entrepreneurs still have many opportunities to create truly useful AI products for the whole world, thereby enhancing overall human happiness, efficiency, and sense of value. Ultimately, future globalized AI giants will be born from this batch of Chinese entrepreneurs.
Huang Sheng: I believe whatever the industry, it must meet certain standards. First, it must be in a vertical domain with unique data advantages. Second, the industry must know how AI can help it improve efficiency. Finally, it must be able to address AI's negative effects while possessing commercialization value.
Personally, I'm quite optimistic about AI applications in healthcare or the broader health domain. It receives much attention, has broad受众, and once deployed, efficiency will improve very quickly with wider普惠 coverage.
AI is our infrastructure; it will change our lifestyle and work style. I hope our domestic ecosystem — from regulation, practitioners, technology and other dimensions — can create a favorable environment for AI development. Promote marketization, rule of law, and internationalization of the environment. Because without marketization there's no innovation; rule of law can规避 AI's negative factors. And we need internationalization, promoting exchange in technology, products and other dimensions. Finally, don't intervene too much — it's still too early. If directions are predetermined from the start, innovation's possibility is lost.
Source: Huxiu APP
Author: Yu Yang
The full suite of Unity Ventures portfolio robots, "competing on落地" at WRC