Linear Capital Perspectives | AI Enters Industry: Is the Current Moment a Window of Opportunity or a Bubble, and Which Investment Areas Does Linear Capital Favor?

At the end of June, Harry Wang, founder and CEO of Linear Capital, was invited to speak at the Amazon Web Services (AWS) Innovation Conference. He joined Dr. Xia Zhang, chief enterprise strategist at AWS, for a conversation about recent AI investment insights. They discussed Linear Capital's analysis of AI's integration into industry, the firm's investment focus areas, technical bottlenecks to overcome in the next phase, and related recommendations. Below is a summary of their dialogue.

Dr. Xia Zhang: With capital markets fluctuating and the global economy moving through uncertainty, looking back at the past decade, what are your reflections? What stage are we in now?

Harry Wang: Let me start with my conclusion. I believe we're in an adjustment phase — coming down from a peak and preparing for the next one. Compared to the peak two or three years ago, we've definitely come down. But compared to where we'll be two or three years from now, today is a preparation stage.

Let me briefly map out the past decade. I returned to China in 2012, right at the beginning of the past ten years that Dr. Zhang mentioned. I was previously an engineer. At Facebook, I led the payments security team and was among the core engineers on the first backend versions of Facebook Newsfeed and the ads system. That experience was primarily from a technology practice perspective. The first two or three years after I returned, in my view, were the most critical period for mobile internet's rise.

By 2015 and 2016, mobile internet investment reached peak levels. Those years, I believe, were about the mobilization of business models. Smartphones entered millions of households, bringing a wave of tech product investment and model innovation built on new products. Companies like Dianping — I served as a CEO advisor there for a year — Meituan, WeChat, these names everyone knows, all rose during that era. At a higher level, what we saw was the digitization of ordinary people's lives.

Then, from 2015–2016 to around 2018–2019, I believe two major shifts occurred.

First, digital products gradually became familiar consumer goods. New mobile internet-based commercial opportunities dropped significantly. The growth and IPO of companies like Pinduoduo, then considered newcomers, marked in my view the closing chapter of the mobile internet era. You might occasionally stumble upon a gem, but in terms of opportunity density, the peak was 2015–2016.

Second was enterprise digitization. Previously it was ToC — individual digitization. Around 2016 to 2018 was when enterprise digitization and enterprise intelligence began developing in China.

From 2018–2019 to 2021, I see a progression from enterprise digitization to true enterprise intelligence. AI deeply integrated with real problems inside enterprises. This required technology to enter industry, collaborate with industry, and connect previously abstract concepts — AI, digitization, data, algorithms — with concrete industrial problems that enterprises actually faced. At the same time, this digitization and intelligence process surfaced many other technology issues within enterprises. So over the past three to four years, another trend in enterprise intelligence was hard tech. "Hard tech" has become a household term again.

I broadly divide the past decade of domestic tech product innovation into these three stages. The next 3–10 years will be the opportunity for deep tech, hard tech, and AI to deeply integrate with industry — what we consider the golden decade of tech investment.

Dr. Xia Zhang: In recent years, in the field of AI investment, what developments and changes have you observed? There are bubble theories and trend theories in the market. What's your view?

Harry Wang: Let me lead with my conclusion. I believe泡沫 are inevitable, but there are two kinds. One is blown bubbles — when they pop, nothing remains. The other is beer foam. After you drink the foam, there's real beer beneath. This latter kind of foam has flavor, has value.

I believe AI has泡沫, but it's the latter kind — beer foam.

Indeed, many investors and entrepreneurs who rode the hype will exit. Because in the process, they'll discover that landing technology, integrating with industry, and delivering value to industry is difficult and grueling — not a quick-money endeavor. The waves will wash away speculative entrepreneurs and investors. Those who can remain grounded will stay and prove their worth.

For entrepreneurs and investors, I believe two mindsets are needed: patience, and the ability to grasp first principles. Truly understanding, with genuine effort, where industry problems actually lie. Whether there's a connection point between the problem and AI. Where the big data is. So-called artificial intelligence is data processing data. What we call the output is actionable insights — whether it can help enterprises predict, or accelerate decision-making, improving both speed and quality. This requires a mindset of going deep into industry, yet not being constrained by industry's existing thinking patterns. You need a technology mindset, a data mindset, an AI mindset. So here, I believe the new opportunities set higher bars for both investors and entrepreneurs. This is precisely why I view this as the best moment to separate real gold from fool's gold.

Dr. Xia Zhang: Harry, I recall you have a term you care deeply about — "gray tech," defined as black tech that can land in industry. From the technology layer, application layer, and specific investment project perspectives, could you share what investment opportunities you've recently favored in AI, machine learning, and related industry applications in the broader intelligence space?

Harry Wang: On the broad directions our investments focus on, let me give two categories we currently weigh heavily.

One is newer scenarios we've only emphasized in the past 3–4 years. The other is domains we've tracked since Linear's founding that sit very close to people's daily lives.

The first category — what's new? For example, the combination of AI, big data, and biomedicine. Biomedicine involves extensive gene data capture and analysis, with massive data volumes. This data doesn't emerge naturally; it must be obtained through extensive experimentation, high-throughput experimentation. You get massive data and data processing opportunities. This data must also integrate with industry-specific characteristics. Throughout this process, whether data storage or data computing, opportunities exist. These specialized domains have enormous data volumes and professional data, so you must enter industry and obtain it through professional means.

The second category is domains inherently close to people's lives. Marketing-related, for instance. Marketing involves extensive analytical work. Our earlier investments include Sensors Data, Guandata, and Shushu Technology. In recent years, we've invested in Jingshuo Technology and JINGdigital, a company focused on the ToB marketing space. Traditionally, marketing analysis meant defining the product and defining potential customers — this was when data was scarce and granularity was limited. But in the big data era, you can obtain high-granularity data to understand people's preferences and behaviors. This massive, high-granularity data helps enterprises find the intersection between human needs and product positioning. In this space, we've invested heavily in tech enterprises that enable different clients to better understand users, understand products, and understand how to get products to the right customers better, faster, and more accurately — and through what marketing methods to reach them, stringing the entire chain together. Marketing itself represents trillions in annual commercial spending. This has been, in my view, a gold mine over the past decade, and will remain one for the next decade.

Both categories, as I've explained, involve massive data. Only by understanding what this data is can you truly apply algorithms, and connect algorithm-derived conclusions with real problems in industry. This process is what turns so-called black tech into "gray tech" — truly landing, solving concrete industrial problems. Otherwise, divorced from industrial problems, from data, from business decisions, boasting that your algorithm processes however many transactions per second is meaningless.

Dr. Xia Zhang: For the next phase, beyond healthcare and marketing you've already introduced, are there other interesting topics?

Harry Wang: For how frontier technologies represented by big data, AI, and robotics integrate with industrial problems, we have a structured framework we call "1+1+1."

The first 1 is a foundation. This big data and AI methodology requires new data infrastructure. This includes how data is acquired, how it's transmitted — because the granularity of this data, the frequency of collection, is completely different from ten years ago. So to engage with big data and utilize it well, there must be innovation at the foundational layer from acquisition, transmission, and storage through to computing. People using data should experience it like a faucet — turn it on and it's ready. Others can then focus on algorithms above and thinking about how to integrate with business processes.

On top of this foundation, we divide into two major domains. One we call commercial applications — where data exists, insights are found directly within it, and applied directly for commercial purposes. We have a concept called DaaS, Decision as a Service, which we proposed building on SaaS. The next generation of enterprise software should be built on massive, high-quality, high-granularity data as its core — not merely creating new cloud-based software that migrates existing usage habits online. That, in my view, is completely insufficient. Furthermore, in building this new software, you must start from business decisions, considering all necessary data comprehensively. The ultimate purpose of data acquisition is to serve business decisions.

For example, we invested in a recruiting company. It processes and optimizes resumes, matching them to requirements. For those with large recruiting needs, matching becomes machine-driven rather than human-driven. Traditional ATS — applicant tracking systems — merely manage your process, but efficiency remains unchanged. Previously you managed with Excel; now you use cloud software. We believe this is insufficient. It must be core-capability based on data prediction and decision-making ability.

The other category, the third 1, we call industry application — industrial-nature applications. We use this term to describe applications where, beyond big data and AI, mechanistic models are also applied. This involves a deeper concept. Big data AI, deep learning — these deal in correlation. They don't answer causality. Understanding causality requires understanding the principles of things themselves — what Elon Musk particularly emphasizes as "first principles." For example, we invested in a company called Aikennong that uses AI to assist agricultural planting decisions. Beyond massive crop-related data, what's crucial is the natural mechanistic model of crop growth. Without this mechanistic model, big data is extremely inefficient because it must first search for patterns in massive data. But with a mechanistic model, you only need to find conclusions within the data space permitted by that model. So in industry application, we emphasize combining big data models with mechanistic models to find more efficient ways to improve industry productivity with multiplied effect.

Dr. Xia Zhang: For AI's next phase of development, at the technical level, what bottlenecks still need to be overcome?

Harry Wang: Right. One issue we see is that for AI to be used effectively, massive data is needed. But in many practical applications, that much data simply isn't available. This has become a huge constraint on AI landing.

Here I have three recommendations. First, pay attention to academic areas like few-shot learning, including few-shot learning, transfer learning, reinforcement learning, and so on. Their core is minimizing learning costs, training costs, and convergence time as much as possible.

Second, as mentioned, if you want to turn black tech into gray tech, you must go into industry and find effective ways to combine mechanistic models with data. Many of Linear's portfolio companies do this — entering industry to find the intersection point between mechanistic models and data models. This is where AI practitioners need to invest serious effort.

Beyond finding this technical intersection, there's also talent matching. You understand big data and AI — but do you have a COO or partner who understands industry? In previous eras, like the internet era, this wasn't necessary. Nice to have, but not required. Investors could invest regardless. But in today's era of AI entering industry, if such a person doesn't exist, that company or team may become un-investable. This is what I want to emphasize, responding to my earlier point that you must find ways to enter industry, understand its mechanistic models, and seek intersection points.

Third, bring in industry players and customers early. We have a concept that all technology itself has no moat. Useful technology is problem-oriented. If you don't even understand what problems industry has, or what problem your technology solves, even pushing your technology to the pinnacle — except perhaps for general-purpose databases — actually won't work. Even databases are hitting many technical bottlenecks now, because relational databases are failing in many domains under new product demands. So what you see is that many general-purpose technologies carry underlying assumptions about application scenarios, and these assumptions are now entering a state of collapse. So focusing on technology must genuinely start from focusing on the problems technology faces. Where do these problems come from? Bringing customers in early matters because sometimes customers may not give you answers, but they can explain what problems they encounter, what their pain points are. Bringing them into your entrepreneurial process early gives you better opportunity to understand problems. Better problem understanding enables better technology trade-offs — where to go deep and build thickness, where to use what others have already built. This makes many technologists' work more efficient. Otherwise, divorced from this, pursuing absolutely universal technology becomes meaningless. These three recommendations are what I'd propose for technical-level thinking.

Dr. Xia Zhang: As an investor, how do you select your investments, and how do you help them succeed after selecting them?

Harry Wang: We have many detailed criteria, but given time constraints I'll skip those. Let me share our two broad, "principle"-level pursuits. One is right problem. The second is right people.

Right problem means problems we believe, if solved, would bring massive efficiency gains or breakthroughs to industry — where technology can deliver massive leverage. Not all problems merit big data or AI approaches.

Second is right people — the right individuals. Here we believe three things matter most.

One is truly understanding industry, and being professional on that foundation — this is the professionalism dimension.

Second, we really require commitment — genuine dedication, that "if not me, then who" spirit of full investment. The entrepreneurial process is extremely difficult. Having good technology doesn't guarantee winning. Are you prepared to fall seven or eight times before winning? Do you maintain that mentality of "entrepreneurship tortures me a thousand times, I treat entrepreneurship like first love" regardless of difficulties? This commitment is something we value highly.

Third, we have a core requirement on character — we call it consistency. Thoughts, words, and actions aligned. Investing is also making friends. We hope to make friends who are candid and genuine, not people who only tell you good news and hide bad news. If others can truly treat you as a sincere, mutually supportive friend, life loses much of the trouble of pretense. These three broad points are our core criteria in team selection.

Dr. Xia Zhang: What advice do you have for those wanting to work in AI and other technology innovation?

Harry Wang: I think it's very important to find ways to learn well from those who've done much practice in this area and have genuine substance. If you're a practitioner or engineer, find ways to join a company genuinely working on turning data and AI black tech into gray tech. Understand how the entire feedback loop forms effectively, how technology lands. If you're an investor, don't just think about how much money you'll make in two or three years after investing. Genuinely care about industrial problems, and invest in two or three companies landing these technologies.

Simply put, there's only one core thing: practice. Sincerely find people meeting the three criteria I mentioned — either start businesses with them, join them to work, or join them through investment. Making this loop work requires enormous energy and effort on every problem, and the entire cycle is very long. You must have patience. So I hope more people join this practice.

---

About Linear Capital

We are hiring for our investment team. Multiple positions are currently open. For details, please click here.

Linear Capital is a professional investment institution focused on "Data Intelligence" and "Frontier Technology."

Linear Capital currently manages ten funds with total assets under management of approximately US$2 billion.

We focus primarily on early-stage projects in "Data Application," "Data Infrastructure," and "Frontier Technology" application domains. Our investment stage focuses on leading angel to Series A rounds, with typical investment sizes of US$3–8 million or RMB equivalent.

To date, we have invested in over 80 entrepreneurial teams at early stage, including Horizon Robotics (US$3B), Tongdun (>$1B), Kujiale (>$1B), Sensors Data, Tezign, Rokid, Guandata, and Agile Robots. The combined valuation of Linear's portfolio companies is approximately US$15 billion.

In the near term, Linear Capital is working to become the best "Data Intelligence Technology Fund," and in the long term, gradually build into the most influential "Frontier Technology Application Fund."