Yu Chen on Tech Investing: Beneath the Hype, What Is Long-Term Value? | Yunqi Insights
Finding the Constant in Change
Amid the embodied AI frenzy, which developments represent genuine technical progress and which are froth? With world models, VLA, and other approaches still evolving, and millions of hours of high-quality data yet to materialize, how should investors assess a company's long-term value?
Recently, Yu Chen, managing partner at Yunqi Capital, joined the video podcast "Lai Dian Shen Du" for a conversation with Shao Yujia, managing partner at Shen Du Capital, touching on many of the hot-button issues the industry is currently debating.
Behind these judgments lies a methodology forged over 12 years in tech investing: from cloud computing and robotics to large models, and now to embodied AI and future technologies — how to find directions truly worth betting on long-term as technological paradigms keep shifting.
The conversation began with a Doraemon T-shirt. This animated character who accompanied Chen through childhood also carries his earliest imagination of technology and the future.
The following is adapted from the "Lai Dian Shen Du" interview, edited for clarity.
From Doraemon
To the original motivation for tech investing
Shao Yujia: You're wearing a Doraemon T-shirt — are you a big fan?
Yu Chen: Yeah, I've liked Doraemon since I was a kid. You know how he pulls out gadgets from his pocket? I always wanted to have those.
Shao Yujia: Which ones left the deepest impression? Which did you want most?
Yu Chen: The memory bread. I hated memorizing things as a kid. If I could eat memory bread and know everything effortlessly, that would be amazing.
But look at large models now — humans don't need to memorize knowledge themselves anymore. In a sense, large models are a compression and presentation of human knowledge. We've achieved Doraemon's "memory bread" through another means.
If I had to pick what I want most now, it'd definitely be the anywhere door. I travel so much for work these days. If I could get from Shanghai to Singapore in five seconds instead of five hours, that'd be great.
Shao Yujia: I get the sense you're a very curious person. What else do you and Doraemon have in common?
Yu Chen: Doraemon was my window into the future, and also my original motivation for doing tech investing — hoping to bring some positive impact to the world through technology investment.
Shao Yujia: So what traits make a good investor?
Yu Chen: First, definitely curiosity. Second, you have to have conviction.
Especially for early-stage investors — you accompany founders for at least 8 to 10 years. If you don't truly believe in what the founder is doing, don't believe that good things will inevitably happen in the future, it's very hard to stay the course.
Shao Yujia: Does conviction ever collapse?
Yu Chen: Definitely. The world doesn't always develop the way you expect. Investing itself is a probability game. As long as your odds are better than others', you can survive in this market.
Shao Yujia: What kept you going for 12 years?
Yu Chen: There were moments of doubt. Especially around 2021 and 2022, when technology suddenly hit a bottleneck — there wasn't much exciting stuff coming out. Combined with the whole pandemic situation, not just me but probably the entire industry was in a low period.
Back then I asked myself: if technology in this world stops developing, what's the point of doing tech investing? Where's the meaning?
But very fortunately, at the end of 2022, ChatGPT burst onto the scene and opened up a whole new world for everyone. People started doing AI investing, from AI applications to physical AI, and now to what I'm looking at — Space AI. There's constantly new, exciting technology emerging.
That's what keeps me motivated to continue investing.
No such thing as "good years" or "bad years"
Every year brings new opportunities
Shao Yujia: Over your 12-year investing career, how many cycles have you experienced? What major changes have there been in investment strategy across different stages?
Yu Chen: From a broad capital markets perspective, there's generally a cycle every 4 to 5 years — you could broadly define this as bull and bear markets in the secondary market, and primary and secondary markets are somewhat linked.
But from an industry perspective, I don't think there have been so-called "good years" or "bad years" in the past 12 years. Because every year, certain industries develop well; in other years, different industries develop well. You can always find industries that are doing well.
From a broad capital markets perspective, there's generally a cycle every 4 to 5 years — you could broadly define this as bull and bear markets in the secondary market, and primary and secondary markets are somewhat linked.
But from an industry perspective, I don't think there have been "good years" or "bad years" in the past 12 years. Every year there are industries growing fast — just the distribution of opportunities differs.
Personally, I first focused on cloud computing opportunities. Back then enterprise services were migrating to the cloud at scale, driving demand for cloud computing infrastructure.
Around 2016, I started paying attention to robotics. Because demographic trends were gradually becoming visible, I realized China might face declining birth rates and aging populations, and robots could become an important way to supplement the labor force.
Following the robotics direction, I also started looking at autonomous driving. In a sense, autonomous driving was among the earliest embodied AI opportunities.
But as everyone can see, it typically takes nearly 10 years for a technology to go from lab demo to commercialization. Throughout this process, technology keeps maturing and costs keep falling. Only when something becomes cheap and good does commercialization truly happen.
So when we look at embodied AI now, we also believe it needs at least another 5 to 10 years before large-scale adoption.
What early-stage investing competes on
Isn't information asymmetry, but cognition asymmetry
Shao Yujia: So at different stages, what were you focusing on in the mainstream areas?
Yu Chen: Personally, I first focused on cloud computing opportunities. Back then enterprise services were migrating to the cloud at scale, driving demand for cloud computing infrastructure.
Around 2016, I started paying attention to robots. As everyone knows, in 2015 the country abolished the family planning policy that had been in place for nearly 30 years. More sensitive people might have thought: China's future population structure could see declining birth rates and aging, so robots might have opportunities to supplement the labor force.
Following this direction, I also started looking at autonomous driving. In a sense, autonomous driving was among the earliest embodied AI opportunities. But as everyone can see, it actually takes nearly 10 years for a technology to go from lab demo to final deployment.
During those 10 years, technology keeps developing and maturing, and costs fall rapidly. Only when something becomes cheap and good does commercialization truly land. So when we look at embodied AI now, we also believe it needs at least 5 to 10 years before adoption.
Shao Yujia: You just talked about enterprise services and that wave of cloud computing — do you think that wave had quite a bit of bubble?
Yu Chen: Actually every wave has bubbles. Without bubbles, it's hard for an industry to achieve major development. Because of bubbles, capital pours in resources at scale, attracts more talented people, and drives industry development.
Shao Yujia: Compared to other investors, what traits do you have that enable you to invest in better companies?
Yu Chen: I think it's being pragmatic. I come from an engineering background, so I'm not easily fooled on the technology side.
A lot of people are starting companies now, with new buzzwords constantly emerging and very impressive resumes. If you truly don't understand technology or the industry, it's easy to be influenced by these backgrounds and the visions founders paint — easy to get carried away.
Shao Yujia: Has your Google engineer background played a role in specific project judgments?
Yu Chen: I think a classic example is PingCAP. PingCAP builds distributed databases — essentially an open-source implementation of the distributed database Google uses internally. When it was raising its Series A, it talked to almost all the mainstream funds, but many institutions didn't have relevant background, didn't know what they were doing, or understand the importance of this.
When I came across this project, I roughly knew that it corresponded to a very important category of storage systems inside Google. This was very challenging, but once built, could become very large.
Shao Yujia: So because you were once on the front lines of the business, you could get closer to where the real demand is.
Yu Chen: Right. We often say early-stage investing is about information asymmetry, but I think fundamentally it's more about cognition asymmetry.
Shao Yujia: What other projects do you feel your experience gave you a significant judgment advantage on?
Yu Chen: MiniMax might be an example. When I first encountered the company in 2021, I had an intuition: large models might represent a paradigm shift in technology, even disrupting the then-dominant small model or proprietary model approaches.
Later the industry's development exceeded expectations. But at the time, past experience gave me a certain intuition about what they were doing, and helped me make the investment decision relatively quickly.
Betting on people, but also seeing whether they're
Standing in the right direction
Shao Yujia: Many people wonder why Yunqi is often able to find and invest in excellent companies in the first round or very early stages?
Yu Chen: First, we insist on investing early and small. Also, my advantage still lies in accumulated information and cognition.
On one hand, my engineering background brings technical cognition advantages; on the other hand, over ten years of investing has also built personal brand. Many founders are more willing to talk with me early on. With shared cognition, it's easy to build mutual trust.
Shao Yujia: So on average, how long have you known most founders when you make investment decisions?
Yu Chen: This varies. But many of the larger investments are with people I've known for 5 or 10 years or more.
Shao Yujia: That's actually somewhat surprising. Because many investors may need to complete a formal decision in relatively short time, but some of your important projects are actually judged on the basis of long-term understanding.
Yu Chen: In a sense, this is also a form of due diligence. But when you first meet someone, you're not necessarily aiming to invest in them in the future. If you think this person has potential and is worth long-term relationship, then build that relationship — don't be too calculating.
Shao Yujia: So returning to the essence of investing, it's still about betting on people?
Yu Chen: The赛道 still has to be chosen correctly. Perhaps in less technology-driven eras, the quality of the person determined a lot. People might assume: as long as it's a good person, whatever they do could succeed, they could keep changing directions.
But now, especially in hard tech investing, much entrepreneurship requires accumulated knowledge and experience. This also means the difficulty for founders to switch tracks is much greater than before.
So now you have to look not just at the person, but also whether their initial赛道 choice is correct, and whether their experience matches the capabilities needed for success in that赛道.
Shao Yujia: Over the past 12 years, has your ability to read people improved?
Yu Chen:
I think it's like machine learning. The more positive and negative samples you see, the more accurate your model becomes.
From our perspective, we look at several characteristics in founders: first, whether they're smart enough; second, whether their ambition is large enough; third, and most importantly, whether they have solid character. It's not that you can't succeed with bad character, but usually for investors, it's very hard to make money on founders with bad character.
Embodied AI:
Technical approaches change, but scenarios don't
Shao Yujia:
Getting back to embodied AI. How do you think about judging which embodied AI company is more likely to succeed?
Yu Chen:
I think embodied AI entrepreneurship has actually gone through two waves. The first wave was around 2023 and 2024.
At that time embodied wasn't a hot赛道, with relatively few people paying attention and fewer targets. Our judgment then was actually the same as when looking at other tech companies — we'd deeply explore technical details and judgments about the future.
Because back then these founders couldn't reference and learn from each other like today, and there weren't so many同质化 routes, so you could still see differences between teams and pick the ones you believed in.
But by 2026, you see everyone piling into world models. Everyone's background looks great, with many papers published and many citations. But at this point, investing actually becomes more difficult. From our perspective, we're more cautious about any single technical approach. Because like autonomous driving, over the past 10 years from lab demo to final commercialization, the technical approach changed several times in between.
Similarly, we feel VLA, world models — they may just be one technical means on the way to the endgame, not the final answer. If you invest in a specific technical approach now, and it switches in the future, much of that investment could be wasted.
But when an industry has already surfaced and formed consensus, you need to find true differentiation within it.
Models change
Data is the long-term moat
Shao Yujia:
Many believe the next stage of core competition in embodied AI will come down to data and models. What's your view?
Yu Chen:
I think data's value is always important. Including now, the more advanced the model, the higher the requirements for high-quality data.
Looking back at large language models, everyone is vigorously purchasing user behavior data, and at high prices too. Because as foundational data gradually approaches limits, people pay more attention to so-called expert data, or real human usage data. I believe embodied AI will also reach this point.
Shao Yujia:
Why is robot data harder than large model data?
Yu Chen:
For language models and video models, acquiring large-scale foundational training data is relatively lower cost. But in the embodied robotics domain, whether collecting real robot data or current Ego-centric data collection, both require labor costs and time costs. So if you want to collect sufficiently scaled foundational training data, the cycle will be very long.
Shao Yujia:
When do you think the "GPT moment" for robots will arrive?
Yu Chen:
I might be relatively conservative. I think perhaps around 2030. Of course, if technology develops very fast, it could be earlier, but at least for now, it still needs some time.
Shao Yujia:
If you want to achieve truly general-purpose robot models, roughly how much high-quality data is needed?
Yu Chen:
I think it might be at the tens of millions of hours level. But this depends on how general and generalizable you want your model to be. If today you're just doing a constrained scenario, like household services or logistics, the data needed would definitely be much less. But if you want to do truly general models covering common human scenarios, I think tens of millions of hours is unavoidable.
Shao Yujia:
If a model has 1 million hours of high-quality data, would it be better than what many embodied brain companies in China have produced?
Yu Chen:
I think definitely. Because many embodied brain solutions in China haven't actually used 1 million hours of data yet.
Shao Yujia:
So to some extent, do many companies actually underestimate data's value?
Yu Chen:
I think data's value is always important. Many companies now focus on model architecture, but actually data quality directly affects model intelligence. For example, feeding 1,000 hours of data into a model for post-training — the final results could be completely different.
Shao Yujia:
But if you don't do pre-training, don't robot companies need to spend that much money?
Yu Chen:
But the risk of not doing pre-training is: if advanced models like Pi stop being open in the future, you might be stuck at your current level forever.
Shao Yujia:
What's your view on opportunities for data companies? You've also invested in a data company.
Yu Chen:
It was actually relatively early in using Ego-centric approaches for data collection. When I invested, it had certain scarcity value. Additionally, a good data company needs to meet several conditions: first, it needs to be independent — only with independence can it serve a larger customer base; second, it must itself understand models — only by understanding how customers use data can it produce higher quality data.
Shao Yujia:
If major tech companies open up high-quality data in the future, will data companies still have opportunities?
Yu Chen:
I'm actually not worried about major tech companies. They may not invest heavily in human resources for this kind of highly customized data collection work. Look at the large language model space — companies like ByteDance and Alibaba also purchase data externally, because for them the ROI is higher; they shouldn't invest large amounts of high-cost human resources in these things.
Shao Yujia:
So will the next few years be a window of exploding data demand?
Yu Chen:
Right. But in the process, many low-quality suppliers will definitely be eliminated.
Shao Yujia:
Then what do you think about many leading embodied companies building their own world models?
Yu Chen:
The answer to this is affirmative. For leading companies, VLA, world models — these are tools in the toolbox. If they think a tool is useful, they'll definitely build it themselves. So I don't think there will be many independent third-party world model companies. In China in the future, things will likely converge. Like Apple — it does both hardware and the brain inside.
Amid the embodied frenzy
How to stay rational?
Shao Yujia:
How long do you think this wave of embodied AI funding frenzy can last?
Yu Chen:
This depends on the secondary market. Most likely by the end of this year or early next year, the first batch of embodied companies will go public. If these companies perform poorly in the capital markets, it will quickly transmit to the primary market.
Shao Yujia:
So for many embodied companies now, the competition is essentially also a capital competition?
Yu Chen:
Actually embodied AI and large models are somewhat different. Large models are now an "arms race." Everyone might train a 3T model this year, then a 10T model next year. The demand for GPUs and data is very clear — without these resources, you can't produce SOTA models.
But embodied AI is still in early exploration. People haven't truly entered large-scale investment stages yet, and don't know where money should actually be spent. There are also many embodied brain solutions on the market, mainly doing post-training based on open-source models.
Shao Yujia:
So how do you view this situation?
Yu Chen:
In a sense, for mid-to-late stage investors, this is also a due diligence method. Look at their bills: how much spent on data? How much on GPU investment? You can basically judge whether they're seriously doing pre-training or mainly doing post-training.
Shao Yujia:
In your fund, what's the ratio between first checks and follow-on investments?
Yu Chen:
First checks definitely make up the bulk. One important reason is that in the current market environment, valuations are growing too fast. Only by investing early can you get enough ownership.
Shao Yujia:
What's the most expensive company you've invested in at the first round, roughly what valuation?
Yu Chen:
We generally cap around $200 million. Above $200 million means when you go public, you have to reach over $10 billion.
Shao Yujia:
Equivalent to needing 50x growth space.
Yu Chen:
Right, so founders themselves also need to weigh whether they can ultimately achieve that scale.
AI hardware returns to rationality:
From subscription imagination back to business fundamentals
Shao Yujia:
Smart hardware went from booming last year to gradually returning to rationality this year. What's behind this shift in logic?
Yu Chen:
I think there are two main reasons. First, many new AI hardware products that emerged last year didn't actually meet end consumer needs, so subsequent sales didn't meet expectations. Second, supply chain issues. Memory prices rose relatively fast, which affects delivery of many smart hardware products.
First, can you even get the components? Second, even if you can, what premium do you have to pay? These all impact smart hardware.
Shao Yujia:
Some overseas AI hardware companies are valued on SaaS subscription logic, but many domestic hardware companies are still on product sales logic. Are these valuation systems different?
Yu Chen:
Right. If it's a one-time sale, it essentially returns to pricing logic similar to Roborock or ECOVACS — maybe 20-30x PE. It shouldn't be valued on SaaS logic. Ultimately it comes back to profitability.
Facing the future:
AI, Space AI
And new capabilities for everyone
Shao Yujia:
What directions are you focusing on now?
Yu Chen:
From my personal and team perspective, there are several key directions. First is embodied AI; second is smart hardware; third is AI for Science; fourth is Space AI. Especially Space AI — I think this is a future direction.
Shao Yujia:
What opportunities does AI have in the space domain?
Yu Chen:
First you need to solve rocket launch capacity. Because if you want to develop computing satellites in the future, you must rely on reusable heavy-lift rockets to bring costs down, otherwise the entire economic model doesn't work. On this foundation, there are many directions: computing satellites, satellite communications, space embodied AI, space 3D printing, thermal management solutions, etc. These are all directions worth exploring in the future.
Shao Yujia:
As a veteran AI investor, what's your advice for ordinary people?
Yu Chen:
First, you need to master AI as a tool. For example, can you clearly express your ideas, and when AI gives you a result, can you judge whether that result is usable? I think the capabilities ordinary people most need to master in the AI era are structured expression ability and judgment.
Shao Yujia:
This year's Agent development has many white-collar workers worried about whether their jobs will be replaced. How do you think people should face the robot era?
Yu Chen:
As long as you make good use of AI, making your output higher than your colleagues, you can establish yourself in the company.
Shao Yujia:
If robots become more mature in the future, what changes will happen to human society?
Yu Chen:
I think human productivity in the future will improve by 1 to 2 orders of magnitude. Because with AI and robot加持, people can gain greater income with less working time, spending more time on things outside of work.
For individuals, the most important thing is learning how to skillfully use AI tools. This is a continuous, lifelong learning process, because AI itself is also constantly changing.
