Consensus Within Non-Consensus: A Conversation Among China's LLM Investors | Yunqi Tech π

"The reason we invested in large models before ChatGPT took off is that I believe the world will always need general-purpose intelligent agents. While today's large models may not be the final solution, they will take us a big step in the right direction."

These remarks were made by Chen Yu, partner at Yunqi Capital, at 36Kr's WAVES 2024 conference. During the event, investors behind China's large model unicorns engaged in a lively exchange of perspectives. As an angel investor in MiniMax, Chen also shared his observations and reflections on large model entrepreneurship.

This edition of "Yunqi Tech π" brings you highlights from that conversation.

Three bullet points

How do you maintain your own rhythm amid industry volatility?

Chen Yu: Think independently. Don't let others set your pace. Start from first principles and consider what's actually useful and feasible.

What views on today's AI differ from market consensus?

Chen Yu: For some markets, energy is more important than compute power at this stage.

Where is the survival space for large model startups competing against tech giants?

Chen Yu: You have to find differentiation in product and business model from the major players. Differentiated competition is the crucial direction for startup survival.

Two more things

WAVES also announced the "36Kr China Equity Investment Annual Rankings," formed by merging 36Kr's two flagship lists: "Most Entrepreneur-Friendly Investors" and "Most LP-Recognized Investors."

Yunqi was selected for "Top 15 Early-Stage Investment Firms in China's Equity Investment Industry."

Additionally, three Yunqi portfolio companies riding the AI wave were named to 36Kr's 2024 Under 36 class:

MiniMax founder and CEO Junjie Yan

Sobot co-founder and CEO Yi Xu

Yunqi Capital CTO Tongyi Cao

The following content is from Dark Currents WAVES

Title: When the Investors Behind Moonshot AI and MiniMax Sat Down Together

Over the past 18 months, AI has raced ahead at breakneck speed in China. Moonshot AI and MiniMax in particular are often compared: both appeal more to dollar-fund sensibilities, and both have moved faster on fundraising.

At last week's WAVES 2024, we invited investors from both companies to join us on stage: Yusen Dai, managing partner at ZhenFund; Yungang Huang, managing partner at Source Code Capital; Rui Han, partner at Gaorong Capital; Chen Yu, partner at Yunqi Capital; and Ling Xia, partner at Future Capital. The first three invested in Moonshot AI; the latter two invested in MiniMax.

From this conversation, we see the fascinating intersection of technology and investment: the hunger and non-consensus that comes with everything still being early. When discussing their "initial decisions," each person's reasoning was completely different: some believed in AGI, some believed in "end-to-end," some simply "believed in this person."

Though they repeatedly emphasized they weren't trying to "stoke flames," you could feel the divergence in perspectives. On general-purpose robotics — an industry tied to the AGI concept — some investors on stage had already deployed capital in multiple companies, yet Yusen Dai took a clear stance that "the industry is too early for VC."

The AGI industry is indeed still in its early days. And precisely because everything is still so early, some questions are genuinely difficult to answer for entrepreneurs and investors in the thick of it — after all, whatever they say becomes part of the historical record.

WAVES is a new summit IP launched by 36Kr last year; this was its second edition. The conversation was moderated by Liu Jing, editor-in-chief of Dark Currents WAVES.

The following is an edited transcript:

Part 1

Dark Currents: Hello everyone! Welcome to WAVES. Before coming on stage, I promised our five guests that we definitely won't be stoking any flames today. Because everyone here is either an investor in Moonshot AI or MiniMax. But we also want world peace, so we deliberately mixed up your seating to avoid things getting too tense. Let's start with each of you introducing yourselves in your own way.

Yusen Dai: We specialize in angel investing, targeting Chinese entrepreneurs globally. We've made many angel investments in AI, plus AI applications — we're prepared.

Chen Yu: I'm your typical engineering nerd. I hit it off with IO (MiniMax founder Junjie Yan) when we met in 2021, because we were both talking about computer science stuff and really clicked, so I invested in Junjie. We focus on early-stage tech investing and have invested in many AI projects. The projects Yusen invests in, we have different benchmarks for each.

Rui Han: Over the past decade, Gaorong Capital has focused on projects that make people's lives and the world better. In the past year, we've made very significant investments in AI. We believe this is just the beginning, and hope more good things are yet to come.

Ling Xia: For the past decade, we've focused on just one thing: early-stage investing in tech. We were fortunate to catch the biggest opportunity of the past 10 years starting in 2014 — intelligent electric vehicles. We were Li Auto's earliest institutional investor.

We also positioned early for this AI wave. In early and mid-2022, we invested in large language model company MiniMax and humanoid robotics company LimX Dynamics. This morning, Mr. Huang and MiniMax founder Junjie Yan had an excellent conversation. The biggest investment opportunity over the next 8-10 years will be AI, and we'll continue systematic investment and positioning in the AI track.

Yungang Huang: We particularly value AI and everything AI enables, both software and hardware. We've invested in Moonshot AI and Galaxy Universal, an embodied intelligence robotics company.

Dark Currents: Our theme is large language models, and we can start by discussing these two companies. Could each of you recall your investment process for MiniMax and Moonshot? Because there are voices suggesting large language models may not be a VC game — most VCs in Silicon Valley missed OpenAI's earliest stage. When you went through IC, what was the single most important thing you said?

Yusen Dai: We've always invested in people, and we're very focused on investing in young people — fitting today's theme. We believe major waves are often driven by young people. OpenAI's founders, for example — one was 30, one 29, one 28, very young. We've always sought out the most outstanding young people. Zhilin Yang — we noticed him back at Tsinghua University. He was the god among gods across several cohorts there, and at CMU he was a very well-known international first-tier AI researcher with excellent work.

When he was doing his PhD at CMU, he participated in founding Recurrent AI. We invested in Recurrent AI because of Zhilin Yang — ZhenFund was Recurrent AI's angel investor. When Zhilin moved into large models, we believed this was a major wave that favors young people. We had already been watching the most outstanding people, so we invested. It was an easy decision for us.

Dark Currents: ZhenFund has a theory about investing in people. Does he qualify as a "young genius"?

Yusen Dai: Of course. He fits our definition of a young genius because he's young, so he can be the best in a new wave — this is a new wave. If you're doing real estate now, it's hard to be a young genius. We hope this genius isn't qualified by age — not the best post-90s entrepreneur, or best post-00s entrepreneur, but simply the best entrepreneur, the person who understands AI the most.

When Bill Gates built Microsoft, he was the best programmer. When Mark Zuckerberg built Facebook, he was the best entrepreneur. We hope excellence isn't modified by age. We believe Zhilin fully meets these criteria — he's a world-class AI scientist.

Dark Currents: The key is the person.

Chen Yu: When we talked with Junjie Yan, he said he wanted to build the Chinese version of OpenAI. He spent a lot of time explaining what large models are and why large models would be more effective than countless small models combined. I also have a programming background, so we talked very well together — a bit of kindred spirits.

When I invest in tech companies, I'm used to gauging technical grasp through conversation. Even though they're working on very cutting-edge technology, the fundamental grasp of technical details — this is something I value highly. When I worked at Google, I saw what some of the world's top programmers were like, and I compare founders against that standard. It's easy to filter out the technically strongest entrepreneurs.

Beyond his technical pursuit, Junjie started thinking about productization and commercialization very early. He's not someone who purely focuses on technology — this is something I admire about him.

Rui Han: We've been waiting to see whether any large language model product could truly enter and stay in people's lives. We conducted an internal survey about Kimi and found that not only Gaorong's investment and research teams, but also colleagues in IT, HR, PR, and other departments were all using Kimi. And quite a few of them had unsubscribed from GPT to switch to Kimi. This was one small piece of our decision-making.

Is VC a suitable player in the large model game? As Yusen mentioned, VC funding is a small piece in both China and the US. Let me stitch together two phrases to express my view: don't fail to do good because it seems small; when everyone adds firewood, the flames rise higher.

If I had to say one thing at IC, my answer would be: China must and will have its own AGI. If this doesn't work out, any valuation is expensive; if it works out, none are expensive.

Dark Currents: The trend itself matters a lot.

Rui Han: Yes.

Ling Xia: Mr. Huang mentioned this morning that in October 2021, three of us from Future Capital met Junjie Yan, and I was the only one who understood what he was talking about. There's a reason behind this. As is well known, we invested very well along the intelligent electric vehicle line, and I was responsible for that track.

If you follow autonomous driving development, you probably know that in 2021, the industry had a landmark change: Tesla's transformer-based BEV and the end-to-end, data-driven perception approach. This was relatively unfamiliar to investors in other sectors. But for an investor focused on vehicles and autonomous driving, this was not unfamiliar at all. The question was simply whether migrating an end-to-end, data-driven paradigm from autonomous driving to NLP would work.

Coincidentally, a core member of MiniMax's early team had previously worked on autonomous driving at Uber. When I pushed this internally, I never viewed MiniMax as China's version of OpenAI. We saw this as a new wave of technology. The "large" in large models isn't essential; China's OpenAI or China's version of any company isn't essential. What's essential is the new paradigm of end-to-end, data-driven approaches behind it.

My internal logic has been consistent: how we viewed autonomous driving, how we viewed NLP, and starting in mid-2022 how we viewed end-to-end, data-driven robotics — the logic behind this is coherent and continuous. We're investing in the new generation of end-to-end, data-driven technology paradigms. This is the consensus I hoped to build internally.

Dark Currents: Ling Xia is being very honest. Mr. Chen views these companies from an AGI perspective, but Ling Xia is talking about end-to-end — this is the insight drawn from Tesla in the automotive industry.

Ling Xia: Precisely because I'm viewing this from the end-to-end, data-driven new paradigm perspective, I pay more attention to the team's engineering capabilities. Junjie has been through 0-1-100 in research, engineering, and commercialization. He's always thinking about how to do things in an engineering-driven way — this is a very prominent trait of his.

Including how to use one-tenth the cost to acquire ten times the data, and a series of similar thought patterns — these are all very engineering-oriented. And at this particular moment, LLMs should indeed be advanced from research to engineering-driven development. Our perspective matches his.

Yungang Huang: We invested in Moonshot AI during the Lunar New Year, but we'd actually been talking for a long time before that. When we spoke, besides discussing who would be the AGI leader in China and what the vision was, we also had to talk about concrete execution — the LLM landscape, how they understood future productization and commercialization. We're fairly pragmatic. What they described at the time gradually materialized over the following six months, including product capabilities. Once they reached that point.

Because LLMs are an extremely capital-intensive business, the team needed to prove strategic and execution capabilities, while also having more investors come in to support them. Reaching that point was the optimal moment, so we invested.

Part 2

"Dark Tides": For ZhenFund, reading people is a crucial part of investing. You're all early-stage investors, and you more or less have your own methodologies for assessing people. Now the first and second tiers of AI LLM teams have developed their own styles.

I'd like to ask Ling Xia and Yungang — what distinctive traits do you think teams that ultimately achieve great success in AI might possess?

Ling Xia: Future Capital has invested in tech startups for the past decade. The profile of tech entrepreneurs we favor is quite distinctive and easy to summarize: focused, resilient, low-key. The founders we back all display these traits very prominently. This morning's conversation between Mr. Huang and Junjie also showed focus, resilience, and low-key temperament.

Another dimension we've consistently adhered to is our belief that what truly builds a tech company is a technologist who becomes an entrepreneur, not a scientist. We're very clear that we should invest in someone who can grow into a technology-savvy entrepreneur, not simply a scientist.

Because scientists and entrepreneurs think in fundamentally different ways. At Future Capital, we evaluate through the complete lens of an entrepreneur, not merely because someone is an outstanding scientist in a particular field.

Yungang Huang: I agree with what Mr. Xia said. Regardless of era or industry — even today's LLMs — entrepreneurs who truly understand product have two meanings behind that: they deeply understand user needs, and they understand the path that could lead to future commercialization. "Product" encompasses a great deal, with technological variables as the prerequisite; the feel for product is especially important.

Beyond Zhilin's widely recognized technical abilities, he has excellent product intuition. With that intuition, you can truly move toward commercial success. What ultimately creates distance is whether you have unique products that generate unique data — that's how you achieve final success. At the end of the day, this person must be well-rounded, understanding both technology and product.

"Dark Tides": Among the five of you here today, how many can code? Does being able to code affect how you approach AI investing?

Yusen Dai: We early-stage investors are particularly prone to "blind spots under the lamp" — when you understand a domain, and paradigm shifts occur in that domain, you're actually prone to making premature negative or critical judgments early on.

The domains we invest in are quite diverse, so any specific background isn't necessarily the core factor here. When I started my business selling cosmetics, I didn't actually need cosmetics — it was about understanding business operations and entrepreneurship itself.

"Dark Tides": Chen, as someone who can code, do you have "blind spots under the lamp"?

Yu Chen: I don't have blind spots under the lamp — it's about maintaining an open mindset. You may know how to code, but you realize that what you learned in school is completely different from what's happening now. This technology changes very rapidly; computer science essentially sees its entire knowledge structure reconstituted every four or five years.

I also agree with what Yusen said about why people prefer to invest in young people around 30 — this implicitly means their knowledge structures are relatively newer, enabling them to produce cutting-edge work. I don't have blind spots under the lamp; I approach the latest technology with an open mindset, and I'm continuously learning about it myself.

Rui Han: Investors need to grasp the big picture and let small details go — you don't need to know how to cook to invest in restaurants.

Ling Xia: I've had blind spots under the lamp in my past. During graduate school I worked on image recognition, so when I first entered VC, I missed Face++ precisely because of this blind spot. After ten years of investing, I believe whether someone can code is also quite superficial. The value of early-stage investors lies in being able to discover future opportunities earlier than others in the market.

The recognition and grasp of opportunities is not equivalent to whether you can code — curiosity and learning ability matter more. It depends on where you're willing to invest your energy, and where your understanding comes from. If you understand technology, or knowing how to code naturally brings you closer to the frontier of technological development, but that alone is far from sufficient.

Yungang Huang: Having knowledge doesn't equal having judgment — knowledge and judgment are two different things. We still need to diligently acquire knowledge, especially now that AI has helped us solve many knowledge problems. What we need more is to think problems through clearly, and AI will increasingly help us solve them. AI can now complete most programming code, so whether you can code or not isn't important.

Part 3

"Dark Tides": A sense of rhythm is particularly important in early-stage investing. It's been a year and a half since ChatGPT's release in October 2022 — several "springs and autumns" have passed in China's AI investment industry, with many players coming and going and rankings shifting. Facing such massive and volatile waves, how do you find your own rhythm?

Yusen Dai: Most of us invested during the internet and mobile internet eras, which was already investing when internet technology had reached a relatively mature stage. AI is still in a relatively early stage, so often we look to history for lessons — seeking analogies from the history of angel investing starting in Silicon Valley to VC over the years. You need patience. The "hundred-group wars," "hundred-model wars," "hundred-X wars" — people are accustomed to flocking to something new.

But today's AI or LLMs are more like chip-making back then — there's a significant research component involved, not just building on open-source code. You need patience. In the year and a half since ChatGPT's release, many new application innovations and scenarios have already emerged.

Many say AI evolution has slowed, but that's because expectations have risen — people are thinking of the internet's explosive proliferation of new applications. We need to recognize this is still relatively early-stage technology. Many investors and entrepreneurs say this is the Apple era, urging rapid large-scale action. But it's hard to say whether this is the Apple era or the BlackBerry era. Those of us born in the '80s lived through the BlackBerry era; those born in the '90s don't even know what BlackBerry did. BlackBerry-era technology had limitations — you couldn't build Douyin in the BlackBerry era, even if you knew it would be huge, the technology then couldn't support it.

A technology's development goes through stages where technology is the bottleneck — you can imagine it but can't build it — then reaches a stage where technology becomes relatively mature and enters an idea-constrained phase: if you can imagine it, you can build it; if you can't imagine it, you can't. Right now, we can imagine it but can't build it. Our technical founders need to judge technology, to predict how far technology can advance in the next year or two with existing resources. If you think too far ahead, you may not be able to execute. At this point, understanding technology — especially having frontline knowledge from a research perspective — is extremely important. Have patience. BlackBerry wasn't about a hundred flowers blooming; there were three flowers to pick, and which one you picked mattered enormously.

Regarding users and monetization — because we've all lived through the relatively mature mobile internet era — people ask about business models. Because we've all seen how the mobile internet ended, we ask what AI's endgame is. There's no answer to that now. Anyone who claims to have an answer is definitely a fraud.

I posted on Moments before: Google launched in 1998, and only by its 2004 IPO did it find a relatively core business model. Facebook launched in 2004, and only in 2012 when it went public did it introduce feed ads and find its business model.

If Google's and Facebook's two "money-printing machines" both took six and eight years respectively to find business models, then demanding that an even earlier technology — LLM applications — find clear business models and explore endgames and landscapes in just over a year is premature. What matters is penetration rate. Before technology penetration reaches a certain level, everything is in rapid expansion phase.

Thinking about monetization too early is actually a stage mismatch. We should focus on how to get more people using AI products, what products modern technology can enable, what scenarios have genuine user adoption rather than being conceptual stages with no real users — especially given that capital markets aren't particularly optimistic right now.

"Dark Tides": Learn from history, respect patterns, maintain patience.

Yusen Dai: Patient capital.

Yu Chen: In this era of information overload, it's easy to get swayed by media and peer momentum. I think the most important thing is maintaining independent thinking — don't let others set your pace. Start from first principles: consider what's useful and feasible. ChatGPT exploded in December 2022, but we began engaging with LLMs in early 2021 and decided to invest then.

Why? Because I believe the world always needs general-purpose intelligent agents. While current LLM solutions may not be the final solution, they will lead us a major step in the right direction. It's somewhat like Goldbach's conjecture — it may not solve the problem, but it can bring us close enough. Independent thinking in investing is crucial.

Rui Han: We've looked at different industries over the years; each has its own cycle and pace. I understand Mr. Liu's question as: how do we as investors respond when looking at different industries?

If the endgame is very distant, don't try to commit precise errors — fuzzy correctness is sufficient. Perhaps we can place key thermometers along the path. For example, at Gaorong, our research staff and middle/back office have all started using Kimi. We see this as AI products entering life and staying in life for the first time — this thermometer is quite important for us along the path.

If we return to the narrow definition of rhythm: the market doesn't care about your rhythm, and the market won't accommodate your rhythm. Before rhythm becomes a logically self-consistent trap, or a constraint on self-modification, what you need to do is interrupt it.

"Dark Tides": Vigilantly detect changes.

Rui Han: Don't let your rhythm become your main theme — your rhythm means nothing to this world.

Part 4

"Dark Tides": Successful VC investments are basically non-consensus in early stages. I'd like to ask: facing today's AI, what views do you hold that differ from the market?

Yusen Dai: Right now I don't know what's consensus and what's non-consensus.

"Dark Tides: Haha, I knew you'd say that. Then just share your own views.**

Yusen Dai: Don't discuss AI monetization. Also, general-purpose humanoid robots are still too early — so we haven't invested in a single one. This is very worth studying, but it's genuinely too early, definitely not at the GPT moment.

"Dark Tides": Do you think it's so early that even VCs shouldn't invest?

Yusen Dai: Depends on whether you're extraordinarily patient. Or whether you're coming from a research or investment perspective.

Yu Chen: At this stage, energy is more important than so-called compute. Many major US tech companies have halted data center investments — not because they lack money to buy chips, but because the US no longer has sufficient energy supply. Even if you started building nuclear power plants now — and you've seen the news, Microsoft partnering with Westinghouse to build nuclear plants — this isn't something that happens overnight. Solving the energy problem is extremely important, even more important than the current compute problem.

Rui Han: Building on Yu Chen's point — I was joking with colleagues the other day: there are so many time-travel dramas now. Suppose a modern person with abundant knowledge traveled back to ancient times — what ultimately limits your ability to change the world? Materials science.

To circle back to Mr. Liu's question — non-consensus is partly about perspective, but there's another easily overlooked dimension: degree. Today, I believe degree of conviction can create enormous distance between people on AGI. "Belief" and "belief" can be worlds apart. Is your belief conditional or unconditional? And if conditional, what are the conditions? Lip service, genuine conviction, or putting your money where your mouth is?

If we use climbing Everest as an analogy, where exactly are we? Have we just bought the gear? Or only half the right gear? Have we reached Tibet? Base camp? Or are we actually attempting the summit? Of course, the more optimistic view is that this Everest has no peak — you can keep climbing indefinitely. I think today, if we debate granular points of non-consensus, things change too fast. But degree of conviction — that can create real distance between practitioners and investors.

"Waves": What stage of conviction are you at?

Rui Han: Like Moonshot AI's English name — Moonshot. It's a moon landing.

Ling Xia: Let me respond to the previous two speakers. In 2022, we invested in LimX Dynamics, a general-purpose robotics company. Our core thesis was that the end-to-end, data-driven paradigm could migrate to this domain. If we treat humanoid robots as our desired endgame — analogous to Level 4 autonomous driving — then however long L4 autonomy took from conception to deployment, that's how long general-purpose robots will take to enter millions of households.

The "ChatGPT Moment" for humanoid robots achieving generalizable intelligent manipulation is at least 2-3 years away. But are there companies, operating under the end-to-end data-driven paradigm, that can find their equivalent of L2++ autonomous driving — achieving their own data and commercial flywheel first, while their underlying tech stack continues evolving toward their ultimate L4 goal? Tesla and Li Auto have done exactly this over the past few years. I believe such companies exist in the market today, and the timing is right.

Energy means something different for China versus the US. The US grid and utilities are extremely fragmented — power infrastructure simply cannot keep pace with compute cluster deployment. That's their problem. In China, this isn't an issue. We've run the numbers: today's AI challenge isn't absolute compute capacity, but energy density within constrained physical space — not absolute energy supply. China's infrastructure is exceptionally strong, with ultra-high-voltage transmission. This isn't our core bottleneck.

To respond to Mr. Liu's question — I think there's an important trend closely related to AI that hasn't been deeply discussed domestically. Because of this generation of AI, the very essence of "chips" has fundamentally changed. In the PC and digital eras, chips meant large-scale integrated circuits. In the smartphone era, chips meant a single SoC.

But in today's AI era, the chip is a system. Chip design should be design for the system. Moore's Law doubles SoC performance every 18 months, but at the system level, improvement happens by an order of magnitude every two years. This is rarely discussed in China today.

Yungang Huang: We've looked at a lot of humanoid robots and invested in some. The world and humanity really need them, because what AI can do right now is extremely limited. We need the physical world to fold our blankets and do our laundry. We need physical companionship, not AI virtual companionship. It's a question of how far away.

Why do robots need two legs? Why legged rather than wheeled? Wheeled is faster. The most intelligent, most general-purpose agents are sitting right here — us humans. We're the most general, the most powerful. If you need to go fast, take a car — robots can take cars too. If you need to move in a small space, we have stairs, where wheeled designs are at a disadvantage. Why did Elon Musk build humanoid robots? The person who thinks most in first principles — how did he choose? Very simple.

Part 5

"Waves": An investor shared this question with me: Alibaba's Qwen as an open-source model is competitive with, or even better than, what startups have built. In this race, where exactly can startups compete with or find breathing room against big tech?

Yu Chen: You have to compete on differentiation. Right now there are clearly several business models or product forms for large models. Take APIs — Alibaba Cloud, Volcano Engine have advantages there. They can offer APIs at cost or below, making money from cloud computing instead. But for startups, competing on cost and pricing with big tech is obviously hard.

Or take productivity tools, whether chat or search — if ByteDance can't charge for it, you can't either. You can't burn compute, pay a lot of money, and fail to make it back. Large model startups need to find product and business model differentiation from big tech. Differentiated competition is the crucial direction for startup survival.

Rui Han: It's about the people. When we use a company name or a big tech name to represent model capability, I don't think that's fully accurate. We should perhaps pay more attention to the granularity of that small core team — specifically, who built it.

Today's Chinese large model companies are all still in catch-up mode — big tech and startups alike. From our perspective, this process is more like a leapfrog race, where new entrants can jump ahead. Snap a photo at any given second and someone is in first place, but a few seconds later the leader may have changed.

If the answer to a question changes every month in real life, then that answer probably has no meaning over a three-to-five-year horizon or longer. What matters more is the leaping — who can consistently stay in the first tier.

We've observed, not just in AI but in other industries, that the path to breaking through requires entrepreneurs to first do something unexpected before they can do something conventional. Without the unexpected, there's no chance to be conventional. Find a point where you can do something unexpected and go all in. Big tech is also run by specific people, also flesh and blood — nothing is unbreakable.

Part 6

"Waves": Returning to Moonshot AI and MiniMax — both have reached valuations of $2.5 billion or higher. Are you worried the market is overheating and prematurely consuming these companies' valuation growth room? Can you predict where the ceiling might be for this wave of Chinese large language model companies?

Yusen Dai: If they succeed, they're all cheap. If they fail, $300 million or $3 billion is all expensive. With such massive uncertainty, valuation is hard. Capital markets are bipolar — $3 billion used to be nothing, many companies were like that, even coffee shops. Valuation is relative. Whether you can ultimately turn technology into product and deliver it is what matters.

I remember the early mobile internet era — Instagram sold for $1 billion, 13 people. Everyone said the company was so early, so small, no revenue, and sold for so much. Turns out it was worth hundreds of billions.

In the early stage of a technology revolution, if you can genuinely be among the most leading players, whoever you are, I don't think tens of billions of dollars is particularly high. Right now, does China's capital market have a bubble? There isn't even any foaming agent left — who knows where the bubble is.

Ling Xia: Today in China, whether MiniMax or Moonshot AI, their valuations are to some degree undervalued, not overvalued. If you compare with European and American peers, given these two companies' technical level, their valuations should absolutely be at least 2-3x higher.

Conversely, if America's challenge is energy, then Chinese AI's biggest challenge today is capital. If we want to build a 100,000 H100 compute cluster, you need $4-5 billion in capex. Even leasing, you're paying over $1 billion in rent annually. Working backward from that to required annual fundraising and then to valuation — that's challenging.

Long-term, I still lean extremely optimistic. Asked about the ceiling — in 2014, someone asked Mingming Huang, after you invested in Li Auto, what kind of company will it become? He said floor in the tens of billions, ceiling in the hundreds of billions. That was heavily challenged at the time.

For China's top AI companies, as Junjie said this morning, if among the world's top five AI companies the second place is Chinese, then I believe over a ten-year horizon this should be a trillion-dollar company. The first wave of Chinese internet companies reached tens of billions. The second wave, like ByteDance and Alibaba, basically reached hundreds of billions. Regrettably for various reasons, ByteDance should logically be China's first trillion-dollar company today.

If we look ten, fifteen years out — if America today has the Magnificent 7, I think China will certainly have 6-7 trillion-dollar market cap companies by then, with no fewer than two being AI companies. That would mean something for China.

I think AI is more like a productivity revolution — the best analogy is still the invention of the computer. In 1967, IBM's market cap was $150 billion, when US GDP was over $800 billion. IBM alone, as the leading computer company, was one-quarter of US GDP. If China's GDP is 180 trillion RMB or even 200 trillion RMB in ten years, how big is one-quarter of that?

Why was it worth so much? Because 20-30% of US GDP at the time was generated by the computer industry. If AI can drive 20-30% incremental GDP for China in the future, such enterprises will absolutely be worth it.

Yungang Huang: Large model investment is not very VC-friendly — the capital requirements are enormous. In fact, neither startups nor investors want such high valuations. It's because they need to spend money that valuations get pushed up. Actually, each round's valuation increase isn't that much — they just need the money. I believe something massive will definitely emerge, hundreds of billions or trillions of dollars.

The most important thing for China is: don't spread investments too thin. Unite and support the best two or three companies — that's the most effective approach. I went to Silicon Valley last year, and every time I'm struck by how American large model companies or teams all want to build AGI for all humanity, to be leaders.

Every Chinese team, regardless of capability or position, says "I want to build my AGI." If we could unite and support the best team together, that would be better — rather than wasting money.

"Waves": To borrow Wang Chuanfu's words: together, that is China's AGI.

Yu Chen: Everyone answered really well. Being able to sit here today means absolute conviction in AGI. Since we have absolute conviction, the number itself doesn't matter.

Rui Han: Though at this stage I believe the qualitative outweighs the quantitative, I sincerely hope all the numbers everyone mentioned come true.

"Waves": Thank you, everyone.