Xiaohu Zhu: No Bubble in Sight for Three Years — When Consensus Is Too Concentrated, Offset by Fifteen Degrees

The show opens cold. Zhang Xiaojun asks whether, standing at the tail end of 2025, Xiaohu Zhu (朱啸虎) can smell an AI bubble. Zhu says he doesn't see one for at least three years: when everyone is talking about a bubble, the bubble definitely hasn't arrived.

Where does the gloom about an AI bubble come from, and how does it spread? Zhu thinks a lot of public-market investors are steering it on purpose, hoping for a pullback. The pullback of the past month or two has actually laid the ground for another explosion next year.

NVIDIA's pullback, he thinks, is healthier. Still holding it? He says there should be better names than NVIDIA. As for Sam Altman today: he says one thing and does another. He may talk AGI; in practice it's all on the application side.

The formal open: Zhang Xiaojun asks Zhu to say hello to the audience. She says this is their third conversation in nearly two years, and the third installment of Xiaohu Zhu's China realist AI story — she wants a running record of his notes on this wave. From the start of the year to now, has the call changed at the root? Anything that surprised him?

Zhu says everyone has gotten more and more realistic. Look at Sam Altman: this year he barely mentions AGI. What ChatGPT has shipped is more about turning toward applications — Pulse, a browser, even the group chat that just landed. A company chasing AGI building a group-chat tool would have been unthinkable before.

OpenAI's shift this year, he thinks, is excellent: the strategy is clear, and you can see how Sam, a founder who came out of investing, is positioning the company. Search and ChatGPT conversation are both weekly-active scenarios, hard to hold against a Big Tech assault; daily usage has a much thicker defense. Products like Pulse — a Toutiao-like feed — plus the browser and group chat are all about refocusing frequency from weekly to daily. That's a major strategic turn.

Will ChatGPT become the new super entry? Yes. The reason it is focusing on daily usage is that it wants to be a super-app entry. Once group chat is done well, does it then need to build a new social graph and rebuild social relationships? It really could hit META hard. At the very least it's a new super entry.

An IPO coming: he thinks it may be a bit early. The scale is already huge, but there's no necessity today — primary fundraising is still very easy, and the $300 billion round was far oversubscribed. They don't have to list as fast as possible. Starting the prep may be because both the growth rate and the scale are already large enough.

On the business model, overseas users have a strong willingness to pay and very good stickiness: unless you're clearly better or very different, they won't switch. That's what's especially good about overseas consumer. A ChatGPT-like endgame may not be a monopoly; once you cut to daily usage there will be at most two or three, like WeChat, WhatsApp, Instagram, and Messenger among chat apps.

Who is China's OpenAI? He feels Doubao is ahead. Alibaba only recently started pushing on consumer, a bit late, but Qwen and Ant's Lingguang both have bright spots. He's never thought Google was behind — the reserves run deep, and the models have always been strong. Chatbots replacing search is already obvious in China; a lot of users just search with a chatbot. Google has also been determined to disrupt itself, and search is embracing AI too.

Google and OpenAI are of course competitors. The key is who cuts toward daily usage more decisively. ChatGPT has at least one sharp move: group chat. After group chat you build social relationships — that's hitting a place Google isn't great at, and hitting META as well, which actually lets ChatGPT dig its own moat. Competition overseas is heating up further, everyone inside everyone else, and in the end it's still a fight over daily usage.

Domestically it's only just begun. Alibaba has just started pushing on consumer; Tencent hasn't pushed yet. For the past twenty years Tencent has never been the one burning money to trial-and-error out in front — it always waits until everyone has fought it out, then chases from fifth or sixth up to first. Zhang Xiaojun mentions he said on Xiaohongshu that Qwen can go toe-to-toe with OpenAI. Zhu says Alibaba has the model capability, and the consumer app isn't bad — it's just a bit late, and still worth watching.

Among ByteDance, Tencent, and Alibaba, in the new fight for the entry he currently likes ByteDance more: Doubao's experience is very good, and the traffic advantage is obvious. Alibaba needs to refocus — on the latest earnings, instant retail is burning money on both ends, and the pressure will be heavy. Whether instant retail should refocus onto consumer AI is a strategic test for Alibaba. The war with Meituan, as it stands, looks like mutual destruction; the spend may not mean much. Relatively speaking, an AI super entry matters more to Qwen or Lingguang, and that competition has only just started.

Coming into 2025, a new round of the entry fight has started at home and abroad — it's no longer a foundation-model fight. People have stopped talking AGI, because everyone can feel how hard it is. Turning to applications, it has to be a fight over the super entry, over daily usage, over time spent. The rhythm is like fighting for WeChat in mobile internet: a daily-active, high-duration super entry that can accumulate user data. Add social relationships on top and it's extremely powerful. ChatGPT group chat has only just shown a hint of possibility: it claps and likes underneath on its own — a kind of personification that didn't exist before.

Will the new AI entry replace traditional social networks? It's still very early today; you can't see nearly far enough. In the mobile-internet era a lot of social products later added group chat, and daily usage and time spent jumped. ChatGPT's cut is different: it starts with group chat. Whether that cuts the other way into social networks, you can't tell today, but it's at least a possibility. For OpenAI to become a trillion-dollar company, it definitely has to cut into social networks.

Zhang Xiaojun turns to Ilya Sutskever's comments: the past few years Scaling was the main theme; now they're entering the age of research, pre-training data is about to run out, and you can't just blindly scale. Does that mean the pace of AI progress ahead isn't optimistic? Zhu says Yann LeCun has been saying this all along. When everyone was charging forward single-mindedly, they thought he was old and throwing cold water on the young. Ilya says it today, and people start to think it might be right.

The past year has already made it obvious: going further is marginal improvement, high cost, limited gain. At least under the Transformer architecture, it may have basically topped out. But is the current technical level enough to carry a new wave? Application competition has already started exploding. Industry-wide token consumption jumped more than tenfold this year, which shows this wave of technology is already good enough in many scenarios.

The bubble smell, asked again. Zhu says Alibaba's earnings put it well: he doesn't see a bubble for at least three years, and when everyone is talking about it, it definitely hasn't arrived. Those arguments don't understand the industry — for example that GPUs fully depreciate in two or three years. How could that be? Cards are nowhere near enough right now; five-, six-, seven-, eight-year-old cards are all still in use, and if you have a card you can use it. Token consumption exploded too fast — more than tenfold this year, and it can be another tenfold next year. Data-center buildout is nowhere close to keeping up. Power, especially in the US, is the bottleneck. China putting up a dozen nuclear plants at once and a lot of offshore solar, he thinks, is the right call: the AI contest later is a data-center contest, a power contest, and China has a big opening.

How does the gloom spread? He still thinks public-market investors are steering it on purpose, the goal being a pullback so they can lay the ground for another explosion next year. Look at token consumption and you know: China is even faster than the US. Very small companies burn tens of billions of tokens a day. OpenAI says a trillion tokens a month counts as a big customer; in China, companies doing tens of billions a day — close to or over a trillion a month — are common. A million DAU can burn tens of billions of tokens, which a lot of startups can reach, and tokens per DAU per day are still rising. The bottleneck later really is data centers, really is power.

Who carries the cost of this wave? Domestic foundation models price tokens very low, so even tens of billions of tokens a day doesn't cost much — most of it is input tokens, which are cheap to begin with, and output volume is relatively small. If a giant thinks the scenario has strategic value, it can discount further, almost for free. That's why startups are willing to charge ahead. Right now token cost is about 0.2 yuan RMB per DAU per day; ads and simple monetization can earn back 0.1 yuan, so the subsidy is actually manageable.

If a bubble really shows up later, what would the tell be? Zhu says: when nobody is talking about a bubble, that may actually be the bubble. Like consumer last year — when nobody said they were investing in consumer, that was the bottom. In a real bubble nobody says bubble; every last cent piles in. Another tell is dark fiber: more than ninety percent of the fiber laid out sits unused. Today's AI cluster — chips and data centers — is nowhere near enough.

Zhang Xiaojun quotes him: the 2000 internet bubble was price and volume coming apart; 2025 AI has no bubble because price and volume move as one. Zhu grants that revenue has ramped fast, but the core difference is that US AI valuations are a hundred times China's. One end has to be wrong: either that side's revenue isn't sustainable, or China's revenue is undervalued. How this divergence evolves later, you still can't see.

The so-called GPU flywheel — OpenAI and Oracle trading with each other, inflating valuations — does look a bit like the internet bubble back then, but today you can understand it: depreciation needs to go off balance sheet. It's a listed-company accounting trick; it doesn't look good on your own books. The core question isn't on- or off-balance-sheet; it's whether anyone is actually using it, and whether there's enough. That doesn't dent his optimism: cards aren't enough today, data centers aren't enough — that's the most core point.

NVIDIA's pullback is healthier, laying a better base for next year's explosion. He still thinks there are better names than NVIDIA. OpenAI at this moment? The $300 billion round was oversubscribed; a lot of friends didn't get in. The big test ahead is whether the consumer super entry can hold. At least Sam has shown that the strategic choices and product definitions are set well. Next is whether a social network grows after group chat.

How does he see Sam as a person today? He says one thing and does another — he may talk API; in practice it's all on the application side. Zhang Xiaojun says that fits his view fairly well. Zhu thinks Sam is also a very pragmatic person — a company built by an investor has to answer to investors and to employees. That's fine.

If he had $500 million now to put into those Silicon Valley AI giants, how would he place it? Conservatively, still Google and Microsoft. For a steady book, Google, Microsoft, and Apple are relatively more stable. OpenAI wouldn't be his pick: whether it can hold a consumer super entry is a huge test, and the opponents it faces are different. The strategic choices are very well placed; execution will keep being tested. Whether social relationships can be built after group chat is a big gate. You can't call OpenAI a bubble today.

The path to profit splits in two. One is whether they keep training foundation models — that burns a lot of money. The other is inference: costs are controllable, revenue relative to any bubble isn't huge, and profitability may not be that far off. Anthropic? He doesn't much like the API business — API users have zero loyalty. Whichever model is cheaper and better, they switch all of it at once. xAI and Thinking Machines Lab also wouldn't be his pick; Microsoft and Google are still more solid.

Today every AI company talks about huge future revenue. Where does the money come from? Not out of the traditional giants' pockets, but from replacing white-collar jobs: this role was going to hire someone, and now it doesn't. By Office user counts there are at least a billion white-collar workers worldwide. Even at $40,000 cost per person, replacing one or two percent is enormous revenue. You also have to look at gross margin: a lot of products like Cursor are negative-margin, which is subsidizing programmers. Revenue with a positive margin is what comes from replacing white-collar roles. That will bring huge social problems, very complicated ones.

How would he place the domestic giants? The same. ByteDance is a good name, Alibaba is, Tencent is too. At the start of the year people said meeting DeepSeek was him meeting his AI true love — how does he see it now? He still thinks people underestimate how it changed humanity, how it changed history. Ten years from now, looking back, DeepSeek will be a very important turning point in the AI process: otherwise China wouldn't have been this committed to open source. Without it, humanity's AI would likely have been controlled by a few private companies, which, honestly, is dangerous. With DeepSeek, it can help Chinese model companies lock in a long-term edge.

Has it expanded into an Android-like ecosystem the way he originally expected? Zhu thinks DeepSeek hasn't considered commercialization at all, hasn't even considered the ecosystem — it's single-mindedly looking ahead at whether there's AGI and where the moat is. He's very eager to see how far R2 will go. But both China and the US have shown that going further toward AGI is extremely hard. Today every open-source ecosystem is still based on DeepSeek's architecture, and a few newly released domestic open-source models probably are too. That is still very valuable.

API isn't a good business; later it will become infrastructure, like water and electricity. So how do you commercialize? Commercialization is too far away — they haven't even thought about the ecosystem. Zhang Xiaojun says AGI has peaked now. Zhu says at Chinese New Year this year people still didn't think so; by now both China and the US feel going further is hard — at least language models themselves have peaked, and multimodal may still have a couple of steps left. Advice for DeepSeek? They don't need advice; they have their own dream. He didn't see Liang Wenfeng later either.

Why have most domestic models gone open source, with only ByteDance choosing closed source? He thinks it's a commercialization choice: ByteDance's own ecosystem is too rich, with all kinds of products and entries, so not opening the source doesn't matter. The Six Little Dragons? Last year he said wait another year and see who's still standing. Today they're all still around, but they've started to diverge. Looking at it now, the ability to train a foundation model is still scarce. Even continuing to train on DeepSeek's open-source architecture, the diffusion has been slower than he imagined. The hardest part domestically is commercialization: the giants give out tokens almost free or at very low prices, so startups can barely make it work.

He can't see the Six Little Dragons' endgame. The giants' tokens are too cheap — how do startups commercialize? Zhipu and MiniMax are racing to list. Listing or not listing doesn't mean much; the key is how they eventually make money. Volcano Engine and Alibaba's token pricing is really fierce. The enterprises they see are basically either Volcano or Alibaba — cheap prices, and the models are good enough. Startups have no niche in foundation models: the model itself is part of cloud services. What do you compete on?

Everyone calls them old-school investors. The upside is a slightly longer memory window. What's the same in this wave as before, and what can't you reference? Zhu says it's too similar: from the chips at the bottom, to infra, to the application layer, every cycle has the exact same rhythm. What's different is the speed is faster — maybe 6x PC internet, 3x mobile internet. At the tail end of 2025, mass adoption has indeed started to accelerate.

One-sentence recap of 2025, one-sentence call for 2026? This year people have already stopped talking the AGI dream and are all focused on the real application layer, which is a very good thing. Next year is when the application layer truly explodes further. After tokens explode again, whether the underlying infrastructure can keep up is a very important test — mainly a test for the US; China is relatively easier to solve. Second, China's open-source ecosystem is worth watching: if Gemini or ChatGPT stall going forward, another burst from Chinese open source still has a lot of potential. A lot of startups take Qwen's small open-source models, do private deployments, get good results, low cost, and revenue ramps fast.

Zhang Xiaojun sums it up: he's not optimistic about technical progress, but he is optimistic about AI continuing to explode. Zhu says he's very optimistic. In a lot of scenarios, token consumption has ramped faster than expected. A tiny company burning tens of billions of tokens a day — he was stunned when he heard it. A million DAU can burn tens of billions of tokens in a day, and a million DAU is nothing in mobile internet. Consumption is at least an AI-content indicator, though there's water in it too — seventy or eighty percent is input tokens, a bit lower in technical content.

What new companies did he back this year? More than a dozen, extremely busy, mainly AI applications, plus some workhorse robots that actually do the job. Humanoid robots have no scenario; they barely look at them. All-in on humanoids? The talk is meaningless, and they're all expensive. Tracks where consensus is too concentrated aren't interesting to invest in. They're now offsetting the highly concentrated consensus tracks, shifting about fifteen degrees: robots that do work, and the intelligence is pretty good too.

What kind of work? They backed a robot that cleans ships at sea, and it's doing very well; and a massage robot that not only massages but talks to the customer — "Sis, how's my massage? Want to top up?" — and can even close the sale. How much did they put in? All cheap, very high value for money. Offset consensus by fifteen degrees and the value for money opens up at once. The founders are all very grassroots, especially interested in the track, having studied it for years. How long did they talk? About half an hour. Half an hour and they invested.

Besides robots they also backed AI companions, also decided in a little over ten minutes, under twenty, this year. It's grown far beyond expectations; in less than six months it's already a clear leader in the track — shipments, activity, token burn, all strong. The moat is still taking user experience to the extreme: a toy this small, even getting Wi-Fi connected properly is hard. Will it be a very fragmented market? He thinks it's like game companies: some games are especially good, others exist too, and the playbooks are different.

Of the dozen-plus, more are software-led, some software-hardware together. Bambu Lab is doing very well. Why he's especially optimistic this year: in every niche track Chinese founders are far ahead, and there are too many openings. Plaud is also very good — a lot of moves this year, but US market share is very stable. Products like this are basically Chinese founders competing with each other; foreigners don't even need to think about it. Your competitors are all within twenty kilometers. This is a hardware-manufacturing startup wave with Shenzhen as the emblem. He goes to Shenzhen, Hangzhou, Beijing, Shanghai to look at deals, quite a few, and looping around once a week is convenient too.

How big is China's supply-chain advantage? Enormous. He cites a men's athletic-wear company they backed in Vietnam. The T-shirts sell very well; they asked it to extend into electric shavers — simple, and it still couldn't make them, and had to come to the Greater Bay Area. A systemic supply-chain advantage is immense: every single part looks simple, and once you have a hundred components it gets formidable, and anyone else trying will be exhausted. In wearables they backed a company that provides solutions for third parties. This track is extremely crowded: glasses, earbuds, rings, watches will all definitely get AI'd, first-, second-, and third-tier companies are all doing it, and competition will be fierce. There are other deals he wants to keep quiet for now and talk about next year.

Zhang Xiaojun asks about Chinese agent products like Manus and Liblib. Zhu says on the application side, on user experience, Chinese founders are definitely more competitive. Why didn't they back these? Liblib is one of theirs; with Manus they've always had some reservations about pure tools. In the mobile-internet era, tools started with huge user growth and might eventually fade away — the giants or the model companies or the OS didn't have time to nail it then, and once they have spare capacity they will inevitably take it over.

Is Liblib's niche stable? At least it has a designer ecosystem: almost every image designer in China is on it, and it's trying to build its own ecosystem. That entry point has a bit of a moat, relatively speaking. What kind of agent sits outside the giants' narrative, that he'd rather back? You have to go vertical, maybe three blocks away from Big Tech. In mobile internet people said one block away; in the AI era it may take three. Or do private deployments: take Qwen's 30B small model and build a vertical application. The giants don't want to do it, and there's opportunity in all of that.

They backed several good vertical scenarios this year — all kinds of vertical engines and optimized deployments, growing fast: tens of millions last year, over a hundred million this year, two or three hundred million next year. How do you tell you're three blocks from Big Tech? The giants don't want this kind of vertical: you have to do vertical optimization, you have to do deployment, and you still have to go sell. Giants will generally look for startups to help with the last mile. Agent startups this year look a lot like investing in apps back then, so you learn from the last cycle: which apps had huge growth at first and later disappeared.

Too many. The simplest is the flashlight: Android and iOS didn't have one at the start, and a flashlight app could get a lot of users. Then perpetual calendars, browsers, AI-native email. They all raised plenty of money back then — where are those companies today? He adds, "Sorry, Fu Sheng" — Android Optimization Master had hundreds of millions of users worldwide, and Fu Sheng's pivot actually went pretty well. It wasn't a fake need, it was a real need, but you can't hold it; you're just a tool. The OS or the underlying model doesn't have time at first, which gives startups some openings; once they have spare capacity, nailing these is easy. Apps that can actually get that much DAU are only a handful.

So don't build a pure tool. That's the lesson he took from the last cycle. They once backed a company that wanted to be China's OS; luckily it later sold to a giant, so at least they didn't lose money, even made a little. The ones that survived the last cycle and survived well: in the US it's clear — Amazon, Facebook, Google, still those. Among pure mobile-internet companies, Uber, DoorDash, and Airbnb all have half of it as offline grind the giants don't want. Same in China: Didi and Meituan both have half grind; Douyin, Kuaishou, and Xiaohongshu exist because the giants kept making strategic mistakes and left a niche.

Will any giant keep making mistakes in this AI wave? He can't see it. Can't see it in China or the US — he wouldn't dare bet on it. Not just the first-tier giants; even second- and third-tier companies are all-in. META looks a bit slower, and it did make some mistakes; ChatGPT found those strategic positions and filled them. For startups it's really not easy. Does he see a gap like that in China? He can't see one at all. So you back the hard, dirty vertical scenarios the giants don't want — three blocks away from Big Tech.

Will there be a general-purpose agent? There definitely will, but who provides that service? Foundation models definitely will, WeChat definitely will, Douyin will too. ChatGPT cutting from weekly to daily this decisively is because it has to cut to daily: a scenario that isn't daily-active, you simply can't hold. The boundary between foundation-model companies and agent startups: there will definitely be a lot of vertical agents in the industry, all different scenarios. The US is very clear — Abridge does the consult between doctor and patient, a vertical engine, and the core is pushing accuracy to the extreme, above ninety-five percent.

Will multimodal change the ecosystem? No — the giants will still be the giants. How many startups will Nano Banana kill? A lot. One launch event wipes out a pile, including after Google Gemini came out you no longer needed RAG, and that pile of startups was gone. Startups' niches are even harder to find now, especially if you want a hundred-billion-dollar one — almost none. Finding a several-billion or ten-billion-dollar opening in the cracks, maybe there's still some chance. There's an Ele.me-sized possibility, not necessarily a Douyin — though on Douyin's first day nobody knew it would be this big either.

So it has to be an opening nobody looks up to on day one, or it probably won't work. Grow slowly, stay under the radar for three years. For three years other people don't get it, all think it's a small opening, and then you get a chance to explode. Investing in agents looks a lot like investing in apps. The difference is that one block was enough back then — fewer giants, domestically just BAT. Now giants everywhere, big countries and small countries everywhere, finding a crack is really too hard. In the AI era the giants getting bigger is inevitable: mobile internet already made the giants more than 10x bigger than PC internet. In the AI era a ten-trillion-dollar company will very surely appear, even within three years.

The key is whether a hundred-billion-dollar opening can come out of startups. In the US there may already be many at a hundred billion, and a trillion may come out too; whether a hundred billion can come out domestically is still a question mark. For application-layer AI agents, what North Star metrics does he watch? One is growth rate, one is retention — the same core logic as the app era. This year's fastest revenue growers are all up more than tenfold, though the scale is all small; a bit larger and you might 3x in a year. Overseas isn't necessarily faster than China. Domestically a lot of small-scale names are growing fast too. B2B applications at tens of millions last year, over a hundred million this year, two or three hundred million next year — that still surprises them.

If the openings get smaller in the AI era, will investors get less excited? No. When they were investing in mobile internet they didn't know it was a big opening either: when they backed Ele.me, Didi, Xiaohongshu, nobody was bullish — even giving away half the allocation, nobody wanted it. Today it's the same: let the startup stay under the radar for about three years before it has a chance to get big, and of course you also need the giants to keep making mistakes. Today that possibility has dropped by an order of magnitude: not only the giants, a lot of smaller shops will fill the gap too. Startups have too few openings.

Has he spent the past three years avoiding every consensus track? Not avoiding it entirely — offsetting by fifteen degrees. If you avoid it completely, basically nobody pays attention; offset by fifteen degrees and you find a lot of high-value-for-money deals. That degree is hard to judge. He didn't back foundation models, didn't back humanoid robots — a lot of Chinese and US investors are backing humanoids with their eyes closed. He doesn't get it, so he stayed out. Founders have it this hard today — where does the excitement still come from? Once you offset by fifteen degrees, you find a lot of interesting openings. They may not be that sexy, but they're solid.

The most exciting deal of the past year? Honestly, quite a few. Cleaning ships at sea is one: once you really understand it, the market is pretty big, and the moat is high. Seawater is extremely corrosive, the waves are huge, and China's seas are all turbid. Being able to see the hull clearly in that water and strip off all the fouling — the bar is very, very high.

A BP lands — how does he decide whether to take the meeting? Whether it's interesting. A lot of the time you still have to sit down and talk before you can feel the potential of the project and the person. If there's something there, he's willing to meet. Everyone knows his ten minutes: invest or not, it's ten minutes. If he's not interested it ends in ten, and nobody feels too bad. He can meet a lot of people in a week, and you still have to meet more and talk more to get a feel. If the talk hits thirty minutes, that's already long, he's already quite interested, and he's basically going to invest. The key is you have to make clear where the market is, where the moat is, and why you can win.

The founders he's backing now are more in their thirties, more post-90s, because you still need some industry experience, at least five or six years in. Ten years ago consumer was relatively younger; B2B without industry experience is barely possible. This wave is a lot of B2B, so it leans more experienced; ten years ago mobile internet was mostly consumer, more youthful. What kind of founder should take VC money? A lot of small-and-beautiful companies have no reason to take VC, especially RMB funds, which may need to redeem in five years.

How big a story should raise? The Hong Kong listing bar is fairly clear: a consumer company needs at least $100 million in profit for a Hong Kong IPO to make much sense; B2B software SaaS or an AI application needs at least RMB 1 billion in revenue. You have to measure whether this scenario, this industry, can hit those two bars. If it can't — especially a consumer company at one or two hundred million in profit — raising has no point. Early-stage ratchets are relatively rare; later-stage they're still fairly common.

GSR Ventures' twentieth year — any postmortem? They recap every six months, and they have to reflect often. The past two or three years the pace was relatively slow; this year it clearly sped up, a lot like mobile internet back then. When a new cycle has just started, the application side isn't ready, the openings haven't shown, and the pace is slow. Once the application side starts exploding, the pace starts to accelerate. So it still looks a lot like the past few cycles.

Any change in daily life? Much busier this year: one, more deals invested; two, a lot of follow-on financing for those deals. A lot of investors ask how they see it, whether to add. Newly invested deals also need frequent contact with the team, to see whether the big direction needs a turn. Zhang Xiaojun says have applications exploded — these deals still look small. Zhu says for the public a lot of it is B2B, they don't notice, but looking at North Star metrics — growth rate, retention, repurchase — they're all very good. From an early-stage VC angle it's still very exciting.

But you don't start by telling a huge story. If someone really pitches a huge story he actually doesn't dare invest. He's different from other people: if you start by telling him a huge story, he actually doesn't dare invest. Now you have to find a crack among large, mid, and small companies, and it's not easy. The boldest deal of the year? AI toys might count — a relatively Red Ocean track. Zhang Xiaojun says this deal doesn't look like one he'd back, because it's more emotional value. Zhu says he's backed quite a bit of emotional value; he did in the mobile-internet era too.

What did it feel like meeting this founder? The other side had actually thought of every question he asked — not only thought of them, but implemented them all in the app. How do you stand out in a Red Ocean? The core is still taking user experience to the extreme. It looks like a very simple product; getting the experience right is really not easy. Quite a few Hong Kong IPOs lately, and he thinks that's a very important reason the market is active this year: the exit channel is clear, so investors dare to invest.

Some funds have started doing primary and secondary at the same time — would they make that choice? They're considering it, but the most comfortable place is still definitely the primary market. Primary is high risk, but the return is high too; the main thing is the risk is controllable and the volatility is small. Secondary, you have to benchmark the sector index every day, mark yourself against other people. The pressure is completely different; you need a completely different angle and mindset.

Is it possible this AI cycle is very unfriendly to startups, that in the end the giants just get bigger and founders' niches get badly squeezed? Zhu thinks that's quite likely. So for early-stage funds, you still have to control size and control entry cost. They've always kept the same scale, basically unchanged. The first check is around twenty to thirty million, and if it goes well they can keep adding. About the same as in the internet days.

Any industries hiding in plain sight, already changed a lot by AI but barely discussed? Plenty. People don't want to say it, don't want others to notice. Today a lot of Chinese startups burn tens of billions of tokens a day, all in places nobody is watching. Look at the volume and you know: why can they burn that many tokens? The scenario must have value, there must be users. An example: AI companion toys. They look unremarkable, and token consumption is huge. Introverts and extroverts really are different; some people really like that kind of companion. Definitely introverts like it more. A lot of companies aren't big yet, but the token burn surprised him: you don't need that much data to get the effect, so the data-flywheel moat is actually small.

Has his taste in people changed? About the same, hard to change. The core is still having thought the thing through. What kind of founder does he meet once and never meet again? The ones who aren't quite honest. You can tell they're catering to you, changing the way they tell it and the story based on what they think you like. You can definitely see that.

US funds believe more in concentrated positions, swinging hard — has he shifted that way? VC making money definitely depends on having enough ownership in the companies that hit. Last week at VC Week in Hong Kong, a lot of LPs complained that GP consensus this cycle is too concentrated: every deal is a cap deal, a dozen funds in together, each with a tiny slice — how do you make money? GPs can't make a big score, and LPs are unhappy too. And backing different VCs feels about the same; they all run about the same, so for LPs there's no diversification either. Every shop's share is tiny. A very obvious problem this cycle is that consensus is too concentrated.

Why did it turn out this way? Taste got too similar. So they want some differentiation, offset about fifteen degrees, not too far. Why don't other people offset? Consensus is the easiest choice. And there aren't really rounds anymore — everyone piles into the earliest round, A+++, Pre-A+++, endless pluses after that. Has he talked this through with peers? That's a long conversation.

Some people say VC makes money on consensus, because you need someone to keep taking the next round. Zhu says that's another topic. In the internet and mobile-internet eras, if there was consensus it eventually proved it could commercialize. The "attention economy" people laughed at later proved that if you have attention you can commercialize; shared bikes and shared power banks were laughed at too, and now they've proven they can commercialize. Consensus and commercialization may have been a bit disconnected then, but they eventually caught up. Today some of the consensus really is a long way from commercialization, and commercialization may have a hard time catching up. That's a big risk. If it never materializes, never cashes in, the consensus is a bubble.

What about the people already in? He doesn't know. So they keep asking: where's your commercialization. Today's foundation-model companies look like which companies from the last cycle? He thinks a lot like the Four Little Dragons of computer vision, or even worse. The Four Little Dragons didn't have a commercialization problem back then; it was just that everyone caught up on the technology fast, so margins were very low, and those were businesses the giants weren't that willing to do. Today even commercialization may not be easy, and the giants are all-in on foundation models.

So the hot deals these past few years are just these few: foundation models, Manus, that wave of humanoid robots — all possible bubbles? Zhu says that's not him talking, that's Zhang Xiaojun talking. He only says consensus in these tracks is all too concentrated, so they offset by fifteen degrees and leave the tracks where consensus is too tight. Why is VC consensus so high today? Benchmarking the US. These tracks are all hot in the US, and except for robots, commercialization is going well. The US is highly consensual too. Is their consensus justified? Their commercialization works; OpenAI and the rest have commercialized well. Will Chinese companies just need more time? He doesn't know. Time will tell the final result.

Has he ever given peers a heartfelt warning? No. He wouldn't dare warn them; everyone talks. Which earlier calls blew up in his face? He has gotten things wrong. Electric vehicles, for example: he'd been burned, been taken in, laid in way too early, invested five years before this cycle, lost a lot of money. He'd had the lesson, so when the cycle actually started they didn't dare invest. Are foundation models and humanoid robots that type today? He doesn't know. Only time can tell them.

A GPT-4 moment in his life this year? Not a single moment, but ChatGPT stunned him. One is Pulse — the product is thought through so well, an AI version of Toutiao. Group chat also went beyond what he'd imagined, and it's done very well: he put the company group into ChatGPT, and it can help them summarize what they've discussed. The user scenario is defined especially well. Search is defined well too.

Then Zhang Xiaojun says, so you still say you wouldn't allocate to OpenAI. Zhu's point is: from zero to one he thought through these scenarios especially well, but whether he can hold them is a different question. That's a big confusion of his. Should DeepSeek follow that path? DeepSeek never took that path to begin with; it never considered DAU, never considered user experience at all. Zhang Xiaojun asks whether it can turn after AGI has peaked. That's a completely different way of thinking, a completely different DNA and operations. And that's exactly why he still has questions about whether OpenAI can hold: that takes a completely different DNA and operating experience.

If OpenAI hit a tipping point that would make him have to bet, when would it be? Once a social network is built, that's formidable. If it uses group chat the other way to build social relationships and a social network, the moat would be extremely deep, and META would actually be in danger. Does charging SOTA, charging the rankings, still matter this year? Meaningless. A lot of scenarios don't need SOTA; they need value for money, a model that fits the scenario. For a lot of scenarios, Qwen's 30B small model is the best: cost is very low, accuracy is very high, you don't need SOTA at all.

And SOTA rankings have been swinging even harder this year; Google has caught up too. Spending a few hundred million dollars to train a SOTA that leads for less than a month — what's the point? For a startup, definitely none. For a giant maybe a bit, because it has to provide cloud services, and a foundation model has value for the cloud. Startups still have to look at daily usage and retention.

Any other calls for 2026? Token consumption will definitely explode further, and a lot of scenarios will be very interesting. The key for the US is how data centers and power keep up. For Chinese founders, it's seize the opening and go full speed — the next few years are too good an opportunity. Don't go looking for the big opening; find a living space in the cracks. Then maybe they don't even need to raise? Not necessarily — you still need the early money.

What kind of deal should come find him? The scenario thought through clearly, where the users are thought through clearly, how to do it thought through clearly — that's the founder he likes most. Grand narratives, he likes least. What's the grandest narrative he's heard? Too many, a lot of them ridiculous. Do post-95s and post-00s have a shot in VC? They do; there are still a lot of openings. A few of their young people are very fired up; they backed a lot of good deals this year. One deal did three add-on rounds in a year — for a young person that's a very exciting thing. Foundation-model investing was all GPs investing; young investors had no part. They still encourage the young people to swing.

Advice for young investors? Is VC still a good career today? The core is still having your own view. Don't echo everyone else; don't chase every hotspot everyone is chasing — then you have no value. You have to have your own view; that's when you have value. How does he see next year's investment openings? Honestly he's very optimistic. Two years ago at ChinaVenture he said stay at the table; today he says seize the opening, go full speed. There really are too many openings. In every niche track Chinese founders are leading.

Did he go to the US much this year? A lot — he's going again next week. What's different about the China and US startup ecosystems now? It's Chinese people, and Chinese people in the US, starting companies. He goes to the US a lot, also to talk with these Chinese founders in the US: he takes Chinese students at Stanford to dinner, has backed a lot of Stanford PhDs, and this time he's going to take the business-school students to a separate dinner. Is the China-US AI gap widening or shrinking? There's still some gap, but it's about the same, stable at three to six months, at least not widening anymore. The next few years there's a chance to shrink it; as long as the gap is shrinking, China's open-source ecosystem will definitely lead. He's especially bullish on China's open-source AI ecosystem three to five years out.

After talking he still feels there's a lot of uncertainty. At this point in time, what is he very sure of? The key is the time horizon: one or two years may have a lot of uncertainty. Stretch it to five or ten years, and he strongly believes China's open-source AI ecosystem will definitely be far ahead. Ten years from now China's AI will definitely lead the US, because US data centers and US power can't keep up. Catching up on algorithms, catching up on models, is all much easier than building data centers and building power — especially in the US, putting up a nuclear plant is a hassle.

China's AI will overtake the US, China's open-source ecosystem will overtake the US? If you stretch the time horizon, China's national strength will definitely lead the US. There's a WeChat article he especially agrees with: China and the US have already passed each other. What does that mean for founders? Chinese founders have too many openings. That's why they're excited and busy today: truly, in any country there's a Chinese startup opening.

What's a typical day this year? Talking, drinking tea and talking. The density is different. He likes packing meetings into certain days, and some days just for reading, resting, thinking. Travel is still pretty frequent; he's tired of flying, and after this US trip he wants to go on vacation. That's about it for this year.

Any new movies or shows lately? More shows, watching on planes. There's a recent one on Apple TV that's interesting: humans catch an alien virus, most people's minds get unified, memories all shared, and twelve people left un-unified are trying to figure out how to save humanity. Zhang Xiaojun asks whether this scene of AI development could become real. Zhu says that's a different concept: humanity's shared consciousness. Still gaming? Yes — phone games, Honor of Kings. Twenty or thirty minutes in the middle is enough.