No AI bubble in sight for three years, but concentrated VC consensus is the real risk
Zhu Xiaohu says the bubble talk is secondary-market investors deliberately steering a correction; the real problem is that every primary-market deal has become a Club Deal, so GPs can't make big money and LPs aren't happy either.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
OpenAI's shift from weekly to daily actives is a strategic turning point
Zhu Xiaohu believes OpenAI has all but stopped talking about AGI this year, shifting strategically from weekly-active scenarios to daily-active scenarios. Weekly-active scenarios are hard to defend against big-platform attacks, while daily-active defenses will only get thicker. Pulse, the browser, group chat — all of it is about increasing usage frequency, refocusing from weekly to daily actives. He judges that ChatGPT will become a new super-app entry point, and that once group chat is established it may rebuild social relationships, with a big impact on Meta.
— Zhu XiaohuThe Transformer architecture has basically hit its ceiling
Zhu Xiaohu says LeCun has been talking all along about scaling hitting a wall, and over the past year it has become obvious; going further is marginal improvement at high cost, with very limited gains. At least on the Transformer architecture, you could say it's hit the ceiling. But he thinks today's technology level is already enough to support an application explosion — industry-wide token consumption is up more than tenfold, proving this wave of AI tech is already good enough for many scenarios.
— Zhu XiaohuThe competition ahead isn't AI competition, it's electricity competition
Zhu Xiaohu says depreciating GPUs over two or three years is simply impossible — plenty of five-, six-, even seven- or eight-year-old cards are still in use, and as long as you have a card you can use it. Token consumption is exploding too fast: more than tenfold this year, and it could explode another tenfold next year, while data center construction lags far behind. He believes the competition ahead isn't AI competition but data center construction, electricity competition, and that China building a dozen nuclear plants and lots of solar farms is absolutely the right call.
— Zhu XiaohuUS AI valuations are 100x China's — one end must be wrong
Zhu Xiaohu says US AI valuations are 100 times China's, and one end must be wrong: either its revenue is unsustainable, or China's revenue is undervalued. He thinks there are some points of divergence here, and how it evolves from here is still unknown. On the GPU flywheel, he says it's a bit like the internet bubble back then, but understandable, because it needs to be off-balance-sheet — an accounting trick by some listed companies to make the financials look better. The core question is still whether anyone is actually using this thing and whether it's good enough.
— Zhu XiaohuWithout DeepSeek, human AI might be controlled by a few private companies
Zhu Xiaohu says people still underestimate how much DeepSeek changed all of humanity, changed history. He believes that looking back ten years from now, DeepSeek will still be a very important turning point in the whole AI competition — otherwise China's open-source ecosystem today wouldn't be so resolutely open. Without DeepSeek, it's quite possible that humanity's AI would be controlled by the AI models of a few private companies, and that would be dangerous for all of humanity.
— Zhu XiaohuIn the AI era you need to be three streets away from the big platforms
Zhu Xiaohu says that in the mobile internet era people said stay one street away from the big platforms; in the AI era you may need to stay three streets away. Concretely: first, go vertical on applications; second, do private deployment, for example taking Qwen's 30B small model for private deployment; third, go do sales. Big platforms generally also look for startups to help them with last-mile implementation. He has invested in several agents for vertical scenarios like this, growing at a pace of tens of millions last year, over 100 million this year, and two or three hundred million next year.
— Zhu XiaohuVC consensus is too concentrated — GPs can't make big money and LPs aren't happy
Zhu Xiaohu says that at a VC meeting in Hong Kong last week, many LPs were complaining that in this cycle GP consensus is too concentrated: every deal is a Club Deal, several funds investing together, each taking a very small stake — so how do you make money? GPs can't make big money, and LPs aren't happy either. And investing with different VCs feels about the same, with similar portfolios, so there's no risk diversification for LPs either. He thinks an obvious problem this cycle is that consensus is too concentrated, because everyone's aesthetic has become too similar.
— Zhu XiaohuLarge model companies are worse off than Cisco — even commercialization isn't easy
Zhu Xiaohu says these large model companies are a lot like Cisco in the last cycle, or even worse off than Cisco. For Cisco back then, commercialization at least wasn't a problem — it was just that other technologies quickly caught up, so gross margins were very low. Today it may genuinely be hard even to commercialize. Big platforms weren't too keen on Cisco's businesses, but today the big foundation-model players are all-in desperately. He believes that if commercialization never materializes, then the consensus is a bubble.
— Zhu XiaohuIn their own words · checked verbatim
I think there's no bubble in sight for at least three years — and when everyone is talking about a bubble, the bubble definitely hasn't arrived yet.
我觉得至少三年内看不到泡沫 而且大家都在讲泡沫的时候 泡沫肯定是没到的
Zhu Xiaohu0:00
The competition ahead isn't AI competition — it's the data center explanation, it's electricity competition.
后面竞争不是AI竞争 是数据中心的解释 是电的竞争
Zhu Xiaohu10:07
US AI valuations are 100 times China's — one end must be wrong: either its revenue is unsustainable, or China's revenue is undervalued.
美国的AI的估值 是中国的100倍 肯定有一端是错的 要么它的收入是不可持续的 要么是中国的收入被低估了
Zhu Xiaohu12:08
Figures
| Annual growth in AI industry token consumption | More than tenfold | 9:06 |
| Ratio of US AI valuations to China AI valuations | 100x | 12:08 |
| Daily token consumption of a small company | Tens of billions | 10:07 |
| Daily token consumption per million DAU | Tens of billions | 20:12 |
| Revenue growth trajectory of vertical agent companies | Tens of millions last year, over 100 million this year, two or three hundred million next year | 25:14 |
| Profit threshold for consumer companies listing in Hong Kong | Over 100 million USD | 33:14 |
| Revenue threshold for 2B companies listing in Hong Kong | Over 100 million RMB | 34:14 |
| China-US AI gap | Stable at three to six months | 45:17 |
Glossary
- Club Deal
- Multiple funds investing jointly in the same project, each taking a very small stake.
- Dark Fiber
- Fiber laid but unused during the 2000 internet bubble, one of the bubble's indicators.
- SOTA
- State of the Art, meaning the current highest-performing model.
- Token
- The basic unit of text processed by large models; consumption is often used as an indicator of AI application activity.
How to listen
Founders and investors watching for opportunities in the AI application layer, especially those trying to understand the concentrated-consensus problem in the primary market and looking for a differentiated ecological niche.
The opening pleasantries and closing chit-chat can be skipped; the bubble argument and investment-strategy sections in the middle have the highest information density.