MiniMax's first investor: Big tech's price cuts are peeking at cards, startups can only eat the loss
Big tech can run models at a loss and make the money back from cloud computing; startups dragged into a price war have nowhere to recoup. MiniMax's answer is model quality first, and monetizing through overseas 2C products.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Big tech can balance the books on price cuts; startups cannot
Chen Yu lays out the mechanism of the price-cut wave plainly: ByteDance and Alibaba have cloud computing businesses behind them, so models can be priced extremely low and the money made back from the cloud—the math works. Startups dragged into the price war may be losing money, and have no second place to earn that loss back. So his judgment is that this is especially bad for startups. MiniMax's response is not to fight the price war and to insist on model quality first—his reason: in China's large-model API market, GPT4's share is still over 80%, and people are willing to pay a premium for quality.
— Chen Yu2C apps can't collect money because switching costs are too low
Large-model applications fall into roughly two types: ChatGPT-style search and productivity tools, and Talkie-style social apps. The vast majority of companies only built the former—Kimi, Hailuo, Doubao are all this type. The problem is that if Doubao doesn't charge, Kimi can't charge, because quality isn't high enough to make people use only you, and switching costs are low—a phone can hold three apps at once, plus ChatGPT, and GPT4O is now free too. An ad model works, but the experience degrades, and you have to balance user experience against revenue—a problem the previous generation of search engines already explored.
— Chen YuTalkie is a rare startup product that can collect money
Chen Yu's numbers for Talkie: daily active users in the millions, registered users in the tens of millions, and this year's revenue already in the tens of millions of dollars. He treats this as MiniMax's differentiation—others don't have this product, and this product can collect money. By contrast, he splits MiniMax and Moonshot into two strategies: Moonshot focuses more on brand awareness, spending a lot on user acquisition, but model quality is not first-tier and is a lesson to be made up; MiniMax focuses heavily on models, does little PR, but product quality and numbers are better, plus the differentiation of Talkie, making its product and business model relatively leading among startups.
— Chen YuValuations of two to three billion dollars, but revenue extremely limited
Chen Yu says financing is not easy now: the top two are both valued at the two-to-three-billion-dollar level, but revenue is extremely limited, and with big tech competition, people are especially hesitant to enter now. The next few—Step, Baichuan, 01.AI—most need to solve financing, because only enough money can keep you alive at the table. He also says where the later money will come from, we have no answer, and believes these startups have no answer either.
— Chen YuClub Deals exist because no one can see who the winner is
Tencent, Sequoia, Shunwei all invested in several, and Alibaba too. Chen Yu's explanation is that everyone still can't see clearly: if there were a clear winner now, all the money would be drawn to that one, and everyone else would be running along. He adds something colder—is it possible none of them have a chance? That's also possible, and in the end big tech might harvest everyone. So to this day, one cannot say this opportunity definitely has a startup opportunity; even after more than a year of running, the answer is still unclear.
— Chen YuBig tech investing in you is peeking at cards; we are not
Chen Yu defines "peeking at cards" clearly: invest a little to get information, then either make a huge follow-on bet or help your own business. Big tech investing in startups is peeking at cards, because even if you raise a billion dollars, it's only a small part of big tech's profit. But independent investment institutions don't have those two follow-up moves, so he says we are definitely not peeking at cards. As for what investors do if startups have no chance, his answer is that you just have to eat the loss—otherwise it wouldn't be called venture capital, it's high risk high reward.
— Chen YuQualification line: cannot be worse than the leading open-source models
Chen Yu gives three layers for the qualification line of domestic large-model startups: first, as a closed-source company, model quality cannot be worse than the currently leading open-source companies—if you can't even reach that, there's no point existing; second, model quality and features must match the currently leading models, and multimodal features must come out this year, otherwise market competitiveness is insufficient; third, and most important, the company must find a business model to sustain development—whether 2C products or 2B revenue, doing projects or APIs, it must make money and cannot lose money forever.
— Chen YuChina has never seen a merger this large in its history
Asked how he would feel if a giant acquired MiniMax, Chen Yu first says he doesn't rule out the possibility, but the probability is relatively low: the valuation is there, and China has never seen a merger this large in its history; in the US, multi-billion-dollar mergers happen every few years, but not in China. The only case he cites is Ele.me, and when pressed, he says $9.1 billion. He also adds that the team may not be willing to be acquired.
— Chen YuIn their own words · checked verbatim
If your model quality isn't good enough, everything above is like a castle in the air—the foundation is unstable, and it collapses easily.
你的模型的质量不够好的话呢 上面都是像这种空中楼阁一样 就是根基不稳 就很容易坍塌的
Chen Yu41:39
I think it's actually bad for startups, because whether it's ByteDance or Alibaba, they actually have a big cloud computing business behind them. I can charge you a very low price for the model, but I can charge you more on cloud computing, and that math works out.
我觉得对创业公司实际上是不利的 因为你像字节也好阿里也好 他们其实背后是有很大的云计算的生意 我可以模型只是收你一个非常低点的价钱 但是我可以在云计算上面去收你更多的费用 然后这笔账是能算得过来的
Chen Yu43:41
You might find it hard to believe, but in China's large-model API market, GPT4's share is still over 80%.
你可能很难相信 就是说在中国大模型API市场里面 GPT4的占有率还是超过80%的
Chen Yu44:43
So to this day, we cannot say this opportunity definitely has a startup opportunity—even after more than a year of running, the answer is still unclear.
所以到今天为止 我们都不能够说 这个机会一定有创业公司的机会 就哪怕是已经跑了一年多了 这个答案还是不清晰的
Chen Yu50:52
Then you just have to eat the loss. No one says when making an investment that it's definitely a sure-win business—otherwise it wouldn't be called venture capital, it's high risk high reward.
那就只能认赔嘛 没有说大家在做一个投资的时候 就肯定说这个是稳赚不赔的生意 不然的话这也不叫风险投资了 是高风险高回报啊
Chen Yu51:52
Yes, I think all large-model startups now should have no illusions—without a business model, in the end you are unsustainable.
对我觉得是现在所有大模型创业公司 就不应该有任何幻想 没有商业模式到最后的话 你都是不可持续的
Chen Yu55:56
First, I think the possibility of a giant acquiring MiniMax is not very high, because after all the valuation is there, and China has never seen a merger this large in its history.
首先我觉得巨头收购minimax的可能性不太大 因为毕竟来说估值是放在那儿 中国历史上是没有出现过这么大的并购案了
Chen Yu56:56
Figures
| Talkie daily active users | in the millions | 46:46 |
| Talkie registered users | in the tens of millions | 46:46 |
| Talkie revenue this year | in the tens of millions of dollars | 46:46 |
| GPT4's share in China's large-model API market | over 80% | 44:43 |
| Valuation level of top large-model startups | two to three billion dollars | 49:51 |
| Number of MiniMax funding rounds | about five rounds counting small ones | 51:52 |
| GPT4 parameter count | 1.8T | 22:24 |
| Cost of Yuanrong's software-hardware solution | about $2,000 | 17:24 |
| Two-year decline in large-model inference cost | one to two orders of magnitude | 21:24 |
Glossary
- Club Deal
- Multiple institutions jointly invest in the same company, spreading bets because the winner is unclear.
- corner cases
- Extreme scenarios in autonomous driving that can only be discovered through actual operations.
- MOE
- Mixture of Experts: a large model assembled from several smaller expert models, activating only some parameters during inference.
- scale up
- Improving model capability by stacking parameters and compute.
- L4
- High-level autonomous driving: fully driverless level, requiring extremely high safety, difficult to deploy.
How to listen
Investors and founders watching large-model startup financing and business models, especially those who want to know how the price-cut wave is accounted for and where the qualification line for the six companies lies.
The opening 1:55–11:00 on GenAI conference observations and Sino-US capital topics—low information density.