Yunqi Capital's Chen Yu: How He Became the Earliest Investor in MiniMax | "Attent!on" Crossover with "Jun Zhang's Business Talks"

Generative AI has been boiling over for more than a year now, and the main contenders in China's large model startup race can be counted on one hand.

Yunqi Capital's angel-round bet MiniMax is one of them. This company, which set its sights on multimodal large models before ChatGPT exploded into the mainstream, is currently one of the highest-valued startups in its track. Its model and product moves are also seen as a window into the progress of domestic large models.

As the first angel investor to bet on MiniMax, Yunqi Capital partner Chen Yu recently sat down with Zhang Xiaojun to share the story behind that investment, and to discuss real questions like the differences between Chinese and American AI venture capital, and where China's large model poker game is heading.

In this episode, Yunqi Capital's podcast "Attent!on" joins forces with "Zhang Xiaojun Jùn Business Talks" — here's what we covered.

Scan the QR code above or follow "Attent!on" on Xiaoyuzhou to listen to this episode.

Timeline

GenAI Conference

02:16 Chen Yu's self-introduction

03:09 Silicon Valley impressions

04:42 What are the current differences between Chinese and American AI venture capital environments?

11:00 American AI startup directions aren't that different from China's overall — mostly software

AI's Evolution Over the Past 10 Years

12:38 How AI has evolved technologically and from an investment perspective over the past decade

13:42 Three waves of AI application: computer vision (which birthed the "AI Four Dragons"), autonomous driving (the difficulty of L4 deployment), and large models

20:51 After going all-in on large models, there were several distinct phases

The MiniMax Backstory

25:16 The story behind being the first investor to bet on MiniMax

30:14 Founder Junjie Yan's nickname was "IO" (Input/Output) — I took a liking to him after one meeting

38:02 MiniMax's technology, application, and product iterations (digital humans — Glow — Hailuo AI, STARFIELD, etc.)

39:52 About Junjie Yan, Yewen Ye, and the team?

42:50 How to view the large model price wars triggered by tech giants, particularly their impact on startups

"The Poker Game of Large Models"

46:07 The players: tech giants are the whales with the most chips, followed by startups with the most funding

48:02 The primary homework for the remaining startups is fundraising

48:51 Why is the Club Deal phenomenon so prominent in the large model wave?

51:18 The qualification threshold for the six large model startups this year

53:39 Tech giants' large model products vs. startups' products

Below is an excerpt from this episode

#The Story of Investing in MiniMax

Zhang Xiaojun: I introduced you earlier as MiniMax's first investor. How did you land this deal?

Chen Yu:

I studied computer science at Johns Hopkins. Our school has produced countless entrepreneurs in the NLP field. One of MiniMax's co-founders, Yewen Ye, is also a Johns Hopkins graduate. I actually met her when she returned to work at SenseTime after graduating senior year. So five, six, seven years before she actually started her company, we had already become good friends. I watched her grow inside SenseTime the whole time.

Actually, Yewen already had entrepreneurial ambitions back in 2018 or 2019, but never acted on them.

It wasn't until early 2021 that Yewen came to me and said she was finally ready to start something. That's when we began discussing what would become MiniMax. The name wasn't set yet, but the direction was clear from the start: use foundation models to solve AI problems, ultimately achieve artificial general intelligence, and build China's OpenAI. That broad direction was established.

Zhang Xiaojun: Did you already frame it as "China's OpenAI" back then?

Chen Yu:

Yes. For a long time, there were really only two companies that everyone in the industry most admired: DeepMind and OpenAI. The reason for that admiration was that both were dedicated to solving AGI problems, without being constrained by commercial considerations. Because, for example, Google actually had another organization around the same time called Google Brain. That one had much heavier commercial constraints. It had to align with Google's commercial strategy, so it wasn't quite as idealistic an organization as DeepMind.

For those of us in technology, we all have a bit of technological idealism to some degree. So we prefer companies like DeepMind and OpenAI that explore unknown territories for humanity.

Zhang Xiaojun: What did you talk about? Besides building China's OpenAI, what else came up in that first meeting?

Chen Yu:

She said she wanted to introduce me to a co-founder. So at our second meeting, I met Junjie Yan. They call him IO. IO is a common abbreviation in computer science — input, output. We started discussing foundation models then. As an investor, you have to be on roughly the same wavelength as the founder. You have to buy into the direction the founder has chosen, otherwise you can't make an investment decision. So we mainly exchanged views on foundation models and AGI.

Zhang Xiaojun: After those first two conversations, what judgment did you make at the time? Because OpenAI wasn't nearly as hot then as it is today.

Chen Yu:

Actually, there wasn't really a "judgment" per se. But Junjie made a very strong positive impression on me, because he was the first person to talk to me this deeply about foundation models. At the time, BERT was still mainstream. Almost every AI project I met with told me they were using BERT or some BERT variant.

Another thing that set him apart from the vast majority of entrepreneurs was that he wanted to solve an enormous problem. Everyone knows AGI is an incredibly difficult problem to solve, but if you do solve it, you're looking at a hundred-billion or trillion-dollar market. Without sufficient ambition and self-confidence, it's hard to even conceive of tackling something like that.

#Silicon Valley Impressions

Zhang Xiaojun: Did you gain any new information or insights about the venture capital environment over there?

Chen Yu:

Right now, AI investment in the US is still relatively FOMO-driven. Everyone is hustling to find the next company that can scale big, and valuations are relatively high. There are some objective reasons for this. AI engineers are quite expensive right now, so their funding rounds can't be too small — otherwise they can't afford a strong AI team. You can see from this that valuations won't be particularly low either.

Zhang Xiaojun: Are many Chinese VCs going to the US to look at this wave of AI projects? Have they seen anything good?

Chen Yu:

Currently, Chinese investors are looking at quite a lot of projects in the US, but the vast majority are still limited to investing in Chinese-founded teams. If you're asking about standout examples, everyone would point to HeyGen as the benchmark — a标杆 for Chinese companies going global.

Zhang Xiaojun: Did you also meet some Chinese entrepreneurs this time? How do their current stage and direction compare to domestic AI entrepreneurs?

Chen Yu:

I don't think the difference is that large. Everyone is all-in on AI, and the vast majority of companies are still software-focused. Because US VCs have historically not liked investing in hardware, and hardware advantages actually lie with Chinese domestic companies. China has complete supply chains locally. If you want to do hardware in the US, costs are relatively high. So apart from a few particularly well-known robotics companies, the US rarely invests in hardware. The vast majority is software, and enterprise services tends to dominate.

#On the Large Model Poker Game

Zhang Xiaojun: I wrote an article recently called The Poker Game of Large Models. From your perspective, how do you see this poker game playing out?

Chen Yu:

There are clearly several categories of players here. First, the tech giants. For them, the advantage is definitely abundant resources and strong financial power. Second, they have business synergy advantages — for example, the synergy between cloud computing and large model businesses. You can charge nothing for the large model and make your money from cloud computing. From the poker table perspective, they're definitely the whales with the most chips, so relatively speaking, they're in a very favorable position.

The second category is startups with relatively leading funding. Like Moonshot and MiniMax — I think they've pursued different strategies. Moonshot actually places more emphasis on brand awareness, and has done relatively well on visibility and active user numbers. But the homework they need to make up is model quality — we'll have to see how they handle that going forward. MiniMax is very focused on models. They don't do much PR — relatively little, in fact. But in terms of product quality and metrics, they're actually quite good.

Behind them you have Stepfun, Baichuan, and 01.AI. For them, the most pressing problem to solve is fundraising. Because only with enough money can you continue surviving at the table.

Zhang Xiaojun: What do you think the goal is for domestic large model startups this year? What's the qualification threshold to advance to the next round — whether in funding, technology, or product?

Chen Yu:

As a closed-source large model company, the qualification threshold is definitely that you can't be worse than the current leading open-source models. If you can't even hit that bar, then you basically have no reason to exist. Beyond that, whether in model quality or capabilities, you'll need to match the current leading models. So these companies' multimodal capabilities really need to ship this year too — otherwise their competitiveness in the market won't be strong enough.

Most importantly, for a company to sustain itself, it still needs to find a business model. Whether you're doing consumer products or enterprise revenue, whether through projects or APIs, you have to make money. You can't keep losing money forever.