Yunqi Capital's Chen Yu: AI Entrepreneurship Isn't About Speed — Direction and Endurance Matter More | Partner Perspectives

AI is evolving at breakneck speed. Where do China's large model unicorns stand now? Agent is the buzzword of the moment — but do horizontal or vertical approaches have better odds? Embodied intelligence is white-hot — how do you separate real value from hype?

These were the pointed questions on the table at 36Kr's recent Waves 2025 conference, where Chen Yu, partner at Yunqi Capital, joined the investor roundtable "One Year On" to share his candid take. Here's what he had to say.

Key takeaways:

1

MiniMax's sustained exploration in multimodal capabilities and large model architecture reflects the clear judgment and technical conviction of a startup operating with constrained resources.

2

The essence of Agent isn't tool upgrade — it's a restructuring of delivery models. From SaaS-era "delivering software" to Agent-era "delivering results," the way customers perceive value has shifted. They're now willing to pay for efficiency and output.

3

Embodied intelligence remains in a high-cost, low-certainty phase. But like early autonomous driving, the growth phase of this赛道 tends to bring both outsize expectations and disappointments. The key to early-stage investing: finding long-term players with technical chops, fundraising stamina, and the patience to endure.

4

Rapid technological evolution has dramatically compressed the growth window for high-potential companies. By the end of 2026 at the latest, the companies with real potential to become AI giants will likely have already emerged.

The following is an edited transcript from the roundtable, originally published by "Waves":

On MiniMax: Sustained Innovation Within Constraints

Waves: How do you see the large model landscape reshaped by DeepSeek, and what are the odds for the "Six Little Tigers"? Also — any insight into what Kimi and MiniMax have been up to? They've been unusually quiet for a while.

Chen Yu: AI is an industry where change happens in an instant — one year here equals three, five, even ten years in other sectors. After DeepSeek's release before the Lunar New Year, both its V3 model and R1's reasoning capabilities were stunning. That's how it seized user mindshare so rapidly. In search-app market share, it's now in a league of its own.

But that doesn't mean large models are only about LLMs or only about reasoning. Over the past year, MiniMax has delivered genuine surprises in multimodal. Hailuo's video generation, for instance, is visually impressive — and more importantly, it's achieved commercial traction.

Another standout for me was its voice synthesis model. Have you seen "Daniel Wu teaches you English" on Douyin lately? MiniMax powers that behind the scenes. The first time I heard it, I genuinely couldn't tell if it was a real person or AI. Only later did I learn it was a MiniMax client — pretty remarkable.

These details show that MiniMax has been doing interesting work all year, just not every project has been in the spotlight.

Startup resources are always finite. Even a company like DeepSeek can't cover everything. With limited resources, Junjie Yan chose to bet on video and voice models. On architecture, he's particularly focused on linear attention mechanisms — gambling on near-unlimited context length in the future. That's critical for social companion apps, and for building Agent systems.

Of course, this means he may have deprioritized reasoning models slightly. Whether that call proves right or wrong, time will tell. But what matters for a startup is this: how do you keep innovating and breaking through with finite resources?

On Agent: From "Delivering Software" to "Delivering Results"

Waves: Many say the Agent era has arrived. What signals confirm this? Compared to tech giants and model companies, where do application-layer Agent startups actually have a shot?

Chen Yu: What we've been tracking more closely this year are vertical, domain-specific Agents. And the first scene to see real deployment is likely automotive.

For example, ByteDance's Doubao automotive product, built on its large model, has already rolled out to over a million vehicles. Users can interact with the Agent system directly in their cars — so people will soon experience firsthand how Agent differs from conventional productivity tools.

Other sectors are Agent-ifying quickly too. The previous wave of SaaS companies, for instance, are gradually incorporating Agent tech — often pushed by customer demand.

The biggest difference between Agent and last-gen SaaS: Before, SaaS mostly delivered software. Whether it worked well depended on how the customer used it. So willingness to pay wasn't always strong. Now, Agent delivers results directly — the outcome is visible at a glance. If the result is good enough, customers naturally pay.

Take something like the Manus Agent system. Why has it grown so fast? Because users genuinely feel the efficiency gains. Productivity improvement is tangible, results are concrete — so they're willing to pay. That's where Agent's real value lies.

On Embodied Intelligence: Picking Promising Horses Before the Industry Settles

Waves: Another hot investment theme this year is embodied intelligence. Early in the year, Allen Zhu's exit from some projects sparked debate about whether the sector is in a bubble. What's your view? Which opportunities are near, which are far?

Chen Yu: There's definitely froth in embodied intelligence right now. Measured by traditional PE or PS metrics, it mostly doesn't hold up. But it reminds me of autonomous driving roughly a decade ago.

The 2016–2017 wave of autonomous driving startups, many going after L4 — most were still lab-bound, capable of only basic tasks. Today's embodied intelligence is similar: high costs, unclear technical paths, yet valuations and funding rounds climbing fast.

Autonomous driving then went through a 3–5 year plateau. Funding nearly dried up because early money had been poured in without the expected technical breakthroughs or commercial progress. Startups began to diverge: players with sufficient capital and solid tech stayed at the table, waiting for a commercial inflection point that might come a decade later; those underfunded or technically behind mostly didn't survive. When real commercialization finally arrived, the technology had matured, costs had dropped, and products were ready for scaled deployment.

So for us as early-stage investors, if we don't invest now, there won't be many opportunities later. We want to pick a few promising horses while the industry is still taking shape — strong technically, strong at fundraising. It's fine if the path isn't clear yet. We can place bets across differentiated routes in hardware, software, and see who emerges as the real winner.

On Investment Judgment: The Early AI Window May Be Just 3–4 Years

Waves: You've all lived through the mobile internet era. Have you noticed differences in investment paradigm between then and now? In AI's technical progress, people often cite the "bitter lesson." Any equivalent lessons in investing?

Chen Yu: Every firm, every investor has their own ingrained paradigm. Yusen is the classic "bet on people" type; we're more "bet on technology."

We've always focused on how AI development is deeply tied to the evolution of foundation models. As long as the base models keep advancing rapidly, they'll inevitably create new demands across the infrastructure stack — and may even unlock application scenarios that were previously impossible.

We pay close attention to shifts in underlying technical models. The pace of evolution now is fundamentally different from before — much faster. We sometimes joke that in the mobile internet era, from the iPhone's debut to the rise of a cohort of giants, there was roughly a 6–7 year window; in AI, that cycle may be cut in half.

In other words, the genuine early-stage AI investment window may be just 3–4 years. By the end of 2026 at the latest, the companies with potential to become AI giants will likely have already surfaced.

As for why companies fail, often it's not technology or direction — it's people. There are real cases where the failed companies weren't actually that bad. The hardest ones are those that "just started to succeed." Once a company achieves modest success, complacency sets in. Everyone feels they personally made it happen, so infighting and credit-grabbing begin — and the company tears itself apart. That's truly a shame.