Reported by |
Edited by | Wen Shuqi
This year, Lenovo Capital has seen its investment returns hit an all-time high, with six IPOs in the past year featuring star tech firms like Zhipu, Moore Threads, and Muxi.
You could say Lenovo Capital’s long-standing bets on computing power, AI, and embodied intelligence are finally paying off big time. But even now, Song Chunyu, Chief Investment Officer and Senior Partner at Lenovo Capital, still looks back at his FOMO (fear of missing out) moment from a few years ago.
Back around 2023, when the “Big Six” large model companies—including Zhipu, Kimi, and DeepSeek—were popping up and raising rounds one after another, Lenovo Capital had talks with a few of them but ended up investing in none. Looking back, Song Chunyu admits their success with AI 1.0 investments like Megvii and 4Paradigm made them way too picky when sizing up the new generation of players.
It wasn’t until early 2025, after four or five deep conversations with Zhipu’s founder Tang Jie, that Lenovo Capital finally decided to jump in. Song Chunyu explained his reasoning: “All six companies were chasing To C (consumer) markets, but only Zhipu was doing To B (enterprise) business. Large models going into every industry will definitely need a To B play, so Zhipu was bound to have its chance.”
Fast forward to earlier this year, Zhipu went public on the Hong Kong Stock Exchange, and Lenovo Capital walked away with returns worth tens of times its initial investment. But Song Chunyu admitted the pressure at the time was intense.
Today, innovation in AI is still moving at lightning speed. After that experience, Song Chunyu has been way more proactive in mapping out the relationship between base models and applications. In his view, the next big frontiers after large language models are AI coding, AI multimodal, and AI for Science—all super promising directions.
Song Chunyu believes the overall trend for AI development is crystal clear, but the exact path technology will take still needs cutting-edge exploration and innovation to prove itself.
Base Models Will Gobble Up General-Purpose Apps
Back in March 2025, the general-purpose AI agent Manus went viral globally, triggering a gold rush in the venture capital market for the next killer AI app. But more than a year later, Manus’s business acquisition hit major bumps, and the product has been squeezed nonstop by the ever-improving capabilities of base models.
And it’s not just teams like Manus that couldn’t hold the line. Even Cursor, which led the AI coding space for a long time, has faced serious pressure from base models like Anthropic and OpenAI.
Song Chunyu sees the biggest shift in the AI app layer over the past year as base models eating up general-purpose applications. For an AI startup to survive, it has to draw a clear line between what it does and what base models can easily swallow.
Based on his analysis of how models and apps evolve, Song Chunyu reckons that any app handling public data on the internet will be instantly replicated by base models the moment it launches. “I’d say 50% of AI apps will eventually be eaten by base models.”
According to his team’s thinking, an AI app that can stand on its own needs three things: really long-term tasks, integration into enterprise workflows, and proprietary data. This year, Lenovo Capital’s investments in AI apps have all followed this logic.
For instance, Lenovo Capital joined Kuaishou’s Kling AI’s first funding round after its spin-off. Song Chunyu explains that AI short films are a long-term task base models can’t swallow. “Every short film needs scene transitions, which are super complex. Base models just can’t train to match the taste of directors like Zhang Yimou, Xu Zheng, Feng Xiaogang, or Tom Hanks.”
The reasoning behind enterprise workflows and proprietary data is more straightforward: if data isn’t available on the public internet, base models can’t train on it, so they can’t beat you at that game.
As for AI coding, Song Chunyu argues that since there’s no flywheel effect between base models and a company’s proprietary code in training, it’s tough for base models to eat up industry-specific code assets. Still, he says, “There’s basically no room for startups there anymore.”
Big Divisions in VC Views on AI’s Future
Right now, every VC firm sees the massive opportunity AI brings. But when it comes to who really owns those opportunities, what startups’ survival boundaries look like, and the value of short versus long paths, the industry is deeply divided.
Take the relationship between base models and AI apps, or whether the scaling law of large models has peaked. On these core issues, investors often hold opposing views.
Song Chunyu points out that one major disagreement is whether AI-driven industrial change can be sustained. AI’s energy consumption is enormous, leading some investors to be pessimistic about its broad adoption, while others believe it’ll push new energy tech like controlled nuclear fusion into widespread use.
On top of that, no one has a clear answer to what the “Token economy” business model really is. Some say it’s delivered through cloud computing, others think it should be bundled with the model itself to calculate economics by selling models. And then there’s the more disruptive view: “You should charge based on the value of the outcome delivered, not by how much computing power you burn.”
Song Chunyu believes the big breakthrough in this AI wave depends heavily on paradigm-shifting scientific achievements—a path that’s naturally uncertain and, at its core, a gamble, totally different from pure business-model-driven projects.
“Everyone’s obsessed with one-to-three-month timelines right now,” he says. “If a company doesn’t ship something impactful in a month, it falls behind. If it can’t do it in three months, it’s basically game over.”
But he also adds that having no consensus on AI’s next phase has its upside. “If all investors agreed, they’d just drive up prices.”
