Linear Zheng Can: AI Apps Are in the "Pre-iPhone 6" Era

"Where the Gaps Are." By Muxin Xu

Since late 2023, Zheng Can of Linear Capital has maintained a pace of at least twenty in-depth conversations with AI startups every month. By his own count, he's now taken a close look at several hundred AI companies. Within Linear, there's a view that competition at the model layer will continue indefinitely — the firm never saw that as the right opportunity or playing field for startups. The wave of innovation unleashed by model improvements has sent a clear signal to entrepreneurs and investors focused on the application layer: a new race is on here, and the first step for investors joining it is confronting a relentless stream of new projects arriving by the hour.

Zheng told Waves: Linear invested in 17 projects last year. "The founders we fear most are those who have no genuine obsession with landing in a specific scenario, and are simply using AI as a buzzword to dress up their pitch decks."

This year, market winds have shifted again. Manus's emergence has made general-purpose agents the hot new direction for entrepreneurs. AI hardware, having completed its fundraising and shipment cycle last year, has entered hand-to-hand combat. Embodied intelligence keeps notching sky-high valuations.

Against this backdrop, we sat down with Zheng to discuss: Is a general-purpose agent something a startup can actually build? Why has Linear Capital increased its check sizes this year? And what gaps remain in the application layer for entrepreneurs to enter?

Here are the core takeaways from Zheng:

1. Agent products have been proliferating lately, but none have matched Manus's early-year buzz on the consumer side. In the end, C-end general-purpose agents may need to consolidate to "just one" to capture all attention and reach critical mass. The enterprise market is a different story. Take Anthropic and OpenAI: the latter has roughly 10x the users and DAU, yet their B-end revenues aren't far apart.

2. The projects that interest us are agents solving specific vertical problems. In concrete scenarios, problems and outcomes are easier to define, and commercialization is more direct — even if the upside isn't as open-ended as general-purpose products.

3. We're currently focused on three directions. First, coding. All existing coding tools have limitations when working with existing codebases, so there's room for many companies to grow here. Second, voice model-related projects. From last year to this year, voice models crossed an important threshold — they can now produce fully human-like speech. There's a U.S. company called Think Labs that's using this kind of voice model to conduct phone interviews; unlike a simple questionnaire, the voice model can follow up based on user responses. Third, AI + silver economy. We're facing the most severe aging demographics in history, so AI must address elder care. For now, this can start with emotional companionship.

4. Our minimum check sizes will be higher this year. If last year's early-stage rounds were $1.5–2 million, this year they'll be $3–5 million. First, startup teams are maturing. Second, many companies are shifting their existing business focus to AI — and these incumbents actually have major advantages over newcomers in terms of processes and data.

5. I think AI applications are in a "pre-iPhone 6" era — there's a lot that can be built on top, but we haven't yet reached the moment when giants like Meituan or Dianping emerge. And AI is productivity, not a new channel. Essentially, everything that could be done before can be redone with AI. From a commercial opportunity standpoint, you can somewhat "copy the homework."

6. Our past AI application investments have been tools, or rather, the application companies that emerged over the past period have all been tools. But we've been thinking about whether there's a larger structural opportunity — not just technology, but technology multiplied by a population. My answer now might be community. Communities come in different forms. For smaller ones, they can increase tool stickiness. Larger communities are a different concept entirely.

7. AI applications and embodied intelligence are completely different in fundraising. With the latter, you need to prepare for "three years" without revenue, funded entirely by investment — so fundraising capability is crucial in that space. AI applications are different. I think AI applications should prove their commercial viability quickly today, rather than needing to raise huge amounts. They iterate fast.

8. From an early-stage investor's perspective, the question I think about most often is: how much will this entrepreneur need to spend to hit the next few milestones, and how much more if they get one or two chances to fail?

9. We currently have little interest in pure generation — that's the model makers' job. I hope entrepreneurs choose directions that benefit from model progress, not the other way around. The moat lies in finding the right scenario to amplify that advantage.

Image source | Unsplash

Recommended Reading

Where the Money Flows, Where People Rise and Fall