Dialogue immediately: 2.5 Lab technical lead: When building AI applications, we prioritize profits first

Author | Huang Yu

Every time an AI large model is upgraded, it further compresses the barriers of a batch of AI app products that rely on single-point capabilities—and may even strip them of the value of existing independently.

This leaves AI app startups shrouded in a constant sense of uncertainty.

For Xia Junchen, the technical负责人 of 2.5 lab under Jike, the AI applications he is focused on developing right now include an AI travel translation app called kulikuli. His biggest worry about the competition is Google Translate.

Xia Junchen frankly told Wall Street Insights, “The team is facing pressure from the rapidly changing market environment every day”—a problem every AI app team has to deal with.

But he believes that big model manufacturers and AI app entrepreneurs are fighting two different battles, with different demands. Small companies care more about how to build a profitable product. In addition, some tracks that startups focus on are easily marginalized by large firms; once marginalized, small companies still have a “gap” to survive.

The most successful product that Jike incubated is the podcast platform Xiaoyuzhou. About four years ago, Jike established 2.5 lab. This department mainly works on AI innovation projects.

Xia Junchen told Wall Street Insights that the name of the department is intended to build products for the 2.5% of early adopters who are eager to try, explore what problems AI can solve that used to be impossible, or re-optimize issues that already existed.

In the past few years, 2.5 lab has incubated multiple products in succession, including the open-source AI client Chatbox, ChatHub, which focuses on “using multiple large models on one interface and comparing outputs side by side,” kulikuli, and the self-discipline tool “Self-discipline Stone.”

Xia Junchen revealed that although the user base is not very large, all of these products launched by 2.5 lab are already profitable.

When it comes to customer acquisition at scale versus profit, Xia Junchen prioritizes profit.

Over the past decade or more, internet products have almost followed the same growth formula: acquire users for free, quickly expand DAU, and then find a business model.

But in the AI era, this approach starts to fail. The reason is simple—cost.

Unlike traditional internet products, AI services cannot expand indefinitely at extremely low marginal cost. Every model call and service delivery incurs real costs. As call volume scales up, Token costs quickly become the “money-eating beast.”

Xia Junchen used his real experience to prove this.

He told Wall Street Insights that now nearly all AI apps face the same problem: Token costs often account for around 70% of total revenue. If daily revenue is 1,000 yuan, maybe 700 yuan has to be paid for model fees. Under this cost structure, trying to grow in the way of the internet era—“burn money to buy users”—is almost impossible to make work.

Meanwhile, the primary market also won’t be willing to pay purely for user scale. Compared with DAU and downloads, investors now care more about whether revenue, ARR, and commercialization efficiency are viable. Only when the business model is proven to work—showing that the product truly has paying users—does further growth become meaningful.

“Business commercialization must be validated first, then growth.”

In Xia Junchen’s view, this has become the new order for AI app startup. And kulikuli, which grew from a once nearly abandoned “toy project” into a profitable product with several million users, is exactly what happened along this path.

It is understood that the first three months after kulikuli launched were free. At that time, due to cost pressure, Xia Junchen was somewhat unsure where this project should go, and even discussed “maybe we should stop this project.” Later, they tried adding paid features, and it turned out that real people were willing to pay—so only then did they get the project running in the right direction.

According to Xia Junchen, kulikuli now has more than 3 million total users, mostly overseas users. The team is about 10 people, with solid profits and revenues.

Xia Junchen believes that because the translation quality of Google Translate still can’t reach kulikuli, kulikuli still has room to survive.

As for why Google Translate delivers worse results than a product from a small company, Xia Junchen thinks that travel translation—like the kind of scenario kulikuli targets—is too remote and not profitable enough for a big company like Google, while kulikuli can do well because “the team puts a lot of effort into it, and we also have some engineering accumulation.”

In addition to product competitiveness, how to reduce model costs as much as possible is also something the development team must consider.

Xia Junchen revealed that while maintaining quality unchanged, it is feasible to cut 50% to 70% of costs for mature scenarios by using services like Tencent Cloud’s TokenHub.

Once survival is solved, ambition will grow out too. Xia Junchen said it’s somewhat like “leveling up monsters”—after climbing a hill like kulikuli, the team doesn’t really want to climb the same hill again, and will want to find a higher one.

In fact, when 2.5Lab does new projects, there is no fixed track and no so-called objective evaluation system. Xia Junchen said, “It’s more subjective.” Nowadays, with agent tools such as WorkBuddy, anyone in any role can—if they come up with a product idea, build a prototype, and can continuously push it forward—have the chance to form a team and truly build the project.

Risk disclosure and disclaimer terms

        The market has risks; investment requires caution. This article does not constitute personal investment advice, and it also does not take into account any special investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article align with their specific circumstances. Invest at your own risk.