Immediate 2.5 lab technical lead: When building AI applications, we will prioritize profit first

Author: Huang Yu; Source: Wall Street Insights

Every time an AI large model is upgraded, it means that a batch of AI application products that rely on single-point capabilities will have their product barriers further compressed, and may even lose the value of existing independently.

This keeps AI application startups shrouded in a sense of uncertainty.

Among the AI applications that Xia Junchen, the technical head of 2.5 lab under Jike, is currently focusing on, is an AI travel translation app called kulikuli. His biggest worry about the competition is Google Translate.

Xia Junchen told Wall Street Insights, “Our team faces pressure from the rapidly changing market environment every day,” which is a problem every AI application team encounters.

But he believes that large model makers and AI application 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 easy for big companies to marginalize; once marginalized, small companies still have some room to survive.

The most successful product incubated by Jike is the podcast platform Xiaoyuzhou. Roughly four years ago, Jike established 2.5 lab, a department mainly dedicated to building some AI innovation projects.

Xia Junchen told Wall Street Insights that the origin of the department’s name is to build products for the 2.5% early adopters who are eager to try, to explore what problems AI can solve that couldn’t be solved before, or to re-optimize problems that already exist.

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

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

When it comes to acquiring customers 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, then look for 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 be expanded indefinitely at extremely low marginal cost. Every model call and service delivery incurs real costs. As usage scales up, Token costs quickly become a “money-eating monster.”

Xia Junchen used real experience to prove this.

He told Wall Street Insights that now almost all AI applications face the same problem: Token costs often account for around 70% of total revenue. If revenue is 1,000 yuan per day, then as much as 700 yuan may have to be paid for model expenses. Under this cost structure, it is almost impossible to grow by following the internet-era approach of “burning money to buy users.”

Meanwhile, the primary market is also unwilling to pay purely for user scale. Compared with DAU and downloads, investors now care more about whether revenue, ARR, and commercial efficiency are actually viable. Only after the business model runs can growth later make sense.

“Growth must come after commercialization is validated.”

In Xia Junchen’s view, this has become the new order for AI application startup development. And kulikuli’s journey—from a “toy project” that once seemed close to being abandoned to a product with several million users and profitability—is exactly following this path.

It’s reported that the first three months after kulikuli’s launch were free. Back then, due to cost pressure, Xia Junchen already felt a bit unsure about where the project should go—he even discussed, “Should we just stop this project?” Later, they tried adding paid tiers; as it turned out, real people were willing to pay. Only then did they get the project to run in a sustainable way.

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

Xia Junchen believes that since Google Translate’s translation quality still cannot reach kulikuli’s level, kulikuli still has room to survive.

As for why Google Translate performs worse than a small company’s product in this regard, Xia Junchen thinks that travel translation is too remote and too unprofitable for a huge company like Google. kulikuli can do well because “a lot of team effort is devoted to this, and we also have some engineering accumulation.”

Besides product competitiveness, reducing model costs as much as possible is also something the development team must consider.

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

Once survival is solved, ambition will emerge. Xia Junchen said it’s a bit like “leveling up and fighting monsters”—after climbing a mountain like kulikuli, the team doesn’t want to climb the same mountain again; they’ll want to find a higher one.

In reality, 2.5Lab’s new projects do not have a fixed track, nor do they establish a so-called objective evaluation system. Xia Junchen said, “It’s more subjective.” Nowadays, with Agent tools like WorkBuddy, people in any role can, as long as they propose a product idea, build a prototype, and can continuously drive it forward, get a chance to form a team and actually bring the project to fruition.