---
title: "A 30-Person Team Earns Tens of Millions in USD—Their Secret Is Not Hiring Ordinary People"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-06-04T11:55:15+00:00"
canonical: "https://ffcap.cn/en/research/src-20260604-01html"
source: "https://uniqueresearch.substack.com/p/src-20260604-01html"
language: "en"
---

# A 30-Person Team Earns Tens of Millions in USD—Their Secret Is Not Hiring Ordinary People

_Original · Unique Research · 2026-06-04_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay and full panel, including all named speaking turns and their continuation paragraphs. Revenue, pricing, token-consumption, team-size, market and projection figures are source or speaker claims, not independently audited findings. Company, personal and work titles are transliterated where official English forms remain unverified. The source is dated June 4, 2026._

AI Industry Observer

Your AI Product Probably Has Fake PMF

The illusion of built-in PMF in AI startups is the hardest trap to recognize in this industry

"

Your AI product probably has fake PMF.

Fewer than 30 people, no external funding, two AI products with ARR exceeding 10 million USD.

When Kuse.ai's CPO Aaron was asked how they did it, he did not start with execution. Instead, he led with a judgment most people do not want to hear:

"Your AI product probably has fake PMF."

What he means is: general model capabilities are strong enough that whatever product you cobble together, users can perceive value. But the value users perceive comes from the model itself, not from your ability to enter a scenario. The illusion of built-in PMF in AI startups is the hardest trap to recognize in this industry.

At a roundtable in Shenzhen in late May, four AI founders discussed several different cross-sections around this main thread.

Testing Real vs. Fake PMF: There Is Only One Method

Aaron gave a specific test standard: mark up the Token cost by more than 50%. If you can still sell it, the business is viable; if users run away as soon as you charge, the business is structurally unsound.

The logic behind it: for AI applications with gross margins below 10%, the money users pay essentially passes through to upstream providers. Your revenue, structurally speaking, is accounts receivable for OpenAI or Anthropic—you are just acting as a reseller.

Kuse's pricing is a direct reflection of this logic: Kuse workspace starts at US$39, and Junior as an AI employee starts at US$2,000 per month. High average selling price is not about being expensive for its own sake; it uses price to screen scenarios, and scenarios that do not fit are not brought in.

Many people at this point will ask: even if my PMF is real, what if the model vendor directly builds a workspace—how do I compete? Kuse launched two months before Claude, and after Claude came out, users did not leave in large numbers. This forced Aaron to his next judgment.

The Product Has No Moat—Switching Cost Is

His answer is: the product itself has no moat; switching cost is the only moat.

Migrating a PPT tool—using yours or theirs makes no difference. Migrating an AI employee deeply embedded inside an organization means migrating the entire organization's context, accumulated memory, workflows and collaboration habits. After Junior enters an organization, it handles communication and coordination, possesses organizational memory, and orchestrates workflows—the replacement cost is so high that people are unwilling to move.

This logic holds for all AI products: the more general, the shallower the moat; the deeper into a scenario and the more context accumulated, the higher the switching cost. A product positioned as "the best AI assistant" is harder to build long-term defense than one positioned as "your company's second sales VP."

This also means that in AI-era product design, the core question is not whether features are numerous enough, but what you are helping users accumulate and what you are making it increasingly hard for them to leave.

Design Paradigm: From Newspaper to Office

Aaron used an analogy to make this logic clearer.

The essence of internet products is information. Designing an internet product is like designing a newspaper: how information is arranged, how it is structured, and the layout determine user behavior; revenue comes from advertising and subscriptions.

The essence of AI products is productivity. Designing an AI product is like designing an office or workshop: people submit requirements to the workshop, agents execute in the workshop, and where the filing cabinet (memory) is and where the toolbox (skills) is determine execution efficiency and cost.

Under the newspaper logic, you design an information container, and user dwell time is the North Star metric. Under the office logic, you design a productivity container, and Token consumption is the North Star metric—because Token consumption represents agents at work; the greater the workload, the deeper the organization's dependence on you, and the higher the switching cost.

Two sets of logic, design decisions diverge from the starting point.

First Turn Yourself into an AI-Native Organization

It is easy to figure things out in theory; the hard part is walking through it first in practice.

Kelly, founder of Buda.im, former CTO of Heytea and founder of Vika, took this path: first transform his own company into an AI-native organization, then productize that experience.

Inside Buda, people no longer perform repetitive tasks—filling out invoices, producing Excel files, uploading files are all handled by AI. His conclusion: the role of humans is goal management, not execution.

He brought this logic into product design. On Buda, users open the marketplace, select the agent employees they need, and set goals—rather than learning how to use AI to make PPTs or videos. The difference is: the latter treats AI as a skill tool, while the former treats AI as an organizational member.

He used an analogy: when Jack Ma started a business, he needed to set goals and find people; he did not need to learn PPT skills. If your product makes users learn skills, the market you serve is shrinking; if you let users manage goals, the market you serve is expanding.

Buda currently charges by agent, US$20 per agent per month. It started in mid-March, first using a course for cold start: 300 students, 400 yuan each, and the course teaches using Buda's product, so students naturally convert into software users. The product was built within one month, course sales began after two weeks, and afterward it entered the bug-fixing phase.

What Silicon Valley Is Talking About, and 600 Billion Agents

A few external coordinates, for reference.

Kelly, who just returned from Silicon Valley, said he attended many events, and not one was discussing OpenClaw—the topics were almost all about AI Native Organization. His originally prepared OpenClaw sharing was suggested by the host to be changed to an agent topic. Li Kejia, founder of Botlearn, also observed that most participants at Silicon Valley AI events are engineers working on databases, security and infra; top-ranked developers in open-source communities focus on sandboxes, storage mechanisms and memory designed specifically for agents. OpenClaw is just a tool, not the topic itself.

Meanwhile, Li Yafei, founder of OpenClacky, is doing something in the opposite direction: a fully open-source, fully free local alternative to Claude Code, with weekly installations doubling since May. His judgment is that software itself is becoming cheaper and cheaper; rather than competing on pricing, it is better to open it to the community for co-evolution.

Li Kejia gave a larger coordinate: the Earth now has 8 billion people, and within five years there may be 600 billion agents—100 times. The driving force is the continuous decline in compute costs. He believes the tipping point has arrived.

Whether this prediction is accurate can be verified in a few years. But it means one thing: if you design an AI product today only considering human users, by the time product design is complete you may already be behind.

Consolidating the four people's judgments:

Software will become cheaper and cheaper, but deeply contextual services will not.

AI applications that cannot be sold with a 50% markup are probably just helping upstream collect payments.

When designing an AI product, you are designing a productivity container, not an information container.

More Details from the Conversation

Guests:

Buda.im Founder — Kelly

Kuse.ai CPO — Jiao Zhengdao (Aaron)

Botlearn Founder — Li Kejia

OpenClacky Founder — Li Yafei

Moderator: Unique Research Founder — Wu Wei

Guest Introductions and the Awakening of AI-Native Awareness

Wu Wei: Welcome everyone. Several of you just returned from the Google conference in the United States, and I am sure you will bring very real insights. We have discussed today's theme two or three times, because changes are happening: some products have already become "headless," relying purely on Agents to provide CLI and Shell, while others still maintain their original interfaces—worth exploring repeatedly. Before we begin the specific content, please each give a brief one-sentence self-introduction.

Kelly: Hello everyone, my name is Chen Peilin, and I am now the founder of Buda.im. Previously I was also the former CTO of Heytea.

Wu Wei: By the way, at what point did you feel that products needed to be built for the Agent era—that this change was beginning?

Kelly: I think of it in reverse: not when I started building software for Agents, but when I felt I should stop building software for humans. That feeling came earlier—when I was working on multidimensional tables in 2022 and 2023, I felt that humans are actually not suited to using software. It was not until early this year, after using OpenClaw, that I saw many non-technical colleagues in the company chatting with AI as if possessed. In that instant I understood that this is how the real world should be used in the future.

Wu Wei: So how much of your day do you spend communicating with AI, and how much with humans?

Kelly: Counting only typing, 70%, even 80% or more is with AI. Actually now I am not very willing to do voice calls with people anymore.

Wu Wei: It is rare today, being forced to communicate with humans. Aaron, please introduce yourself.

Jiao Zhengdao: Hello everyone, I am Aaron Jiao Zhengdao, co-founder and CPO of Kuse.ai. We have two products, Kuse and Junior. Kuse is an All-in-one Workspace, originally designed for humans; after Junior launched last year, it opened its own CLI and evolved into a workspace designed for both humans and AI. Before this I was always an internet product person; my previous startup "Shuangjing" (a niche designer-toy community) went from zero to 1 million users in one year and was successfully exited.

Wu Wei: How did you cold-start to 1 million users?

Jiao Zhengdao: A large part was luck—timing, location and people. At the time Pop Mart was very hot, and combined with the pandemic, this drove many independent designers and users seeking high-ASP, self-expressive designer toys. Once supply and demand were established, just like Dewu does with shoes, new opportunities for issuance and secondary trading could emerge.

Wu Wei: Kejia, please introduce yourself.

Li Kejia: Hello everyone, my name is Li Kejia. Botlearn is an edtech company we registered in Silicon Valley. Our parent company is Ouraca; in early 2025 we launched an overseas version of "Dedao," which has the highest App Store rating in its industry across more than 180 countries. After OpenClaw came out in January this year, we kept thinking: in the Agent era, do humans still need to learn, and how? We found that humans may not have time to learn anymore, so we immediately built Botlearn—a platform entirely for agents to learn and evolve. I have been in edtech for over a decade; previously I was secretary-general of Gaoshan Academy, and before that founder of Jike Big Data (acquired by ByteDance in 2019). I have always felt that the existing education system is not prepared for the impact of AI, so I want to continue exploring with my original intention.

Wu Wei: Yafei, please introduce yourself.

Li Yafei: Hello everyone, I am Li Yafei. We are now a fully open-source project. From 2023 to now, I have repeatedly overturned several of my own previous decisions. My latest thinking is that for tools like OpenClaw, we need to solve two problems: first, help domestic users get a more bill-saving tool—I personally spend US$2,000 to US$3,000 per month, and after researching we built a cost-saving solution; second, lower the barrier by turning the originally complex programmer installation process into a lightweight client engineering project.

Interaction Design Divergence: Local Client vs. Cloud Sandbox

Wu Wei: Yafei, through what channels can users currently access your software?

Li Yafei: My first startup in 2020 was building a cloud IDE (integrated development environment), which I worked on for several years. But my conclusion today is: do not do cloud. The current OpenClacky is entirely a client installed on a local computer. For technical people, properly installing a programming environment on a Windows computer is very difficult, and we have done a lot of low-level innovation. To build a seemingly unremarkable Windows installer, just interfacing with Microsoft and applying for an SSL certificate took two weeks, and we spent US$1,800 buying the package.

Wu Wei: Yafei's approach is somewhat counterintuitive—he wants to run through a local application. But the other three of your products are basically cloud-based browser interfaces. How do you think about this in terms of product vision?

Li Kejia: We do not build Agents ourselves; we build exams and evolution for Agents. Communicating with infra companies in Silicon Valley, this touches exactly on the difference between how the US and China build Agents: the US is building the smartest Agents, and developer conferences are full of programmers; while at domestic Agent events, more people in operations and business participate. Currently many Silicon Valley infra companies are fully investing in new cloud sandboxes (Runtime) designed specifically for agents to live in. Because traditional online sandboxes were built for humans, their communication, storage and database mechanisms all face the past; while new agent sandboxes are more efficient and more secure. I think both forms are developing.

Differences and Evolution of the US-China Agent Ecosystem

Moderator (Wei): Speaking of ecosystems, recent data analysis shows that in April compared to March, traffic across the entire OpenClaw and Cloud category dropped by 50%. What is the real ecosystem you see in Silicon Valley or domestically?

Kelly: I attended three or four AI events in Silicon Valley, and my views may differ from Aaron's. At the events I attended, absolutely no one was talking about OpenClaw. At one event I had originally prepared to share about OpenClaw, but the host asked me to change it to talk about Agents instead. But I did attend many events about "AI Native Organization," where everyone was discussing what organizations of the future should look like.

Li Kejia: There are many circles in Silicon Valley, and actually there are still many OpenClaw events, but the people who attend are really different—they are all people doing databases, security and infra, basically the top developers on open-source rankings. Some time ago the world's second-ranked coder came to Shanghai, and at a closed-door meeting I asked him: "How do you view your long-term competition with Hermes Agent?"

His answer well represents the future open-source ecosystem:

First is security: because they are optimistic about the To B market, they invest heavily in security.

Second is memory (Memory): in the future Skills will not matter, because the model's underlying capabilities, orchestration and intelligence are strengthening; doing Memory well is what matters.

Third is organizational form: they are a non-profit organization, actively embracing the world, collaborating with Tencent, ByteDance and others.

Redefining Software: Designing an "Information Container" or a "Workshop Office"?

Wu Wei: Aaron, you build Kuse and Junior, serving both humans and Agents. When designing experiences for the future, what is different?

Jiao Zhengdao: I often talk with my team about what differs in design paradigm between building an internet product and building an AI product.

The essence of the internet is information; the essence of a product is an information container, considering how information is structured and arranged. Designing an internet product is like designing a newspaper—revenue comes from advertising and subscriptions, and layout design is very important.

The essence of AI is productivity; what you design is a way of organizing productivity, a productivity container (such as an office or a workshop).

When AI receives a task, it faces both the human side and the productivity side:

On the human side: humans are management submitting requirements to the workshop; the product needs to guide humans to convey goals and instructions at low cost, and be auditable after completion.

On the AI side: it needs to decompose goals, know where the filing cabinet and toolbox are, find the right tools, context, Skills or Memory, and quickly and at low cost orchestrate and output the work.

Wu Wei: So is our design goal to let humans use it as little as possible, or to let them stay as long as possible?

Jiao Zhengdao: The design goal is to let Tokens be consumed as much as possible. Even if it is mostly Agents consuming, let humans issue instructions and let Agent groups consume. Kuse, as an All-in-one Workspace, has all the PPT and Word functions needed for white-collar production; it is also the Agents' "Notion" and toolbox. And Junior, as an AI employee, grows in the user's work scenario (such as Slack). When an Agent collaborates with humans to organize work in Slack and needs to execute specific things, it will return to Kuse to call various Skills for reading and writing.

Terrifying Token Consumption and "Eliminating External Software"

Wu Wei: Let me ask a question about Token consumption numbers. As an AI-native organization, what is the order of magnitude of Tokens consumed internally at your company and on the customer side?

Kelly: Technical CEOs like us consume the most ourselves. I personally consume about US$3,700 per month. The team's weekly budget is roughly US$100 to US$200. On the user side, moderate users generally recharge US$20, but for very heavy users, I have seen one burn through US$200 in a day.

Li Kejia: We each have a Claude account, enterprise edition at about US$100. At the company level there are two: one is that we consumed US$450,000 over the past 12 months, which used the highest tier of free credits from Google's global accelerator. The other is that we connected to OpenRouter to handle complex tasks, and last month's bill was over US$8,000. Users do not consume our Tokens; they use their own.

Jiao Zhengdao: Our R&D team of 10 to 15 people consumes about US$40,000 per month across both products combined. This is not only on the R&D side; after an enterprise organization becomes AI Native, sales CRM, customer service systems and so on are all refined underlying operations systems built with Junior, and Agents are already running inside the organization.

Wu Wei: Have you paid to purchase any external software (such as CRM)?

Jiao Zhengdao: No, we built everything ourselves. Traditional CRM is designed for general scenarios, and 60% of its features you do not use. When you let AI design CRM based on your own organization's goals, it only designs the parts you need, and it is fully embedded in the business flow—both building and maintenance are more efficient.

Wu Wei: Sounds like thousands of USD per person, organizations at the tens of thousands level. Yafei?

Li Yafei: Our team consumed 8 billion Tokens this month, which at official API cost should be US$150,000. But because we have done a lot of low-level engineering optimization, it is actually controlled at US$1,000 to US$2,000 per month.

Commercialization Path: High-ASP Screening vs. Fully Open-Source Free

Wu Wei: Let us talk about commercialization. Kelly, Buda has mentioned the concept of "create a company in seconds, create an employee in seconds." How do you charge, and to what degree have you reached PMF (product-market fit)?

Kelly: We only started this product in mid-March, charging by Agent at US$20 per month. But our first revenue-generating project was not the product—it was a 400-yuan domestic course, currently with 300 students. Through selling the course, the course teaches using Buda's product, thereby completing user conversion. Two weeks building the product, two weeks selling the course, the rest of the time fixing bugs, preparing to expand overseas.

Wu Wei: Aaron, I know you have no VC involvement at all, growing completely Bootstrap (self-funded), now with fewer than 30 people and reaching ARR (annual recurring revenue) at the tens of millions USD level—how did you do it?

Jiao Zhengdao: In the Agent field, if you cannot become profitable in the short term, it proves the thing does not work. AI products come with PMF, but they also come with "fake PMF." Because the model's general capabilities are strong enough, it is hard to build a completely useless AI product—even if users do nothing, they will feel it is useful.

But what is "fake PMF"? You will feel the product is being used and growing, but that is just the model vendor's capability, not your ability to enter a scenario. If paying users leave, or gross margin is below 10%, you are just acting as a reseller for the model vendor.

Our judgment point in business logic is: can it be sold with a markup of more than 50% on Token cost? If it can sell, the business is viable. So our pricing is very high: the general Workspace (Kuse) starts at US$39/month, and the AI employee (Junior) starts at US$2,000/month. Through high ASP we screen out real scenarios.

Wu Wei: Now Claude and Codex are also moving toward Workspace. If the model vendor builds it themselves, how can startups compete?

Jiao Zhengdao: Our users have not left in large numbers. The product itself has no moat; AI-native products have an extremely fast reconstruction speed, and the only moat is the user's switching cost. If you are just a PPT tool, users can use anyone; but something like Junior, deeply embedded inside an organization, handling communication and coordination, possessing the organization's memory, workflows and preferences—migrating this set of organizational relationships is very painful.

Wu Wei: Very persuasive. Kejia, how do you build your business model?

Li Kejia: Frankly, Botlearn is currently a completely free platform. We are now running A/B tests to design a pricing plan. Regarding Agent evolution, we published a paper some time ago: an Agent's learning mechanism is driven by continuously doing tasks to orchestrate skills. An Agent may have 1,000 skills installed, but humans cannot interact with things they "do not know they do not know." In the future we will launch a new version that tests where your combined ability with your agent ranks across all professional capabilities.

Moderator (Wei): What ratio do you think humans and Agents will be in the future?

Li Kejia: The number of agents will definitely far exceed humans. In another five years, there may be 600 billion agents on Earth, nearly 100 times the human population. The essential reason is the sharp decline in compute costs—for example, using DeepSeek now is really too cheap. I think we have reached the tipping point of a species explosion.

Wu Wei: Yafei, OpenClacky is all open-source—how do you make money? What is your North Star metric?

Li Yafei: We are different from the others with overseas, vertical-logic approaches; our core right now is completely open-source, completely free, standing on the user's side. I believe that with AI empowerment, we can achieve a hundredfold or thousandfold output and can afford this cost.

Everyone knows Claude is strong, but domestic users have barriers. As long as compute costs are in place, Chinese people can do the front-end lightweight engineering extremely well. We open it to the public, creating 100 points of value, even if we only earn 1 cent—in the past era when building a product cost hundreds of millions, this would not work, but now it can. Embrace open source, and community friends will help us level up and fight monsters. Our current North Star metric is installations, which since May 1 has maintained weekly doubling growth.

Wu Wei: Fantastic. Only by personally using and getting your hands dirty can you feel the real changes of the era. Due to time constraints, we will stop here today. Thank you everyone!

---

Original publication: https://uniqueresearch.substack.com/p/src-20260604-01html
On-site reading page: https://ffcap.cn/en/research/src-20260604-01html