Leaving Tencent to start a new business, a 10-person AI company has achieved annual revenues of tens of millions of yuan.

Text by | Zhou Xinyu

Edited by | Su Jianxun

At the AI company "Aha Moment", monitoring data is the prerequisite for all business:

In the Beijing office with only 10 people, there is a display screen. When employees look up, they can see several groups of real-time updated data: subscription success rate, end-to-end conversion rate, number of weekly subscribed users, and number of daily subscribed users.

At the bottom of the data dashboard, there are clocks for four regions: San Francisco, London, Paris, and Beijing, corresponding to the company's most concerned markets: North America, Europe, and China.

△ Aha Moment's data dashboard. Image source: Provided by the interviewee

Refining data reporting and analysis to the minute level is the daily work rhythm of Aha Moment.

CEO Kang Hongwen described to "Intelligent Emergence" that the team may look at data more carefully than those in the secondary market look at the stock market: "There can be optimization only with data. If you don't measure, you'll never know what needs to be optimized."

"Measuring data" has brought rapid commercial progress: Aha Moment achieved break-even within three months of its establishment; by the end of 2024, the company's ARR (Annual Recurring Revenue) reached nearly ten million US dollars, achieving profitability.

The proof of self-sufficiency also brought financing to Kang Hongwen. "Intelligent Emergence" learned that Aha Moment recently completed an angel round of financing worth tens of millions of RMB. The investors include the US dollar fund Hi2 Capital, as well as angel investors Mike Green, Yipeng Li, etc.

As a veteran in AI technology, Kang Hongwen has accumulated over 20 years of work experience, including stints at large companies, startups, and universities.

In 2004, he conducted research on computer vision (CV) technology at the Microsoft Research Asia, under the guidance of Shen Xiangyang, Tang Xiaoou, and Hua Xiansheng. Later, he obtained a doctorate from Carnegie Mellon University in the United States. After two AI startup experiences, Kang Hongwen became a senior director at Tencent's PCG (Platform and Content Group) following the acquisition of his company.

△ Kang Hongwen. Image source: Provided by the interviewee

After the popularity of ChatGPT, Kang Hongwen saw the new wave of AI brought by large models. In October 2023, he left Tencent and plunged into AI entrepreneurship again.

Aha Moment, originally a psychological term describing the moment of sudden insight and enlightenment, was chosen by Kang Hongwen as the company's name: AhaMomentAI.com.

In the wave of large models, the Aha Moment he felt was finding the intersection between "the explosion of edge computing power" and "the reduction of model size" — Put simply, it means deploying a model with a small enough size and high enough performance on the terminal.

"There will be an intersection point between the two trends. Reaching this point means that a model with quite good performance can run on a single graphics card," he explained to "Intelligent Emergence". "This will definitely bring about innovative product forms and unlock new product experiences and business models."

Today, Kang Hongwen's initial prediction has been confirmed in the industry, such as the rise of lightweight edge models and the new wave of AI PCs.

But in Kang Hongwen's view, the ideal form of large model implementation is to provide users with an app; an app encapsulating a large model can also be flexibly installed on any terminal hardware such as PCs, mobile phones, and smart devices.

In Kang Hongwen's opinion, an AI PC equipped with an edge model can provide limited services: "The AI PC is from the perspective of device manufacturers. An AI PC will come with some underlying and basic general capabilities, but it still needs to run apps to serve the specific scenarios and needs of professional users."

He showed the product demo to "Intelligent Emergence". When 3T of text, image, audio, and video data was dragged into an app with more than a dozen models deployed, in response to the request of "summarize the relevant content about Lady Gaga", the app only took less than a second — this processing speed is much faster than that of large models deployed in the cloud.

Today's users are sensitive to the response speed of AI. A typical example is that whether it's the DeepSeek App or the recently popular Agent product Manus, "slow server response" is a much-criticized issue.

In optimizing the AI response speed, the opportunity Aha Moment found is to provide C-end users with an app matrix with text, voice, and video creation capabilities. And these apps have improved response efficiency because large models are deployed locally.

However, it's no easy feat to fit a large model with dozens or hundreds of gigabytes of video memory into an app of just a few gigabytes. Kang Hongwen also admitted to "Intelligent Emergence": "It may take two or three years for the technology to mature. I believe in this trend and am also making a big bet."

However, company management and business advancement require as much certainty as possible. Back in 2023, when the entire industry was still a believer in the Scaling Law, on the first day of the company's establishment, Kang Hongwen made a decision: Focus on product development and commercialization first.

Within less than a month of its establishment, the team tested several products to evaluate their commercialization effects. By early December 2023, the company selected the products with the potential for scale-up and officially launched them into the market. Throughout 2024, the team's focus was on achieving large-scale profitability.

However, Kang Hongwen is very secretive about the names and forms of these products. "I don't want the products to be overexposed," he said. His reason is: "The current products are all in a transitional state. I hope that what we actively release is a mature product that realizes our vision."

"Making money isn't really that difficult." After consecutive entrepreneurial experiences at home and abroad and the commercialization test at Tencent, Kang Hongwen said a somewhat boastful statement in the current context.

But the original intention of this statement is not to show off, but to demonstrate the hard work required behind it: Precise calculation of PMF (Product-Market Fit).

Aha Moment has only 10 employees, but each of them is connected to a data alarm. Once the business data shows anomalies within a certain period, even in the middle of the night, the corresponding person in charge will be woken up by the phone and will analyze and optimize the product.

"This is the painful period of entrepreneurship. You must maximize commercialization in the short term," Kang Hongwen told "Intelligent Emergence".

The data-driven methodology comes from his former employer, Tencent.

From Tencent, Kang Hongwen directly witnessed how a data-driven team works efficiently: For the core internal products, there are hundreds of data indicators. Without the leader's reminder, employees will automatically optimize the solutions when they see the data feedback.

But in reality, finding PMF is no easy task for most entrepreneurs with a technical background.

Entrepreneurs with a technical background often focus too much on technology and neglect the product. By the time they find the right application for their technology, they are often left behind and miss the best opportunity for product implementation and commercialization.

During the nearly two-hour conversation, a term that Kang Hongwen emphasized was "breaking away from technical inertia".

"At CMU, I was also very focused on technology and always wanted to make innovations in algorithms." It was his then tutor, the famous scientist Takeo Kanade in the CV field, who "poured cold water" on him.

At that time, he often told his students: "Any student who comes to CMU has no problem with technology. At this time, you should instead think about what problems you want to solve with technology."

Kang Hongwen recommended the book "Think Like an Amateur, Practice Like an Expert" written by Takeo Kanade to many people, including his employees: "Thinking like an amateur means finding the 'nail', and then thinking about how to use engineering capabilities to solve the problem, practicing like an expert."

Another term that Kang Hongwen mentioned multiple times was "timing".

To some extent, he thinks he is a technical person favored by timing. At the beginning of each cutting-edge technology, he was lucky enough to be at the forefront.

For example, at the Microsoft Research Asia, he was lucky to be one of the earliest engineers in China to develop on GPUs. At that time, Kang Hongwen's job was to integrate AI vision algorithms into Microsoft's game terminal, the Xbox.

Another example is that in 2012, Kang Hongwen's doctoral project at CMU was the current hot topic: smart glasses. He summarized the first half of his career: He has worked on almost all hardware terminal forms, from the cloud, PCs to mobile devices.

Later, when he started his own business, "timing" meant the window for finding the market and financing, as well as the first-mover advantage.

In 2017, when he was starting a business to develop an AI video editing application, Kang Hongwen was asked in a media interview "how to deal with the competition from BAT". His answer was: Start early enough. He told "Intelligent Emergence": "Looking back, this answer was so right! Less than a year later, short videos became extremely popular, and all the big companies you could think of entered the market!"

This experience makes him stay calm about the current hottest "concepts". "You can't aim at the target's current position. You have to aim at where it will be in the future." He described entrepreneurship as "shooting at a moving target". "Entrepreneurship is not about doing what's current; it's about doing what's in the future. The time difference will turn into your barrier in the future."

He had two judgments about the popularity of large models in 2023:

One was based on his grasp of technology. "Large models are still in their early stages, and there will definitely be several major technological paradigm iterations in the future."

The other was based on his understanding of the market. "By 2024 and 2025, both capital and the market will calm down and will surely return to rationality to evaluate your product and commercialization. What we need to do in 2023 is to prepare for 2025 and get ahead of other companies in commercialization."

As a technical person, Kang Hongwen also has a tinge of regret for not being directly involved in the wave of model pre-training in 2023. During the conversation, he mentioned Liang Wenfeng, "He is still a person writing code on the front line."

On the day "Intelligent Emergence" met him, there was still a dozens-of-pages PPT on DeepSeek technology analysis on Kang Hongwen's display screen.

But as a CEO, he needs to make trade-offs and focus enough. Now, the initial success in commercialization has bought some time for the ambition of "encapsulating large models into apps".

Kang Hongwen told "Intelligent Emergence": "If the product and commercialization achieve breakthroughs now, then we can be patient and wait until the technical route stabilizes. By then, we can directly use the most mature technical route."

Welcome to communicate!

Welcome to follow!