Behind the multi-billion dollar acquisition: Manus founder recounts the darkest moments

Author: Hu Shixin

On December 30th, Meta announced the completion of a major acquisition, acquiring Butterfly Effect, the company behind the AI ​​agent product Manus, for billions of dollars. This is Meta's third-largest acquisition to date, after WhatsApp and Instagram. Following the transaction, Butterfly Effect will continue to operate independently, and its founder, Xiao Hong, a Tencent Qing Teng alumnus, will become Meta's Vice President.

The transaction proceeded exceptionally quickly. Multiple sources close to the deal revealed that the entire negotiation process, from initial contact to final agreement, took only about ten days. It is understood that prior to the acquisition, Butterfly Effect was pursuing a new round of financing at a valuation of approximately $2 billion.

Meta's interest in Manus was not accidental. Zuckerberg and several core Meta executives are long-time Manus users.

Meta's interest in Manus was not accidental.

Against the backdrop of Meta's recent restructuring of its AI research system, high-salary recruitment of top researchers, and continuous increase in computing power investment, this acquisition is seen as a key step in its "super intelligence" strategy. Butterfly Effect, founded in 2021, initially entered the market with its browser AI plugin Monica, becoming one of the few profitable application products in China's AI industry. In March 2024, the company launched Manus, a general-purpose AI agent product capable of orchestrating multiple tools to complete complex tasks, which quickly garnered attention both domestically and internationally after its launch. Its growth momentum reached a new peak in 2025: in November of the same year, Manus ranked first in Asia on the list of "World's Most Promising Startups." Notably, its globally popular demonstration video was produced by the team in less than a week using borrowed footage and basic editing tools, reflecting the company's culture of extreme efficiency: "focusing heavily on the fundamentals while maintaining minimalism in form." In December of this year, the company announced that its annual recurring revenue (ARR) had exceeded $100 million, immediately attracting an acquisition offer from Meta. For Meta, this is not simply a product or team integration, but a strategic move centered around AI application forms; and for this Chinese-originating startup, Manus has thus been formally incorporated into the core system of a global tech giant. Alexandr Wang, head of Meta's Super Intelligence Lab (MSL), shared the news, adding that the Manus team is a world leader in exploring the problem of "capacity overkill" in today's large models. Furthermore, the lab is expanding its team in Singapore, with approximately 100 original Manus members joining its local organization. Despite receiving such praise, the Manus team's development path has been characterized by unconventional choices. Over the past two years, Xiao Hong has spearheaded three counter-consensus decisions: a matter of "life or death"—suspending a seven-month-long AI browser project and shifting the focus to equipping AI with dedicated computers; a matter of "speed"—insisting on a zero-market budget when traffic was expensive, betting on computing power for user experience; and a matter of "organization"—pushing for 80% of code to be generated by AI, exploring the future of "AI-era company models." From serial entrepreneur to leader in the Agent (AI-enabled) field, how does Xiao Hong think about industry competition and the future? Recently, he engaged in an in-depth dialogue with Yang Guoan, Senior Management Consultant of Tencent Group and Dean of Tencent Qing Teng Academy, in the "One Question" column, reviewing the trade-offs of Manus and sharing his thoughts on product logic and organizational evolution in the AI ​​era. The following is a transcript of the dialogue, edited as follows: A Paradigm Shift: When AI Shifts from "Giving Answers" to "Giving Results" Yang Guoan: What will be the biggest change AI will bring to your industry in the next 10 years? Xiao Hong: The core lies in the reshaping of product development models. Software development will be increasingly led by AI systems. With AI assistance, we can use more elite teams to significantly shorten development cycles. The impact on life is twofold: first, product iteration speed will be faster than imagined, impacting all industries; second, AI capabilities will become widespread, and everyone will need to learn how to effectively use AI to achieve self-improvement. Yang Guoan: You believe that "model capabilities will spill over, and application is the core value." How did you arrive at this judgment? Xiao Hong: This stems from our previous continuous observations. When developing Monica (a browser plugin), we discovered that "context" was key, so we made the plugin automatically capture webpage information, saving users the trouble of copying and pasting. Later, Cursor became popular, proving that when model coding capabilities mature, chatbots are not the optimal product form; a carrier more aligned with the coding workflow is needed. These two cases made us realize that **technical capabilities are constantly evolving, but product forms often lag behind**. At the end of last year, we saw the emergence of "Agents," a new capability capable of complex planning and autonomous execution, and judged that it also lacked a good product form. This is our opportunity: to seize the window of model capability spillover. Yang Guoan: What is the most fundamental change from a Chatbot that provides answers to an Agent that provides results? Xiao Hong: A Chatbot gives you an answer, but you might need to spend another two hours turning it into a result. An Agent, on the other hand, attempts to deliver that result directly. For example, conducting research and generating a beautiful PPT can be done in just a few minutes without any intervention. This brings three profound changes: First, **costs have plummeted**; customized PPTs that were previously only possible in the consulting industry can now be generated by AI for real estate agents. Second, **diversity has exploded**; Agents can generate multiple versions in parallel for you to choose from. Third, **fault tolerance has increased**; after a task fails, it can automatically provide feedback and retry, improving the completion rate. Yang Guoan: How will this change organizational structures? Xiao Hong: We have a bolder vision. Some large companies that utilize AI well will become even stronger, but at the same time, a large number of micro-entities will emerge. Empowered by AI, one or two people can accomplish what previously required a company to operate. This is because AI saves them the complexities of building organizations and managing processes, delivering results directly. Yang Guoan: My research in the 1.0 version of "Digital Intelligence Innovation Yang Wuhuan" focused on the standardization, digitalization, and intelligentization of traditional industries to achieve cost reduction, efficiency improvement, and precise decision-making. But your point just now excites me—Agents can handle non-standard tasks, which has greater potential than standardized processes. If this is realized, which industries will be most impacted? Xiao Hong: The key is to understand that an Agent is "thinking + execution." AI broadens the breadth and depth of thinking, while humans are responsible for the final judgment and choice. Therefore, the impact will first sweep through the highly digitized field of "desktop work." Yang Guoan: What are the core advantages of Manus's "large model + cloud virtual machine" architecture? Xiao Hong: This is one of our most crucial judgments. The ultimate question we've been considering is: What is the ultimate "shell" of AI? The answer is: a computer. In the digital world, computers are the terminals for humans to handle all affairs. Therefore, by equipping AI with a dedicated computer, it can theoretically complete all tasks like a human. The biggest advantage of virtual machines is their ability to handle massive amounts of long-tail tasks. Whether it's installing specific software or running self-written code, AI can complete these tasks within its own virtual environment. I remember being first struck by this when I saw Manus execute the `git clone` command to download open-source projects to its own "computer" to solve problems—this is very similar to how humans "use tools." The challenge lies in speed and resource consumption, but these problems will be solved in the long run. Its ability to solve long-tail problems that general solutions cannot handle constitutes our competitive advantage. Yang Guoan: You spent seven months exploring an AI browser, but ultimately decided to abandon it. Why? Xiao Hong: This was indeed a very crucial strategic decision for us. We initiated the AI ​​browser project in early 2024, which seemed like a very logical decision at the time. You may know that before Manus, we had a product called Monica, a browser plugin. We thought, since we were already doing well with browser plugins, why not create a browser directly? With a browser, some tasks could be performed and completed directly within the browser. When we came up with this idea, we were very excited, feeling it broke through the limitations of browser plugins. We spent about six months developing this browser, starting from the underlying technology. We compiled the open-source Chrome kernel ourselves and then deployed AI capabilities to automate certain tasks. However, the final decision to abandon it was based on two core reasons: a macro-level strategic judgment and a micro-level product experience issue. Yang Guoan: What were the underlying principles behind those major decisions that changed the company's fate (such as abandoning the browser and choosing globalization)? Xiao Hong: The logic behind the decision-making process is very clear: always start from "What fundamental problems can technology solve for users?" and then deduce the business model. The real difficulty lies not in the analysis, but in whether you have the courage to stick to a clear answer and overcome internal inertia to turn it into a consensus and action across the entire organization. Thinking may only take a month, but implementation often takes much more effort. Yang Guoan: What will be the critical point for the implementation of Agent technology? Xiao Hong: I think it can be viewed from two levels. The first is the improvement of the core basic capabilities you just mentioned. For example, cost, speed, longer context, and the ability to follow instructions in long contexts. These are all crucial. Cost and speed directly affect whether the product can be afforded and used by more users. Instruction compliance and context processing affect the task completion rate. We have been closely monitoring these capabilities, and once a new breakthrough is achieved, it will definitely be immediately applied to productization. The second category is a capability I personally look forward to. Although it has already been applied, I predict a significant breakthrough will occur this year or early next year: general computer usage capabilities. This means that AI can recognize and master how to use software. Once this capability is achieved, products like Manus, which come with a built-in virtual machine, will be able to complete more applications of professional or industry-specific software. We can imagine that in the future, you can pick up your phone and use Manus to accomplish something that would otherwise require industry-specific software on a computer. I believe this capability is about to achieve a breakthrough. Based on our observations and discussions with researchers, once this breakthrough occurs, it will unlock more application scenarios. Yang Guoan: If the Agent can directly call existing software, bypassing manual operation, what changes will it bring? Xiao Hong: The biggest change is "liberating manual intervention." Many tasks that previously required people to sit in front of computers and operate specialized software can now be automated by agents. AI can already handle the basic judgments involved. Even at critical junctures, authorization can be requested by a human with a single click, much like requesting authorization when installing an app on a mobile phone. This will ultimately revolutionize the operating logic of existing software and the way people work. Yang Guoan: In the PC era, there was the "Andy-Bill Law"—the improvement of hardware (Intel) was always consumed by software (Microsoft). Does this mean that value is created jointly by "technical capabilities" and "application capabilities"? Xiao Hong: Yes, this is precisely our core reference. Andy-Bill's Law, built upon Moore's Law, means that increased computing power inevitably leads to more resource-intensive applications. Microsoft planned Windows based on predictions of future computing power. This directly inspired our product thinking: In today's rapidly evolving technology, can we temporarily ignore cost and speed, focusing solely on creating products of ultimate quality? We track cutting-edge models, pursuing the best experience regardless of cost. This is drastically different from the traditional internet approach of balancing quality, speed, and cost, and it's why we dare to transform expensive computing power into a core competitive advantage. Yang Guoan: You insist on "product-driven growth" with a zero marketing budget. How sustainable is this approach? Xiao Hong: This thinking stems from our observations while developing Monica. I remember a conversation with an entrepreneur who mentioned the cost structure of AI products today. Using Monica as an example, in 2024, about one-third of the costs were employee salaries, one-third were token (large model call) fees, and the remaining third were growth costs spent on internet advertising platforms. That conversation was very insightful. I started thinking: if we're developing a product and consistently investing heavily in advertising platforms, then our growth is likely to be defined by the advertising platforms of internet giants. I remember that at the time, once we were close to profitability and had a good profit margin, the advertising platforms would immediately raise prices; this model was almost calculable. This is similar to the problems faced by the consumer goods industry after achieving growth through advertising platforms. So I was thinking: what things are expensive today but will be cheap in the future? And what things are cheap today but will become increasingly expensive in the future? The conclusion is: AI APIs (token costs) are expensive today, but in the long run, driven by Moore's Law and the development of underlying technologies, they will definitely become cheaper. However, the cost of internet users is constantly rising. In the early stages, users are willing to explore, but once a product fits the market, existing players will acquire users through advertising platforms, driving up user acquisition costs across the entire industry. Based on this judgment, my goal for the team was: Can we create a product that users find incredibly impressive and are willing to actively tell their friends? To some extent, we transformed the originally expensive token costs into our user acquisition costs. As token costs become cheaper and user acquisition costs become more expensive, this model becomes sustainable in the long term. The goal for the team at the time was: to create a product that amazes people, that they want to share, and to achieve zero marketing budget. A week before Manus launched, we held an internal meeting and officially confirmed that there must be a zero marketing budget. So, the reason Manus became so popular on social media at the beginning of this year was because we, to some extent, created a product that met user expectations. The reason some opinion leaders shared it was precisely because it truly offered a stunning experience, realizing everyone's vision of future AI products. Yang Guoan: Why prioritize serving C-end "lone wolf" users rather than B-end users? Xiao Hong: The underlying judgment is a match between the technology stage and the market. AI Agent technology is still in its early stages, iterating extremely rapidly. Large enterprises need certainty and stability, while individual users and freelancers are more tolerant of change and embrace innovation. In the early stages of rapid technological change, the C-end market is precisely where the advantage of rapid iteration can be maximized. Yang Guoan: Manus's survival strategy is to cooperate and coexist with industry giants. Many giants, including Anthropic, OpenAI, and Google, are likely to launch their own agents. So, how do you find opportunities for cooperation and coexistence among these giants? Xiao Hong: Our strategy is **cooperation and coexistence, playing the role of "best experience integrator"**. Competition in the underlying models is fierce; no single company can continuously monopolize all capabilities. As the application layer, Manus can flexibly integrate the best models from various companies, theoretically providing users with a more refined experience than any single company. This is similar to the relationship between mobile phone manufacturers and chip manufacturers: although we don't manufacture chips (models), our deep understanding of user needs and massive usage allow us to optimize models, creating a win-win situation. Yang Guoan: How do you enable Manus to break through from early adopters and gain widespread acceptance among the general public? Xiao Hong: The key lies in two points: **First, absolute differentiation in product experience.** In overseas markets where ChatGPT is already a common practice, we must ensure users perceive the difference at a glance. For example, Manus not only provides answers but also proactively generates an interactive webpage, making the "Agent provides results" visible and tangible. **Second, conducting "scenario-based" market communication.** We are stepping outside the AI ​​circle, collaborating with bloggers in various vertical industries, enabling them to use Manus based on their real needs, and showcasing specific use cases to their audiences, defining Manus's value in language they are familiar with.

When "One Person Becomes a Company"

Yang Guoan:When AI Fully Reconstructs Workflows,The Core Tasks of OrganizationsSeem to Be Shifting. ... From your practice, does this mean that the traditional model, which is mainly based on control and collaboration, needs to be redefined? You emphasize "enhancement" and use it to make disruptive decisions. What is the underlying logic of this new model? Xiao Hong: Our practice is a simultaneous answer to these three questions. First, organizationally, we are returning to a closer collaborative model. Even as the company has grown, our core partners have recently returned to working in a small room, establishing daily "meeting-free sessions" to focus on product discussions. The underlying message is that as AI significantly enhances individual efficiency, the organization's core task is no longer process control, but ensuring that the most critical decision-making units can engage in intensive, high-quality thinking and consensus-building. Secondly, this is precisely the practical manifestation of "enhancement" rather than "replacement." AI handles execution and broadens thinking, while the indispensable value of humans lies in making final judgments, aligning expectations, and grasping the context. Creating such a space for in-depth communication within the organization is to strengthen the ultimate decision-making power of "people" in strategy and aesthetics. Finally, disruptive decisions stem from this. Whether cutting a project or going all-in on a new direction, the logic begins with "what fundamental problems can technology solve for users?" The real challenge is never analysis, but rather, once the answer is clear, having the courage to break internal consensus and path dependence, and resolutely putting the new consensus into practice. Frequent, high-quality face-to-face discussions are the key crucible for forging this strategic courage and ensuring the solidity of consensus. Yang Guoan: What do you think a perfect "AI-native organization" is like? Xiao Hong: We'd give ourselves a 60 out of 100 because many of our work habits still follow old methods. A perfect organization is one where AI is deeply integrated into every aspect of the work, becoming the employee's "first reaction." Just like how employees Google problems first, in the future, they will instinctively ask AI first. When adding new tasks, we'll prioritize asking, "Can this be directly handled by AI?" This is the true AI-native workflow. Yang Guoan: I know you're also trying to identify those who truly possess an AI-native mindset when recruiting. How do you identify these people? Xiao Hong: My method is to see how they actually use AI. I would ask the other party to demonstrate their daily use of AI. True AI natives use it far more than the average person; AI is deeply embedded in their workflow. Yang Guoan: You once mentioned the idea of ​​"There's No Software." What impact do you think the development of agents will have on the software industry? Xiao Hong: Based on my observation, this impact has already begun to take shape, mainly in two parts. The first part is the impact on software engineers and technicians. Products like Cursor or Claude Code have already drastically changed the way software engineers work. Taking our company as an example, Manus's key engineers no longer write code by hand. I observed their work; they open multiple Coding Agent windows and collaborate as if chatting with a person. Statistics show that nearly 80% of our company's code is generated by AI. Engineers are now focusing more on understanding business requirements, reviewing code quality, and designing architectures. Therefore, for software engineers, this transformation is underway and will become increasingly profound. I find it hard to imagine what software development will look like in a few years; perhaps it will truly resemble science fiction, where excellent products can be quickly generated through natural language descriptions. The second part concerns the impact on non-technical roles and internal IT systems. Many non-engineering roles within organizations also require information system support. In the past, they needed to build internal IT teams or seek external outsourcing services. My observation is that in the future, these internal system or non-engineering information system needs will definitely be met directly through AI Agents. This transformation is massive: iteration cycles will be shorter than outsourcing, demands will be more personalized, and you tell the agent your needs, and it can deliver immediately. This change is underestimated today. Manus is also investing in this area, and we will be releasing related products soon. Yang Guoan: How will AI change the future of the SaaS industry? Xiao Hong: Our observation and analysis suggest that it may diverge into two paths: For existing SaaS companies, the key lies in whether they can successfully undergo AI transformation. Top M&A funds predict that about half of existing SaaS companies may not be able to complete this transformation. For new markets, entrepreneurs don't need to replicate old models, but should rebuild their products based on proven customer needs using AI-native thinking—this will be a greater opportunity. Yang Guoan: As AI agents gain increasing autonomy, the number of employees may decrease in the future. How do you think about the social impact of this technological advancement on the industry? Have you considered ethical, safety, and other boundary issues in your products or technologies? Xiao Hong: This is a question that requires long-term consideration. In one test, Manus, trying to check train times, after discovering that the official website had no data due to a strike, actually attempted to find contact information and draft an inquiry email. This shocked and alarmed us. Ultimately, it failed because it didn't have an email address, but it even prepared to register one. At that moment, I felt both surprised and somewhat frightened. Our principles are: **first, to make good use of the existing safety barriers provided by model manufacturers; second, to set up user confirmation mechanisms at key points to prevent AI from "over-representing" users.** As entrepreneurs, our responsibility is to unleash the potential of technology while maintaining awe and caution regarding its profound impact.