In the 130 days since its launch, Manus has suddenly evolved from a top-tier AI Agent with “internal beta code speculation to 100,000” to a storm center of “headquarters relocation and layoffs of 80%”; When the quadruple pressure of capital, compliance, computing power and product boundaries is superimposed, it is not running away, but sounding the alarm for the entire track.
Recently, the spotlight of the market has once again hit Manus, but this time the keyword is not the “hard to find” three months ago, but the current “running away” suspicion.
The AI Agent company, which has become popular around the world with a demonstration video, recently confirmed that it has made drastic optimizations to its domestic business team after officially confirming its headquarters move to Singapore in June – of the 120 employees, only more than 40 core technical personnel were retained to move to Singapore, and the rest were laid off.
For a time, the official social media was emptied, the cooperation with Alibaba was deleted, and the Chinese version of the product could not be used…… All kinds of signs have made speculation that “Manus ran away” rampant.
In stark contrast, its account on the overseas social platform X is updated, with the latest update on July 10, introducing information about a hiking event scheduled for San Francisco on the 13th.
Although a person familiar with the matter clarified to Tiger Sniff that “the company has not run away, the business is normal”, and Xiao Hong, founder and CEO of Manus, also posted on social platforms to express the vision of “doing a good job in global products in the new environment”, the market’s doubts have not dissipated.
From the timeline, the whole process of Manus from the launch of the product on March 6 to the announcement of financing, headquarters relocation, and large-scale layoffs is about 130 days. The story of this star company is more like a microcosm of the entire AI Agent track between ideals and reality.
To understand this “migration” of Manus, we need to start from two key clues: one is the investment from the old VC Benchmark in the United States in April and its chain reaction, and the other is the commercialization pressure faced by Manus’ products themselves and the agent track.
01 Manus “ran away”, rational or helpless?
It started in late April this year. Manus’ parent company, Butterfly Effect, has completed a blockbuster round of financing: a $75 million (about 550 million yuan) Series B financing led by Benchmark, a well-known venture capital firm in Silicon Valley, with a valuation jumping to $500 million. The funds raised from the Series B financing will be mainly used to expand international markets such as the United States, Japan and the Middle East.
What does a product manager need to do?
In the process of a product from scratch, it is not easy to do a good job in the role of product manager, in addition to the well-known writing requirements, writing requirements, writing requirements, there are many things to do. The product manager is not what you think, but will only ask you for trouble, make a request:
View details >
It stands to reason that this should be good news to celebrate, after all, Benchmark is the top VC in Silicon Valley, and has invested in star companies such as Uber, Twitter, and Instagram.
But getting the money also means that Manus has to be more careful about compliance. Under a new U.S. rule that came into effect in January this year, some investments involving cross-border technology companies are subject to strict scrutiny. As an AI company, Manus is naturally within the scope of attention.
An investor who pays attention to overseas projects said that for U.S. investment institutions like Benchmark, subsequent compliance arrangements will be taken into account when making investment decisions, “This is not just a simple investment negotiation, but also a packaging agreement that includes investment, migration, and structural adjustment.” ”
In fact, this kind of operation has been reflected in Benchmark’s previous investment cases. AI video company HeyGen deregistered its Chinese company in 2023, received financing from Benchmark in 2024, and moved its headquarters from Shenzhen to Los Angeles after the financing was completed.
However, it may not be comprehensive to attribute Manus’ migration entirely to capital pressures. A key detail is that the Singaporean entity of Manus, a company named “Butterfly Effect”, was registered in Singapore as early as August 2023 and is wholly controlled by the Cayman Islands entity of the same name.
According to the analysis of the above-mentioned investors, before the video became popular in March this year, and even before the start of investment negotiations with Benchmark, Manus had the possibility of preparing for overseas layout.
Manus needs a global stage, Benchmark needs a compliant investment target, and the cooperation between entrepreneurs and investors is more like a high degree of mutual interest.
Liu Yujia, investment partner of Zhengjing Capital, analyzed that in addition to capital and regulation, practical considerations of computing power and data are also the key forces driving this “migration”.
In recent years, due to geopolitical influences, the access to some high-performance AI chips in some markets has been limited. As one of Asia’s GPU and computing power hubs, Singapore is an ideal location for AI companies like Manus with resource flexibility and international connectivity advantages.
A person close to Manus revealed that Manus has previously encountered a shortage of high-end computing resources, resulting in delays in the iteration of agent products. In this case, migrating to Singapore has the potential to obtain computing resources more efficiently.
Data compliance is also an important consideration. Liu Yujia emphasized that as countries have increasingly strict requirements for data security and cross-border transmission, how to conduct international business under the premise of compliance has become a problem for all AI companies. Singapore’s relatively relaxed data compliance framework provides more room for Manus to serve the global market.
Of course, this also means choice. The above-mentioned people close to Manus said that it is difficult to take into account both the Chinese and American markets with the current volume of Manus. Betting on overseas is the current strategic decision.
From the beginning, it mainly calls the technical route of overseas large models and has a relatively low dependence on domestic large models, which to some extent also lays the groundwork for today’s “going overseas”.
02 B-side to the sea: How is the product and users?
The story at the capital level will eventually return to the product itself. After all, any star company has to rely on products and business models to speak.
Let’s briefly talk about what Manus is. If you haven’t already, you can understand it as a competent assistant – not the kind of robot that only chats, but has the advantage of disassembling and calling various tools to perform complex tasks, and finally gives you the results. The most ingenious thing is its interface design: on the left is the familiar chat box, and on the right is an execution window. This design is particularly intuitive for ordinary users, making you feel like AI is really “operating the computer” for you.
In March this year, Manus was born with a “general AI agent” product, and the internal beta code was once speculated to 100,000 yuan a piece, when Xsignal (singularity factor) data showed that its peak monthly active users (MAU) exceeded 20 million.
However, when the novelty faded, the market began to return to rationality. After opening registration in May, its monthly active data showed that it had plummeted to about 10 million (Xsignal), and the user retention rate faced severe challenges.
This gap in experience stems from the “capability boundary” that is prevalent in AI Agent products at this stage.
The biggest changes at Manus over the past three months are the launch of a free Chat mode and a template library called Playbook in June, which allows users to apply ready-made workflows directly. Manus has also optimized the architecture – it is said to be 5 times cheaper and 2 times faster.
In order to feel its current true capabilities, “Prime One” has designed two tasks for the basic version of Manus: one is close to the life scenes of ordinary people, and the other is biased towards professional fields, which can just see the performance of Manus in different scenarios.
The first task was to have Manus compare four wall-mounted washing machines for consumers. This is a typical consumer decision-making scenario that tests AI’s comprehensive capabilities: it needs to capture price data from multiple e-commerce platforms, organize product parameters, and analyze user reviews.
The first step in Manus is to generate a detailed list of plans and then call various tools to execute them step by step. The whole process took 15 minutes, and the efficiency was really good. But what was the result?
Let’s talk about the highlights first: The PPT generated by Manus looks quite beautiful, not only summarizing user evaluation data, but also giving purchase suggestions classified by family scenes. And when encountering platform reverse crawling mechanisms, such as JD.com’s slider verification, it will actively seek human help, which is much more “smart” than skipping it directly.
But the reliability of the underlying data discounted the report: the task clearly required a comparison of prices on the three platforms, and in the end it only gave data from JD.com; The accuracy of its price capture is not high, and the price gap of some products is still relatively large, or because of platforms like JD.com, a link often corresponds to multiple different models of products, which makes accurate price comparison difficult. In addition, the capture of core parameters is not comprehensive enough.
The second task was highly professional: Manus was asked to do detailed Excel valuation modeling for Microsoft as a secondary market analyst.
Also in 15 minutes, Manus handed over a comprehensive model that included various methods such as DCF valuation, comparable company analysis, and sensitivity analysis. After evaluating this model, an investor in the AI field said that the final output model follows the professional standards of investment banks and data capture is relatively accurate, but AI still cannot replace human decision-making in links that require “human judgment”, such as risk pricing.
More specifically, the model’s access to Microsoft’s β (a measure of stock relative to market volatility) is still stuck at a static data of 0.9 in 2023, while the data for Bloomberg terminals in July 2025 has already been updated to 1.05-1.10. This lack of sensitivity to real-time data directly affects the accuracy of valuation.
Speaking of experience, there is another episode worth mentioning during the test. When the free credits run out, Manus prompts you to pay for an upgrade. When I paid for the cheapest Basic version ($19/month), I failed twice in a row due to “account risk”, and finally succeeded the third time.
These two tests reflect the current situation of Manus and the current AI Agent: they are very good at execution, but they are inexperienced and sometimes even make some low-level mistakes. In other words, you can delegate some repetitive, standardized process work to it, but you still have to make key decisions that require accurate data and real-time judgment.
This gap reflects the boundary between AI tools and professionals. A number of technical people told “Prime One” that the tasks they give Manus will involve programming, such as writing a mini-game. However, Manus still has a high failure rate when handling complex tasks, especially when generating videos, often lagging, data crawling is often limited by the platform, and even refuses to create new tasks when the load is too high.
The product performance that needs to be improved has directly led to the dilemma of user growth. Guosen Securities emphasized in the research report that for AI Agent to become a new paradigm of human-machine collaboration, it needs to have three core capabilities: planning, tool use and memory. At present, Manus still has a long way to go in terms of reliability and accuracy in “tool use”.
03 The cost account behind the AI Agent
In addition to user growth, there is also the pressure of commercialization.
Manus’ business model is tiered subscriptions: $16 per month for the basic version, $33 per month for the Plus version, and $166 per month for the Pro version, starting at an annual subscription. But its model faces cost challenges.
To handle complex tasks, Manus needs to call APIs for multiple large models at the same time. According to a report by The Information on March 18, 2025, Manus’ products are currently limited by both server capacity and high operating costs.
The problem with this cost structure is even more obvious in comparison with competing products. Genspark, for example, claims to have achieved $36 million in annualized recurring revenue (ARR) in 45 days, and with zero ad investment, it achieved this with just a core team of 20 people.
This is largely due to the design of its business model. Genspark focuses on document analysis and knowledge refinement, which are relatively standardized and cost-controllable. At the same time, European and American users are generally more willing to pay, which provides Genspark with a better revenue base.
In contrast, Manus is positioned more broadly, trying to become a general-purpose AI agent, which expands the application scenarios and has more imagination space, but the user needs are very different, and the models and resources that need to be called are also different, which directly increases the difficulty of cost control.
Although moving to Singapore has solved some problems, Liu Yujia believes that the increase in fixed costs such as labor costs and office costs has undoubtedly put forward higher requirements for the business model of a startup like Manus.If domestic losses can be supported by relatively low operating costs, companies must find sustainable profit models faster in Singapore’s high-cost environment.
An analyst who focuses on the AI industry said that for pure to C AI Agent products, this cost pressure may force the team to consider more diversified business models, such as a hybrid model of to B and to C, or to transform into an enterprise-level market. “Pure consumer-grade AI agents may be destined to be a money-burning business, at least at this stage.”
Liu Yujia said, “It is reasonable to give Manus more time.” After all, it has only been a few months since the video became popular in March, and it is indeed a bit too harsh to require a product to achieve commercial success in such a short period of time.
Judging from the current development trend, Manus has chosen the path of improving products through rapid iteration and establishing barriers through deep integration.
Whether this goal can be achieved is too early to tell. But what is certain is that the challenges faced by Manus reflect the current situation of the entire AI Agent track.
From a technical point of view, most of the mainstream AI Agent products on the market (such as Adept, Lindsey, etc.) are in the position of “middlemen”: they do not do the lowest level of model training, nor do they directly target specific vertical scenarios, but are built on third-party large language models (such as GPT series, Anthropic’s Claude series), focusing on the development of “middle-layer” technologies such as scheduling, planning, and tool calling. Although this positioning has its advantages, it also faces “squeeze” from upstream and downstream.
Looking up, it is the “dimensionality reduction strike” of large model manufacturers such as OpenAI and Anthropic. They are constantly integrating agent capabilities into their foundation models, trying to provide end-to-end solutions directly to users.
Looking down, AI applications in various vertical fields are emerging. They can often provide more accurate and in-depth services than general agents in specific scenarios (such as Harvey AI in legal scenarios, BloombergGPT in financial scenarios, etc.).
In this competitive landscape, the Manus, who are in the “middle layer”, must find their own position. As the report of Huachuang Securities said, the core of the practical implementation of AI Agent in To C lies in whether it can truly move from “suggestion” to “delivery”.
The story of Manus is also the story of all Chinese AI companies in the wave of globalization. An investor who has been paying attention to the AI industry for a long time emphasized that the early adopter mentality of young users can indeed bring short-term user growth, but to achieve long-term commercial value, it is still necessary to return to the practicality of the product. In this multiple game of technology, capital and market, who can better balance innovation, business and user needs will have the last laugh.
Of course, we should not be overly pessimistic. As an emerging field, AI Agent still has a lot of room for exploration.
