Manus, standing on the eve of the explosive boom of AI applications
On October 8, Butterfly Effect, the parent company of AI firm Manus, announced the completion of a financing round of over 500 million US dollars. Boyu Capital and IDG Capital co-led the investment, while existing shareholders including Tencent, HSG and ZhenFund continued to participate. The company did not disclose its latest valuation, though prior reports stated that Manus was seeking financing at a 4 billion US dollar valuation.
It has been less than a year since Meta announced the acquisition of Manus. On December 29, 2025, Meta announced the acquisition of Manus in a deal reportedly worth over 2 billion US dollars. In April 2026, the acquisition was called off due to regulatory issues. After months of adjustment, Manus resumed independent operations and secured another large-scale financing.
Manus was launched on March 6, 2025. Nine months later, the company announced that its annual recurring revenue (ARR) exceeded 100 million US dollars, with around 105 employees. According to *The Information*, in June 2026, Manus' annualized revenue run rate reached approximately 500 million US dollars. The company does not independently train cutting-edge foundational large models, yet it has secured a multi-billion dollar capital valuation in less than two years.
Xiao Hong, the founder, has been engaged in application software entrepreneurship for more than a decade. From Yiban, a WeChat Official Account tool, Weiban, a WeCom customer management software, to Monica, an AI browser assistant, the team has continuously leveraged capabilities provided by large technology platforms to develop products. Manus follows a similar entrepreneurial approach, but the market it faces has already changed.
In 2025, US enterprise generative AI spending reached 37 billion US dollars, of which AI application spending was approximately 19 billion US dollars. Enterprises began to procure AI software on a large scale, and agents gradually gained the ability to perform complex tasks. Manus drew capital attention during a period when the AI application market was growing rapidly and the commercialization of general agents had not been fully rolled out.
I. The Business Logic of Application Companies
In 2016, Xiao Hong and several entrepreneurial partners from Huazhong University of Science and Technology were preparing to abandon their startup projects. The group had already received job offers from internet companies in Beijing, but before leaving, they participated in a hackathon and unexpectedly secured a 1 million RMB investment from ZhenFund.
The team stayed in Wuhan to develop Yiban, the WeChat Official Account tool. At that time, products such as 135 Editor and Xiumi had already entered the market. Yiban adopted the form of a browser plugin, embedding editing functions into the Official Account backend to help operators reduce page switching and repetitive operations.
In December 2019, WeCom released version 3.0, further opening up customer connection capabilities. Xiao Hong's team subsequently developed the Weiban Assistant, providing customer management and marketing services for enterprises. In December 2021, Minglue Technology announced a strategic investment in Weiban. According to the data disclosed at that time, Weiban had served a total of 140,000 enterprises, helping them connect with more than 100 million WeChat customers.
In 2022, Xiao Hong started his new business again. The team originally planned to develop scientific research software, but the language processing capabilities demonstrated by GPT-3 changed their original plan. On November 20, the team initiated the development of Monica. Ten days later, OpenAI released ChatGPT.
Monica provides translation, summarization and writing functions through browser plugins. In 2023, the team also acquired the ChatGPT for Google plugin to gain access to its existing user base. After more than a year of operation, Monica accumulated millions of users, with annual recurring revenue reaching millions of US dollars.
During several entrepreneurial attempts, Xiao Hong's team did not participate in the construction of underlying technologies. WeChat is responsible for connecting users, WeCom provides customer management interfaces, and companies including OpenAI provide foundational models. The team leverages the capabilities of existing platforms to develop specific software functions, and then charges users for the services.
Every time a large platform opens up new technical capabilities, it leaves a batch of unmet demands that have not been fully satisfied. WeChat provides the Official Account backend, but operators still need editing tools; WeCom connects customers, but enterprises need to manage customer relationships; ChatGPT can generate content, but users hope to call the model directly when browsing web pages.
What Xiao Hong's team entered is exactly the market where the platform provides basic capabilities but complete products have not yet been fully formed.
The user scale and technical investment of large platforms reduce the R&D cost of startups. The team can allocate more resources to product design, user growth and sales. However, platforms may also add related functions, compressing the operating space of third-party software. Yiban, Weiban and Monica have all faced competition from similar products, as well as pressure brought by platform function adjustments.
In 2024, Monica had already generated revenue, but Xiao Hong's team still spent several months developing an AI browser. Investor Liu Yuan recalled that he had experienced the new product in Wuhan, and he did not even realize that he was already using a brand new browser after the demonstration. The team finally gave up the project.
In March 2025, Manus was launched. The product can call browsers, code execution environments and external tools according to the tasks proposed by users, to complete information collection, data processing and report production.
From Monica to Manus, the team still uses external foundational models, but the work undertaken by the product has increased significantly. Monica mainly helps users process information, while Manus tries to continuously perform multiple operations and deliver results.
Xiao Hong's entrepreneurial experience shows that opportunities for application companies often appear between the technical capabilities of large platforms and the actual needs of users. The stronger the capabilities provided by the platform, the more complex products startups can develop, but they also need to face the possibility that the platform will directly enter the application market.
II. The Commercial Inflection Point of AI Applications
On December 9, 2025, US venture capital institution Menlo Ventures released the report *2025 State of Enterprise Generative AI*. The research team surveyed about 500 US enterprise decision-makers, and combined with market model estimation, it showed that enterprise generative AI spending reached 37 billion US dollars in 2025, an increase of about 2.2 times from 11.5 billion US dollars in 2024. Among them, AI application spending was about 19 billion US dollars, accounting for 51% of the total expenditure.
In 2023, enterprise generative AI spending was only about 1.7 billion US dollars. In two years, the market size has expanded by more than 20 times. Enterprise AI procurement has expanded from early experiments to actual business deployment.
Menlo Ventures also found that in 2025, AI-native startups obtained about 63% of the revenue share in the AI application market, higher than 36% in 2024. The traditional enterprise software market has long been controlled by large manufacturers in terms of customer relationships and sales channels, while a group of new suppliers have emerged in the AI application market.
However, the progress of enterprises in procuring AI products is not consistent.
According to the classification of Menlo Ventures, in 2025, AI applications for different departments of enterprises obtained 7.3 billion US dollars in expenditure, of which programming tools accounted for 4 billion US dollars. Vertical industry applications such as medical care and finance obtained 3.5 billion US dollars in expenditure. The general-purpose AI application market reached 8.4 billion US dollars, which is mainly composed of products such as ChatGPT Enterprise, Claude for Work and Microsoft Copilot.
Programming tools have taken the lead in forming a large paid market, which is related to the characteristics of software development work. Programmers have long used code editors and automated testing tools, and AI-generated code can be verified through compilation, testing and manual review. Enterprises can relatively easily measure the efficiency changes brought by AI tools.
The development of the agent market is relatively lagging behind.
Menlo Ventures estimates that in the general-purpose AI application market in 2025, Copilot-type assistants obtained about 7.2 billion US dollars in expenditure, accounting for 86%; Agent platforms obtained about 750 million US dollars, accounting for about 10%. Enterprises are already willing to pay for AI-assisted work, but general agents that can independently perform multi-step tasks have not yet become the main category of enterprise procurement.
The two sets of data reveal a change in the AI application market. Enterprises have invested a lot of money in purchasing software that can help employees work, but agents that directly undertake work tasks are still in the early stage of commercialization.
In the past, when enterprises purchased software, they mainly paid according to the number of accounts, usage period and function scope. Software improves the work efficiency of employees, but the work itself is still completed by employees. After agents can perform part of the tasks, enterprises can also evaluate product value according to the number of completed tasks, saved manual time and delivery quality.
Taking business research as an example, researchers need to find corporate announcements, verify financing information, organize data, and then produce reports. AI assistants can help read and summarize materials, while agents can try to complete information collection, table sorting and draft production. When evaluating related products, enterprises will not only consider the software cost, but also calculate how much manual time they would have originally invested.
Agents may change the revenue source of software companies. Traditional software mainly obtains the enterprise's informatization budget, while agents that can stably complete work have the opportunity to obtain part of the enterprise's original expenditure for performing related tasks.
However, whether the work budget can be converted into software revenue depends on the actual performance of the agent. Enterprises need to calculate the cost of purchasing services, the cost of manual review and the loss caused by task failure. Only when the comprehensive cost is lower than the original working method, continuous procurement has economic rationale.
In October 2025, a16z cooperated with fintech company Mercury to analyze the payment data of more than 200,000 of its customers, and selected 50 AI-native application companies that obtained actual expenditure. Manus ranks 33rd on the list, which also includes companies such as OpenAI, Anthropic and Perplexity.
Compared with website visits and downloads, enterprise payment records can more directly reflect the commercialization of products. However, the relevant samples mainly come from startups served by Mercury, and cannot represent the entire enterprise AI market.
Enterprise AI application spending is growing rapidly, and general agents have not yet been fully commercialized. The market where Manus is located is in a period of transition from assisting employees in their work to performing part of work tasks.
III. What is Capital Betting On
On December 17, 2025, Manus released a set of operating data. The company's annual recurring revenue exceeded 100 million US dollars, and the overall annualized revenue run rate including usage-based fees and other items exceeded 125 million US dollars. Nine months after the product was launched, it processed more than 147 trillion Tokens in total, and created more than 80 million virtual machine instances.
A few months ago, the capital market's judgment on Manus was mainly based on the team and product prospects. In 2025, Benchmark led a financing of about 75 million US dollars for Manus. Manus disclosed that the investment took place before the official launch of the product. By December, the company had formed an annualized revenue of hundreds of millions of US dollars. Meta announced the acquisition of Manus in a deal reportedly worth over 2 billion US dollars.
In 2026, Manus' revenue continued to grow. According to *The Information*, in June, Manus' annualized revenue run rate reached approximately 500 million US dollars. On October 8, after resuming independent operations, Manus secured a new financing of over 500 million US dollars. The company did not disclose the valuation of this round of financing, and the outside world cannot accurately calculate the revenue multiple based on public information.
Capital's continuous investment in Manus reflects that the AI industry is forming a new division of labor. Large technology companies invest funds to develop foundational models, while application companies use model capabilities to develop products, acquire users and provide services. Manus does not independently train cutting-edge foundational models, but has already generated revenue through subscriptions and usage-based fees.
However, the improvement of foundational model capabilities has two impacts on application companies at the same time. The more complex tasks the model can complete, the more powerful products application companies can develop. At the same time, as the model is opened to the market through APIs, competitors can also purchase similar technical capabilities.
The foundational models that Manus can call are usually also accessible to other agent companies. The popularization of underlying technologies lowers the threshold for application development, and also makes it difficult to establish product differentiation solely by relying on model capabilities.
Therefore, the price capital pays for Manus includes the judgment on the operating capabilities of application companies.
When users put forward demands for recruitment screening, stock research or report production, Manus needs to break down the tasks into multiple steps, call models and tools, handle exceptions that occur during the execution process, and then deliver the results. Products need to continuously improve task completion rate, delivery time and user experience to get users to pay.
More importantly, a new profit relationship may be formed between model cost and product revenue.
Every time an agent performs a task, it needs to consume model Tokens, computing resources and running time. According to the data disclosed by Manus, the product processed 147 trillion Tokens in total and created 80 million virtual machine instances in the nine months after its launch. The relevant figures show that Manus needs to continuously purchase computing resources, but the company has not disclosed the corresponding costs, and the outside world cannot yet judge how much profit each task can generate.
In the future, if the call price of foundational models continues to drop, the cost for agents to complete the same task may decrease. Assuming that customers are willing to pay a stable price for the same work results, the profit margin of application companies will have the opportunity to expand.
However, the decline in technical costs will also reduce the operating costs of competitors, and push similar products to cut prices. Whether the economic benefits created by model progress can finally remain in Manus depends on whether the product can maintain customers' willingness to pay.
The actions of large technology companies further increase the uncertainty of competition. OpenAI, Anthropic and Google continue to develop products with tool calling and task execution capabilities. Foundational model companies not only provide technical services to independent developers, but also can directly sell agents to end users.
For Manus, upstream suppliers are also potential competitors. The company needs to build operating capabilities that can continuously obtain revenue through task execution quality, workflow and customer relationships.
What Manus currently discloses is mainly revenue growth and computing resource usage scale. Indicators such as customer renewal rate, average task execution cost and overall gross profit margin have not been fully disclosed. The annualized revenue run rate is not equal to the full-year operating revenue that has been realized, and the company's future profitability still needs to be verified by more operating data.
The scale of the AI application market is expanding, but the profits created by market growth may not necessarily be distributed according to the amount of technical investment. Foundational model companies have R&D capabilities and computing resources, while application companies try to master customer needs and workflows. As agents take on more tasks, both sides may obtain new revenue.
Manus' high-value financing includes recognition of the company's revenue growth, as well as capital's expectation for the future agent market. After the capabilities of foundational models continue to improve, whether independent application companies can retain sufficient profits will become an important issue to test Manus' long-term commercial value.