With just one day remaining until the end of 2025, the US social media giant Meta suddenly announced the acquisition of the AI agent startup Manus. This move introduces a dynamic and fast-paced Chinese entrepreneurial team, drawing a symbolic conclusion to a year filled with anxiety and a sense of crisis for Mark Zuckerberg.
The acquisition was finalized in a year-end, fast-tracked deal. Meta's head of AI business, Alexandr Wang, announced on social media yesterday that Manus had joined Meta, expressing his anticipation to build outstanding AI products alongside Manus co-founder and CEO Xiao Hong. He praised the Manus team as being world-class in exploring the surplus capabilities of contemporary models and capable of building powerful AI agents.
Xiao Hong of Manus responded, stating, "Joining Meta allows us to develop on a stronger, more sustainable foundation, without changing Manus's operational or decision-making methods." Following the acquisition's completion, he will assume the role of Vice President at Meta, with the team continuing to be based in Singapore. The acquisition of the AI agent product company Manus is not merely another expansion and gap-filling exercise for Meta's AI business portfolio; it represents another high-stakes gamble by Zuckerberg and Alexandr to inject external forces to reshape the internal organization and boost the competitiveness of their own AI products amidst ongoing internal conflicts. Meta stated in its announcement that after the transaction closes, it plans to continue operating and selling Manus's services, integrating them into its social media product suite. AI agent tools have become a fiercely contested area among tech giants. "We plan to expand this service to more enterprises."
The financial terms of the deal were not disclosed by either party, but according to US media reports, the purchase price was between $2 billion and $3 billion. Furthermore, when Meta was negotiating the purchase, Manus was itself seeking a new funding round with a valuation of $2 billion. The entire process, from initial contact to the announcement of the deal, took less than half a month. This is typical of Zuckerberg's deal-making style: identifying value and striking swiftly. Placing this transaction within the context of Meta's series of AI strategic adjustments this year provides greater insight into Zuckerberg's severe anxiety regarding his own AI business: from the dismal performance of Llama 4, to offering sky-high annual salaries to recruit top industry talent, to investing $14.3 billion for a stake in Scale AI and bringing in Alexandr to parachute in and reorganize the AI business, followed by major layoffs in the fundamental research department, and the resignation of AI godfather Yann LeCun. The year concluded with the $2 billion acquisition of Manus, drawing a final line under a tumultuous year of restructuring for Meta's AI operations.
Although nominally a Singaporean company, Manus is essentially an overseas venture of a Chinese entrepreneurial team, having relocated to Singapore only in June of this year. The company behind Manus, Butterfly Effect, was founded in 2022 by Chinese entrepreneurs Xiao Hong, Ji Yichao, and other partners, with offices in Beijing and Wuhan. Butterfly Effect's first product was the AI assistant Monica, a browser extension plugin integrating models like ChatGPT and Claude, which had accumulated over 10 million users and achieved profitability by last year. However, what truly brought Xiao Hong's team to the attention of the global AI industry was the launch of Manus in March of this year.
This product has been dubbed the "world's first true general-purpose AI agent." It can run asynchronously in the cloud, autonomously completing complex tasks such as creating websites, analyzing stocks, formulating travel plans, and screening resumes. Unlike traditional chatbots, Manus employs a multi-agent architecture, with a central "executor" coordinating multiple specialized sub-agents to handle different stages of a task. Manus's technical implementation does not involve developing a new model from scratch; rather, it cleverly orchestrates existing powerful models—primarily using Anthropic's Claude 3.5 Sonnet and a customized version of Alibaba's Qwen, combined with 29 specialized tools for web browsing, API interactions, and script execution. This "integration over innovation" strategy allowed Manus to launch quickly and, in some benchmark tests (like GAIA), surpass OpenAI's Deep Research. What is particularly impressive is Manus's speed of commercialization. According to their recent presentations, within just over eight months since launch, Manus has processed over 147 trillion tokens, created 80 million virtual computers, and serves millions of individual and enterprise users.
Furthermore, Manus's Annual Recurring Revenue (ARR) has already exceeded $100 million, reportedly making it the fastest startup globally to reach this milestone. It's important to clarify that ARR refers to projecting future annual revenue based on the current run rate, not the actual revenue from the past year.
Following this acquisition by Meta, the equity of all Manus investors was bought out. This was the most desirable outcome for Manus's many investors, particularly the US venture capital giant Benchmark. After its founding, Butterfly Effect completed two funding rounds, raising a total of $85 million; domestic investors included Tencent, ZhenFund, and Sequoia Capital China, among others. In April of this year, while still based in China, Manus completed a $75 million Series B funding round led by the US venture capital firm Benchmark, achieving a post-money valuation of approximately $500 million. Benchmark's Managing Director, Chetan Puttagunta, consequently joined Manus's board of directors. According to a source in the Silicon Valley venture capital circle who spoke to Silicon Valley Observer, the Benchmark fund played a key role in brokering the deal between Meta and Manus. Meta's buyout acquisition of Manus means Benchmark's investment yielded a 4-5x return in just eight months, representing a very successful quick-in, quick-out investment.
It is worth noting that, as a well-known Silicon Valley VC firm, this is the second time Benchmark has profited handsomely from an acquisition by Zuckerberg. They were previously the lead Series A investor in Instagram and also participated in the Series B round, ultimately achieving a total return of over 25x on their investment in that deal. In 2012, immediately after Instagram completed a Series B round valuing it at $500 million, Zuckerberg swiftly locked in an acquisition at double that valuation, eliminating what was potentially Facebook's most significant future competitor. Today, Instagram is valued in the hundreds of billions of dollars, and Zuckerberg's bold gamble is considered one of the most successful acquisition deals in internet history.
To understand why Meta was so eager to acquire Manus, one needs to review the numerous setbacks Meta faced in the AI domain throughout 2025, which illuminate Zuckerberg's anxiety and sense of crisis about falling behind in the AI competition. The AI industry in 2025 witnessed the surprising rise of the Chinese team DeepSeek, as well as the chaos and setbacks experienced by the AI giant Meta. In April, Meta released the Llama 4 series of models, including Scout (109B parameters), Maverick (400B parameters), and the in-training Behemoth (planned for 2 trillion parameters). This was supposed to be another victory for Meta in the open-source AI arena, but instead, it turned into an embarrassing disaster.
Almost immediately after release, researchers noted issues with the benchmark scores. Meta had used a specially optimized "experimental chat version" in its tests, rather than the publicly released version. This practice drew skepticism from the community. Worse still, Llama 4 lagged significantly behind competitors, especially China's DeepSeek models, in key areas like coding and complex reasoning. The combination of internal and external criticism deeply affected Zuckerberg. According to insiders, he was particularly angered by the perception that Meta was trying to "cover up" the product's underperformance and was very disappointed with the Meta AI team. Company executives also felt the model's performance fell short of expectations, while external developers accused Meta of overpromising and underdelivering. More worryingly, the release date for Behemoth was postponed indefinitely. As reported by The Wall Street Journal, there was internal frustration at Meta regarding the progress of the Behemoth team, leading Zuckerberg to consider "significant management changes" for the AI product group responsible for developing Llama 4 Behemoth. Open-source large language models were originally Meta's leading track, but they have now been completely overtaken by the "wolf pack" of Chinese models. Chinese models like DeepSeek and Alibaba's Qwen rank at the top in multiple benchmarks, often matching or surpassing top closed-source models (such as OpenAI's o1, Claude 4.5, Gemini 2.5), particularly in mathematics, programming, reasoning, and multilingual tasks.
For developers, Chinese large models also offer a clear cost advantage. A recently published OpenRouter report (in collaboration with a16z, covering 100 trillion tokens) showed that the global usage share of Chinese open-source models has skyrocketed from approximately 1% at the end of last year to a peak of nearly 30%. The overall open-source share is around 33%, with Chinese models dominating the growth.
So, why did Llama 4 perform so poorly? A former Meta employee who was directly involved in the Llama team's development revealed to Silicon Valley Observer that the root cause likely lies more with decision-making issues at the middle and senior management levels, highlighting a serious problem of "amateurs leading professionals." The original Llama team had prioritized investment in the multimodal direction, given Meta's diversified product ecosystem, including the metaverse, smart glasses, and social media. However, after the emergence of DeepSeek earlier this year, whose reasoning capabilities significantly surpassed Llama's, it caused "great panic" within the Meta team. The team tried to excel in both areas, but time was insufficient, leading to product chaos. This former Meta employee disclosed that the root of the problems within the Meta AI team lies in personnel being unfit for their positions. Some mid-level and senior managers at Meta, who originally worked on products, have weak AI technical backgrounds yet lead the actual AI developers—a classic case of "amateurs managing professionals." This misalignment of authority was one of the key reasons prompting his decision to leave. The original Llama team consisted of 14 scholars and researchers holding PhDs, but 11 of them have since left the company. Subsequent versions were developed by entirely different teams, and this brain drain severely impacted product continuity. This former employee also left Meta before Zuckerberg brought in Alexandr.
This series of setbacks plunged Zuckerberg into anxiety, prompting him to decide on radical change. Zuckerberg recognized that serious team leadership issues underpinned the failure of Llama 4 but believed it was difficult to effect change through internal means alone. He felt the need to find an external "catalyst fish" to make the team competitive again. This was the direct reason Zuckerberg brought in Alexandr to lead the Meta AI department. Undoubtedly, Zuckerberg's most significant deal this year was the $14.3 billion investment to acquire a 49% stake in the AI data annotation company Scale AI, simultaneously recruiting its 28-year-old founder and CEO, Alexandr Wang, to lead the newly established Meta Superintelligence Labs (MSL). This represents Meta's largest external investment to date.
Zuckerberg places high hopes on Alexandr, expecting him to bring breakthroughs in data innovation for Meta. According to informed sources, Meta's leadership had complained about the lack of progress in data innovation within the company's leading AI teams. Alexandr's expertise lies precisely in the high-quality data annotation and evaluation required for AI model training, which is exactly what Meta needs. Beyond recruiting Alexandr to parachute in and lead Meta AI, Zuckerberg personally participated in AI talent recruitment, creating an executive group called "Recruiting Party" to discuss and scout for talent around the clock. Over the past six months, Meta has offered compensation packages in the "seven to nine figures" to researchers from renowned AI companies like OpenAI and Google. The company also reconfigured the seating layout at its headquarters to place core new hires near Zuckerberg.
Alexandr's arrival quickly triggered organizational tremors within Meta AI. This was perhaps what Zuckerberg hoped for, but it may have also exceeded his expectations.
First came departmental restructuring and a significant purge. After Alexandr's parachute appointment, Meta's legendary AI scientist Yann LeCun became his subordinate. For a renowned scholar with decades of research experience in AI, widely recognized as one of the "fathers of AI," this was an unacceptable "humiliation," directly foreshadowing his subsequent departure. In October of this year, Alexandr streamlined and integrated the Meta AI team, laying off over 600 people. Except for his directly affiliated TBD Lab, which saw no layoffs, the other three departments were significantly affected, especially the FAIR fundamental research department. Many AI research talents, including the "Chinese AI expert" Yuan Dong Tian, were ruthlessly cut. Second were strategic disagreements and resource conflicts. The New York Times reported that during internal meetings in the fall of 2025, Alexandr had clear disagreements with Meta's veteran executives, including Product Chief Chris Cox and Technology Chief Andrew Bosworth, both long-time confidants of Zuckerberg. Reportedly, the conflict centered on fundamental issues like AI investment priorities, development timelines, and integration with Meta's existing products. The veteran executives advocated leveraging the company's vast social data advantage, using data from Facebook and Instagram to train Meta's new models.
However, Alexandr held a different view. He argued for focusing on catching up with the models from Google and OpenAI, believing competition should happen at the technological frontier rather than relying on existing product data. This divergence reflected two different visions for AI strategy: incremental improvement using existing assets versus boldly pursuing technological breakthroughs. At least for now, Zuckerberg has given Alexandr sufficient support and trust. Third were cultural clashes and talent attrition. On one hand, after his arrival, Alexandr assembled his own direct elite team and poached top AI development talent from competitors with high salaries, directly creating a pay gap with the original Llama team. This disparity affected morale within the Meta AI department, prompting some existing employees to seek opportunities elsewhere.
On the other hand, some of the highly paid talent recruited from outside also sought to leave due to cultural clashes. New hires from OpenAI and Scale AI expressed frustration with the bureaucracy of the large corporation. Ruben Mayer, former Senior Vice President at Scale AI, left Meta after just two months. Generative AI Product Management Director Chaya Nayak and Research Engineer Rohan Varma also announced their departures in recent weeks. According to a mid-December report by the Financial Times, Alexandr privately complained to colleagues that Zuckerberg's "micromanagement" made him feel "suffocated." Meanwhile, some employees questioned his ability to manage a research team of Meta's scale, pointing out that his background is in AI data services rather than cutting-edge AI development. This mutual dissatisfaction signaled fragile internal relationships.
Against this backdrop of reorganization, Meta's swift acquisition of Manus takes on deeper significance. Both Alexandr and Zuckerberg hope the acquisition will complement Meta's AI business, bringing new users and data, while also injecting the competitive drive of a Chinese entrepreneurial team. Meta possesses powerful open-source foundation models in the Llama series, leading in model parameters, reasoning efficiency, and multimodal capabilities. However, Meta's core gap lies in transforming these models into reliable, autonomous systems capable of completing complex tasks—agent systems. Manus is precisely the best supplement for this Meta gap. Manus claims to be the "world's first general-purpose AI agent," capable of independently handling end-to-end tasks like market research, resume screening, travel planning, code writing/deployment, and stock analysis with almost no human intervention. It relies on a multi-agent architecture + third-party models, focusing on execution reliability, tool usage, and context engineering.
Moreover, Manus has already proven its business model. In less than nine months since launch, it has gained millions of paying users and enterprises, with ARR exceeding $100 million. This is a rare case of rapid commercialization success in the AI agent field, demonstrating market demand. Such a business is an excellent complement for Meta. Post-acquisition, the Manus business will continue to operate independently while being integrated into the broader Meta family of products like Meta AI, Facebook, Instagram, and WhatsApp, realizing the vision of "agentic-rich" personal AI. This directly fills Meta's shortcoming in the "agent execution layer," transforming Llama models from "good at chatting" to "good at doing things," and providing automation tools for small and medium-sized businesses (especially on WhatsApp SMB). From a competitive perspective, acquiring Manus prevents competitors from strengthening themselves. AI agents have surpassed chatbots as the focal point of competition in the foreseeable future. In this arena, OpenAI has Deep Research and Operator, Anthropic has Claude's computer use capabilities, and Google has Gemini's multimodal capabilities. To avoid falling behind in the competition, Meta needs similar AI agent capabilities to maintain competitiveness, especially when integrating AI into its products used by billions, like Facebook, Instagram, and WhatsApp. This is the direct significance of acquiring Manus. Additionally, Manus's Chinese team and their experience with the Qwen model might help Meta better understand and respond to AI competition from China.
As a company founded by Chinese talent, relocated to Singapore, backed by US基金 investment, and now acquired by a US tech giant, Manus's development journey epitomizes the globalized nature of the AI era. Whether their technology and culture can survive within the Silicon Valley behemoth will be an important case study for observing the international export path of Chinese AI entrepreneurs. The success story of Manus once again proves that in the AI field, Chinese entrepreneurs have reached world-class levels in execution, productization capability, and commercialization speed. Achieving $100 million ARR within eight months is a commercialization speed rare even in Silicon Valley. Like TikTok, behind Manus's success lies the triumph of methodologies honed over years in China's internet industry: rapid iteration, lean startup principles, and a user-centric approach.
However, the future integration following this acquisition also holds uncertainties. Currently, Manus is built on external models like Claude and Alibaba's Qwen, while Meta naturally wants to promote its own Llama models. If Manus is required to migrate to Llama, it seems almost inevitable. Should Llama's capabilities prove insufficient to support Manus's functionality, this could impact the product's competitiveness.
Manus's founders, Xiao Hong and Ji Yichao, are seen as representatives of the new generation of Chinese AI entrepreneurs, accustomed to rapid iteration and flexible decision-making. This could clash with the established processes of a large company like Meta. By introducing these two Chinese entrepreneurs, Zuckerberg is essentially making another high-stakes bet: wagering that these two "Chinese catalyst fish" can thoroughly invigorate Meta's AI ecosystem before being consumed by the large corporate culture. Currently, Meta has adopted a relatively flexible strategy. The Manus team is joining Meta as independent employees for now, maintaining operations in locations like Singapore, attempting to preserve the rhythm of a startup without fully integrating into Meta's processes. The Manus team hails from a Chinese startup background, accustomed to "996" work schedules and rapid trial-and-error. While Meta also emphasizes "moving fast," as a mature public company, its internal processes, compliance requirements, and decision-making mechanisms are far more complex. This cultural collision could cause the team to lose its original innovative vitality. In a post-acquisition reflection, Xiao Hong mentioned that the real difficulty lies not in technical analysis, but in "overcoming internal inertia and turning it into organizational action." This hints that his current role within Meta involves more of driving cultural change and combating corporate bureaucracy—precisely the changes Zuckerberg and Alexandr hope to see. On the surface, Meta's acquisition of Manus is a routine tech acquisition. In essence, it is a microcosm of the Silicon Valley giant's anxiety in the AI era, another attempt by Zuckerberg to use money and speed to introduce a "Chinese catalyst fish" to combat organizational inertia, and yet another manifestation of Chinese AI entrepreneurs exporting their global influence.
Comments