Summary: Manus AI launched in March 2025 with a demo video that got over a million views in 20 hours. Within eight months it had crossed $100 million in annual recurring revenue. By December 2025, Meta had acquired it for approximately $2 billion. By April 27, 2026, China’s National Development and Reform Commission had ordered the deal canceled with a 54-character statement. Bloomberg reported the model was “officially dead.” The founders had already been banned from leaving China. This is the full story of how the world’s most-hyped AI agent went from viral to vanished in 14 months. Part of our AI Industry Analysis series.
The Demo That Broke the Internet
On March 6, 2025, a startup called Butterfly Effect sent out invitation codes for something called Manus. The demo video — showing the agent autonomously screening resumes, analyzing stock data, conducting market research, building reports — got over one million views in twenty hours. The invite queue crashed. People were selling invite codes online. The AI world called it the first genuinely autonomous general-purpose AI agent.
The name was deliberate. Manus is Latin for hand. The product was positioned as the hand that followed the mind — not just reasoning, but executing. You described a task. Manus completed it. Browser, terminal, code editor, file system — it had access to the tools and it used them without being asked to specify each step.
The benchmark results backed the hype. On GAIA — a comprehensive evaluation of an AI agent’s ability to reason, use tools, and automate real-world tasks — Manus hit state-of-the-art scores, reportedly exceeding the previous champion’s performance by 10 points on Level 3 tasks, where competitors included OpenAI’s best models.
The company behind it, Butterfly Effect, had started in Wuhan and Beijing building Monica, a browser extension that aggregated commercial LLMs for translation, summarization, and writing. Monica was useful. Manus was something different: an attempt to build a general agent that actually got things done.
The Numbers
By the time Meta acquired the company in December 2025, Butterfly Effect had already told backers it crossed $100 million in annual recurring revenue — eight months after launch. For an AI agent startup in a market where most products monetize through developer subscriptions or enterprise deals, $100M ARR at that pace was extraordinary.
The growth trajectory drew Benchmark, which led a $75 million Series B in April 2025 at a $500 million valuation. That was nine months before the Meta deal. The jump from $500 million to approximately $2 billion in under a year reflected both the ARR growth and the strategic value Meta placed on Manus’s agent architecture, which was differentiated: a multi-agent coordination system running in a sandbox with CodeAct capabilities, giving it a profile distinct from what Meta was building internally.
By mid-2025, the three co-founders — including CEO Xiao Hong and chief scientist Ji Yichao — had relocated from China to Singapore. Butterfly Effect established a Singapore entity that operated the AI agent product in markets outside China. The China-to-Singapore relocation was a standard playbook for Chinese AI startups with global ambitions: maintain Chinese engineering talent, move the legal and commercial entity offshore, target Western customers and acquirers.
Meta’s Acquisition
On December 30, 2025, Meta confirmed the acquisition. The company was described as a Singapore-based developer of general-purpose AI agents, and Manus employees were reported to have already joined Meta’s AI team. The deal was structured as an acqui-hire with product acquisition: Meta wanted the team, the architecture, and the $100M ARR product.
The deal was valued at approximately $2 billion — some reports put it at up to $3 billion. Tencent and HongShan Capital, two of the major backers, received their portion of proceeds. Founders were publicly listed as members of Meta’s expanded AI organization.
For Meta, the acquisition fit a pattern. The company had been making aggressive AI moves throughout 2025, and an autonomous agent with demonstrated ARR and novel architecture was a defensible addition — particularly as competitors at Anthropic, OpenAI, and Google were all racing toward agent capabilities.
What Meta may not have fully modeled: the Chinese government would have something to say about it.
The Travel Ban
In January 2026, China’s Ministry of Commerce announced it would conduct an assessment of whether the acquisition complied with laws on export controls, technology import and export, and overseas investment. This was not unexpected. Chinese regulatory review of outbound technology acquisitions had been a known risk since the mid-2020s tightening of export control frameworks.
What followed was less expected.
In March 2026, Euronews reported that Chinese authorities had banned Manus’s co-founders from leaving China. CEO Xiao Hong and chief scientist Ji Yichao were told they could not depart while the regulatory review was ongoing. For founders who had already relocated to Singapore, returning for any reason now meant being grounded.
This was not a negotiating posture. China’s use of exit bans against business figures under regulatory or legal review is a documented enforcement mechanism — used in Ant Group and Didi proceedings, among others. The signal to other Chinese entrepreneurs with cross-border ambitions was unmistakable.
The 54-Character Decree
On April 27, 2026, the National Development and Reform Commission issued its ruling. Bloomberg described it as a “chilling 54-character decree from the top state planner.” The NDRC ordered the acquisition canceled. Meta was required to unwind the deal.
The Ministry of Commerce’s earlier review had reportedly characterized the acquisition as a “conspiratorial” attempt to hollow out China’s technology base. Whether that language reflected the government’s sincere view or was a negotiating position for a different outcome, it had no visible effect on the outcome: the deal was blocked on national security grounds.
The complications were immediate and structural. Tencent and HongShan had already received their proceeds. Manus employees had already joined Meta. The product had been operating under Meta’s umbrella for months. How does an acquisition get unwound when the money has already moved and the team has already merged?
The answer, two days later, was in Bloomberg’s headline: China’s Meta Backlash Renders Manus Model ‘Officially Dead’.
What “Officially Dead” Means
The Manus AI product — the agent people were selling invite codes to access a year earlier — is gone. Not deprecated. Not paused. The model that reached $100M ARR, that led the GAIA benchmark, that Meta paid $2 billion for, has been shut down.
The reasons are intertwined. The regulatory blockage made the acquisition’s legal status uncertain. The Chinese founders are travel-banned. The team is at Meta. The China-based entity that originally built the model is now in an ambiguous legal position with respect to IP, employment, and regulatory obligations. Operating the product, in this context, became untenable.
For users, there is no recourse path that is obvious. Reviews of Manus from early 2026 were already noting customer support failures, billing inconsistencies, and session reliability problems — the signs of a company whose operational focus had moved to the acquisition process. The product kept running. Now it has stopped.
What the Story Means
Manus is not primarily a story about a failed product. It was a successful product: real ARR, real users, real benchmark results, a real acquisition at a real price. The failure is structural and geopolitical.
Three things converged:
China’s tech war posture hardened. The NDRC’s willingness to block a $2 billion transaction and ban the founders of a Singapore-domiciled company from leaving the country reflects an enforcement posture that is more aggressive than most founders building in China — or relocating from China — had priced in. The Butterfly Effect founders did the legally correct thing: established a Singapore entity, relocated, built internationally. The Chinese government still had jurisdiction over them.
The exit path for Chinese AI startups is narrower than the pitch decks assumed. Manus demonstrated ARR, product-market fit, and strategic value to a tier-one acquirer. That should have been a clean exit. Instead, it became a case study in what happens when a Chinese-origin startup with Chinese engineering, Chinese investors, and Chinese-trained founders tries to sell to an American platform company in the current geopolitical environment. The U.S. side wasn’t the blocker. The Chinese side was.
Autonomous agents carry higher regulatory sensitivity than models. Manus was not a foundation model or an API. It was an agent — something that takes actions, uses tools, accesses systems, operates autonomously. That profile — an autonomous system developed by a Chinese company, acquired by a U.S. social media platform — appears to have read differently to Beijing than an LLM licensing deal would have.
The Benchmark That No Longer Matters
Among the more poignant footnotes: Manus’s GAIA benchmark scores remain in the historical leaderboard. The numbers are still there — the 10-point lead on Level 3, the documented outperformance of OpenAI’s models at the time. The benchmark existed to demonstrate that Manus could do things others couldn’t. It did demonstrate that. The model is still gone.
That gap — between demonstrated capability and actual availability — is the defining feature of Manus’s story. The capability was real. The regulatory environment was realer.
What Remains
Manus AI the product is gone. The Butterfly Effect / Monica browser extension product line may continue in some form — Monica was the original product, and it was built for Chinese users on Chinese infrastructure. The Manus brand, the Manus model, and the Manus team are now distributed across Meta’s organization with unclear long-term IP implications.
For the Chinese AI agent ecosystem, the precedent is stark: a company can build a genuinely differentiated product, achieve $100M ARR, relocate to Singapore, close a tier-one acquisition, and still have the deal killed by a 54-character government statement. The founders can be banned from leaving. The product can be shut down. The money, once distributed, may or may not need to be returned.
What Manus actually built still matters. The architecture was real. The approach — multi-agent coordination, sandbox execution, CodeAct — was differentiated enough that Meta was willing to pay $2 billion for it. The knowledge is now inside Meta. That is a form of continuation, even if the product is not.
But the agent that crashed its own invite queue in March 2025, that people were paying for invite codes to access, that hit $100M ARR in eight months — that is gone. The hand that followed the mind has been closed by a decree that, in its 54 characters, said less than the product once did.
This analysis is based on reporting from Bloomberg, CNN, CNBC, Euronews, MIT Technology Review, and TechCrunch. ChatForest researches AI products and developments; we do not conduct independent hands-on product testing.