The Agent Beijing Wouldn't Let Go: Manus and the $4 Billion Second Act Nobody Planned

On September 18, 2026, Bloomberg reported that Manus — the general-purpose AI agent that went viral in China eighteen months ago — is raising $500 million at a $4 billion valuation. The potential investors read like a cross-section of Chinese capital: IDG Capital, Boyu Capital, and Contemporary Amperex Technology Co. (CATL), the world's largest battery maker. Tencent, which already holds stakes from two earlier rounds, is in talks to become the company's largest shareholder.

What makes the number remarkable is not the number. It is everything that happened to get there.

Manus has, in the space of eighteen months, been the viral sensation of Chinese AI, the poster child of the country's agent ambitions, a $2 billion acquisition target of Meta Platforms, a national-security case for Beijing's top economic planner, and a company that had to buy itself back from one of the world's most valuable corporations before it could legally exist on its own again. On September 1, 2026, it resumed independent operations. Seventeen days later, it was reportedly worth $4 billion.

The story of Manus is the story of what happens when the two most powerful governments on Earth decide, almost simultaneously, that a thirty-something entrepreneur's agent startup is strategically important. It is also the cleanest case yet of a new rule in the AI era: in China, the agent is a national asset, whether the founder wants it to be or not.

Manus at a Glance

Phase One: The Butterfly (2022–2024)

Before Manus, there was Monica — and before Monica, there was a serial entrepreneur from Wuhan who kept arriving slightly too early.

Xiao Hong graduated from Huazhong University of Science and Technology in 2015 and founded Wuhan Nightingale Technology the same year. In 2022 — two months before OpenAI would release ChatGPT and upend the industry — he founded Butterfly Effect (蝴蝶效应), keeping offices in Beijing and Wuhan. The company's first product was Monica: an all-in-one AI assistant and Chrome extension that aggregated multiple large language models behind a single sidebar for translation, summarization, and writing help. It was unglamorous, unheralded, and enormously successful — tens of millions of installs across North America, Japan, and South Korea.

The Monica years matter for two reasons. First, they gave the team something almost no Chinese AI startup had: direct, years-long experience shipping consumer AI products to a global, English-first audience. Second, they established a pattern of refusing easy exits. In early 2024, ByteDance reportedly offered roughly $30 million to acquire the team. Xiao Hong declined.

That refusal looks prescient now. It also looks, in hindsight, like the moment the company picked its destiny: it would not be absorbed into a larger platform. It would try to become one.

The funding table tells its own story: a company that went from $14 million to $4 billion in four years — and passed through a $2 billion Meta deal in the middle that no longer exists.

Phase Two: The Thunderclap (March 2025)

On March 6, 2025, Butterfly Effect opened Manus to the world in invitation-only beta. What happened next remains one of the most violent launches in AI history.

Manus was not a chatbot. It was an agent: you gave it a high-level goal — research this market, screen these résumés, plan this trip, build this website — and it planned the subtasks, operated a real browser, wrote and executed code, managed files, and returned a finished deliverable. Under the hood, a planning agent coordinated specialist sub-agents, powered in part by Claude 3.5 Sonnet and fine-tuned versions of Alibaba's Qwen, running in a sandboxed cloud virtual machine.

The launch demo video — produced in under a week, per Baidu Baike's account, with borrowed camera lenses and a paid editing-software membership — passed one million views in twenty hours. Within days, the waitlist reportedly crossed two million sign-ups, and invitation codes were being scalped on Xianyu, Alibaba's second-hand marketplace, for as much as ¥100,000 (about $14,000). "Begging for a Manus invite code" became a national pastime. Rowan Cheung, founder of The Rundown AI newsletter, captured the global reaction: *"It's like Deep Research + Operator + Claude Computer combined, and it's REALLY good."* Chinese state media promptly declared Manus "the next DeepSeek."

It was China's second DeepSeek moment — except this time, the breakthrough was not a cheaper model. It was the proof, eighteen months ahead of most of the industry, that autonomous agents could actually work in the hands of ordinary users.

Phase Three: Nine Figures and a Mirage (December 2025–April 2026)

The product velocity matched the hype. In July 2025, Manus shipped Wide Research, running up to 100 parallel sub-agents on a single task. By late 2025, versions 1.5 and 1.6 could generate full web applications from a single prompt. On December 17, 2025, the company announced it had crossed $100 million in annualized recurring revenue — eight months after launch, a commercialization pace with few precedents in software history.

Twelve days later, on December 29, Manus announced it was joining Meta. The reported price: more than $2 billion. The deal had reportedly come together in about ten days. Manus would operate under Meta's umbrella; the Monica product line would continue separately.

From the outside, it looked like the perfect ending: a Chinese team builds a global AI hit, and a American giant pays billions for it. The ending lasted four months.

In April 2026, Beijing's National Development and Reform Commission (NDRC) moved to block the acquisition on national-security grounds. What followed, per the Wall Street Journal and subsequent reporting, was a sequence without precedent in the AI industry:

- The founders were summoned. Xiao Hong and Ji Yichao were called in by the NDRC and informed they would not be permitted to leave the country.

- The Singapore escape was closed. Butterfly Effect had incorporated in Singapore and moved its headquarters there in mid-2025, cutting China-based staff from more than 120 to roughly 40 core members. Beijing's intervention effectively ended that relocation; the operation would stay on Chinese soil.

- The unwind was ordered, not negotiated. Early backers helped the company buy its shares back from Meta at roughly the $2 billion deal value.

Washington restricts what Chinese companies can buy. Beijing, it turns out, restricts what they can sell.

Phase Four: The Unwind

The buyback itself was an oddity — a startup repurchasing itself from Meta at a valuation higher than any of its venture rounds, financed by investors who were, in effect, paying to keep the asset in China. Tencent was reported in July to be in talks to lead the buyback at the $2 billion level, explicitly not seeking control — a posture that reads, in retrospect, as positioning for the larger stake now under discussion.

Between April and August 2026, Manus operated in a strange limbo: technically under Meta's umbrella, practically under Beijing's supervision. The company told users in August to export their data so that anything accumulated during the post-acquisition period could be wiped — an unusual act of data hygiene that doubled as a signal of the divorce becoming final.

On September 1, 2026, the divorce was final. Manus resumed independent operations. The company that had been worth $500 million in April 2025, sold for $2 billion-plus in December 2025, and bought itself back in mid-2026, was itself again — except that everything had changed. The founder could not leave the country. The headquarters could not stay abroad. And the cap table was about to be rebuilt entirely on Chinese money.

Phase Five: Independence Day and the $4 Billion Question

The September round answers the question the whole saga raised: what is an AI agent company worth when geopolitics forbids it from being acquired abroad and forbids its founders from leaving?

The market's answer: $4 billion — roughly double what Meta agreed to pay, nine months after the deal collapsed. The investor list is a map of China's institutional capital: IDG Capital and Boyu Capital from private equity, CATL from industrial capital, and Tencent, already a two-round veteran of the cap table, now in talks to become the largest shareholder. The company is also reportedly considering a restructuring that would pave the way for a Hong Kong IPO.

Place that number against the global agent landscape, and the picture sharpens:

Manus trades at a steep discount to its American peers. But unlike Sierra, Cognition, or Instinct, Manus carries a constraint none of them face: its exit options are closed. It cannot be acquired by an American company — the Meta episode proved that. Its founders cannot relocate. Its data and operations sit under Chinese regulatory jurisdiction by design, not choice.

And yet the investors are lining up. The reasons are structural. Manus retains something money cannot easily rebuild: eighteen months of consumer-agent product experience, a global English-first user base built during the Monica years, a brand that is still one of the few Chinese AI products with genuine international recognition, and — after the September 1 reboot — a clean regulatory story. It is, almost by accident, the only major agent company on Earth that has been pre-cleared by both superpowers: Washington never objected to its existence, and Beijing has now explicitly asserted it as a domestic asset.

What Comes Next

The Hong Kong IPO path is the logical terminus — and the natural next chapter in a story that has been, at every turn, about where a Chinese AI company is allowed to exist. Hong Kong offers something no other venue can: a global investor base, a listing regime that has already absorbed China's AI champions, and a jurisdiction Beijing demonstrably prefers for its strategic tech assets.

The risks are equally clear:

The company also faces the oldest problem in the agent business: agents are expensive to run, cheap to undercut, and sticky only when they are right. The $100 million ARR announced in December 2025 was real, but it was also the product of a scarcity-driven, invite-code-fueled hype cycle that no longer exists. The new round will buy compute, talent, and time. It will not buy a moat.

But the moat question may be the wrong frame. Manus's deepest asset is not technology. It is a story no competitor can copy — and in an industry where distribution, talent, and capital flows are increasingly shaped by geopolitics, the story is the strategy.

What People Are Saying

The Manus saga has produced one of the most divided reactions in the history of Chinese tech — celebrated as patriotic vindication in some quarters, mourned as a cautionary tale in others. Here is what people are saying:

@AI出海笔记 (AI Going Global Notes, WeChat official account): Manus这波操作可以写进教科书了：Meta收了又退，腾讯接盘，40亿美金估值独立上市。肖弘当年拒绝字节3000万美金的收购，现在回头看简直是神之一手。

*"Manus's moves could go in a textbook: Meta bought in then backed out, Tencent stepped up, and now it's heading to IPO at $4 billion independent. Xiao Hong rejecting ByteDance's $30 million offer back then looks like a stroke of genius in hindsight."* — 48k reads

@TechPolicyWonk (X/Twitter): Beijing blocked a $2B exit and the company re-emerged at $4B with Tencent as the anchor investor. Whatever you think of the intervention, the outcome is a Chinese agent champion with a clean cap table and no foreign control. That's the policy working as designed.

*"Beijing blocked a $2B exit and the company re-emerged at $4B with Tencent as the anchor investor. Whatever you think of the intervention, the outcome is a Chinese agent champion with a clean cap table and no foreign control. That's the policy working as designed."* — 12k likes

@硅谷内参 (Silicon Valley Insider, Zhihu): 别光看到40亿估值，看看代价：创始人出不了国，公司回不去新加坡，全球化故事基本上完了。Manus现在是一个纯中国故事，估值却要用全球agent公司的逻辑来撑，这个折价长期看会越来越大。

*"Don't just look at the $4 billion valuation — look at the cost: the founders can't leave the country, the company can't go back to Singapore, and the globalization story is basically over. Manus is now a purely Chinese story priced on global-agent-company logic. That discount only widens over time."* — 8.2k upvotes

@AgentBuilder_Chen (V2EX): 作为一个用过Manus 1.6的开发者，产品确实能打，Wide Research跑100个子agent的体验到现在没有第二家做到。独立之后如果能把API开放出来，国内agent开发者生态会起来一波。

*"As a developer who's used Manus 1.6, the product genuinely delivers — nobody else has matched the experience of Wide Research running 100 sub-agents. If they open up the API after independence, we'll see a wave of domestic agent developer activity."*

@MidasTouch_2026 (X/Twitter): Sierra at $15.8B, Cognition at $10B+, Instinct raising $1B at $10B in the SAME WEEK as Manus at $4B. Either American agents are wildly overpriced, or Manus is the biggest bargain in AI. There is no third option.

*"Sierra at $15.8B, Cognition at $10B+, Instinct raising $1B at $10B in the SAME WEEK as Manus at $4B. Either American agents are wildly overpriced, or Manus is the biggest bargain in AI. There is no third option."* — 21k likes

@资本市场老周 (Veteran Capital Observer, WeChat): 腾讯这轮要是真成最大股东，Manus实际上就是腾讯系agent了。结合微信小薇、元宝，腾讯的agent版图已经比阿里完整。港股IPO之后，中国AI应用的格局就要重新画了。

*"If Tencent really becomes the largest shareholder this round, Manus effectively becomes the Tencent-system agent. Combined with WeChat Xiaowei and Yuanbao, Tencent's agent map is already more complete than Alibaba's. After the Hong Kong IPO, the map of China's AI application layer gets redrawn."*

The Last Word

The Manus story resists the tidy moral that either Washington or Beijing would prefer. It is not a story about a stolen technology, because nothing was stolen. It is not a story about a failed acquisition, because the company that emerged from the wreckage is reportedly worth twice what Meta agreed to pay. And it is not a story about national champions, because Manus became one by accident — a founder tried to sell his company, and his government told him, in effect, that he couldn't.

What it actually is: the first full-dress rehearsal of a conflict that will define the next decade of AI. As agents graduate from demos to infrastructure — as they begin screening résumés, closing sales, writing code, and moving money — the question of *who owns them* becomes a question of sovereignty. Beijing answered that question in April 2026 with an NDRC summons. The $4 billion valuation is the market's answer back: sovereignty, it turns out, is also a business model.

Manus will list in Hong Kong. The founders will stay in China. Tencent will likely anchor the cap table. And somewhere in a Menlo Park or Seattle boardroom, the next American acquirer of a Chinese AI startup is already being told the Manus story in due-diligence slides — as a deal to avoid, or as the cost of doing business in the age of agent sovereignty. Either way, the price has been set.

*Data notes: Funding and valuation figures derive from reporting by Bloomberg, the Wall Street Journal, TechCrunch, and Yahoo Finance citing sources familiar with the matter; Manus has not publicly confirmed the new round. The Meta acquisition price was reported as "over $2 billion." The $100 million ARR figure was announced by Manus on December 17, 2025. Funding history is compiled from public records and prior TechCrunch and TMTPost reporting. The NDRC's role and the founders' travel restrictions were reported by the Wall Street Journal and have not been officially confirmed by Chinese authorities.*

*What do you think — was Beijing's intervention a defense of national assets or a constraint that will ultimately cap Manus's ceiling? Leave a comment below, and subscribe to the AI in China newsletter for weekly deep dives into the people, labs, and capital reshaping the world's most consequential AI race.*