You’re a Chinese AI startup. You’ve shipped a product, you’ve hit ARR, your users are mostly overseas.
What’s next.
If you just want to keep running, you can keep doing what you’re doing. But if you want to scale, push to the next level, or give your earlier investors something to point at, sooner or later you’ll face the question every founder has to answer: how do you exit, or how do you raise exit-level capital to keep building.
To take in foreign capital, the standard move is to put your holding company outside China. The most common option for the past two decades has been the Cayman Islands — the classic red-chip structure: a Cayman parent, a Hong Kong sub, and onshore VIE arrangements. Regulators are familiar with it, there’s twenty years of case law, the market accepts it. MiniMax went this route, listing on the Hong Kong main board on January 9, 2026, jumping 109% on day one to a market cap around HK$300 billion. Moonshot AI is in early talks with CICC and Goldman about a Hong Kong listing, with valuation rising from about $4.8B at the start of the year to roughly $18B; its main entity is still Cayman. StepFun’s Cayman structure is being unwound as of early April 2026, redirected toward a direct domestic H-share filing.
Singapore is a newer option that’s emerged in the last two years. Unlike Cayman, the Singapore route is more aggressive on regulators and on buyers. The upside: it lets you transact with a buyer at Meta’s level. On the Cayman path, CFIUS asks questions, export controls ask questions, the buyer’s own board asks questions. The Singapore path lets you handle all of that before the deal even starts.
Up until Meta’s announcement of the Manus acquisition in January 2026, both Cayman and Singapore were legally permissible structures. The Cayman route’s ceiling is a Hong Kong IPO retail market cap, in the HK$300-400 billion bracket. The Singapore route’s ceiling is an outright acquisition by a Meta-class global tech giant. Manus took the Singapore route. It’s the only Chinese AI company to walk that route all the way to closing.
What Manus did in June 2025 was execute the Singapore playbook the most thoroughly anyone had. They moved their global HQ from Beijing to Singapore. The operating entity became Butterfly Effect Pte. Ltd. In July, core engineers started relocating to Singapore. Service inside mainland China was discontinued in parallel. At that moment, this set of moves was clean, lawful, and clearly targeted: it prepared the company to transact with a Meta-class buyer.
Reading this, you might wonder: shouldn’t Manus have gone Cayman like MiniMax and listed in Hong Kong instead? It’s a fair question, but the answer doesn’t lie in the cleverness of the compliance strategy. Companies on the Cayman route don’t get unsolicited bids from Meta. The Cayman structure doesn’t clear CFIUS or buy-side legal. MiniMax and Moonshot are pursuing Hong Kong partly because Hong Kong is the only exit available to them. Manus is one of the few Chinese AI companies that received an unsolicited bid from a global tech giant. On December 30, 2025, Meta offered “just over $2 billion” to acquire Manus outright — Meta’s third-largest acquisition in its history, behind only WhatsApp and Scale AI. That bid itself is a signal of international market recognition for the company’s technical capability. It’s also a form of exit unique to the Singapore route.
To be precise, this isn’t a simple “IPO got lucky, M&A got unlucky” comparison. Being acquired by a global giant means the technical ceiling is higher. And precisely because the ceiling is higher, you need a more aggressive offshore route to match it. That route doesn’t have Cayman’s twenty years of precedent underwriting it, and the rules can be rewritten before you finish walking it. The fastest movers, the ones who go furthest, are the first to hit the new wall.
Set aside the block itself for a moment. Chinese opinion has long pinned a different label on Manus, unrelated to this block but directly related to the company’s technical substance.
That label is the “just a wrapper” thesis, attached starting the week Manus launched in March 2025. The claim was that Manus was just a Claude/GPT wrapper with a thin agent layer over it — no in-house model, no real engineering, valuation a bubble. That label hung on the company from launch through the Meta acquisition, almost a full year.
The NDRC’s action this week, incidentally, took that label off.
The logic is simple. If Manus were a wrapper product with no real China-rooted technical IP and no transferable core R&D capability, Meta wouldn’t be paying $2B for a shell, and the NDRC wouldn’t have invoked the Measures for the Security Review of Foreign Investment — for the first time in the regulation’s five-year life — to issue a “prohibit and unwind” order. The FISR (Foreign Investment Security Review) is a tool for blocking “foreign acquisitions of critical technology with national security implications.” Five years in, the NDRC reached for it for the first time, used it on Manus, and chose the strictest setting. The regulator concluded that this company’s core team, R&D capability, training data, and IP constitute national-security-grade assets. You only deploy the strongest legal tool when there’s something real to protect.
This block, then, can be read as carrying a meaning that isn’t on the announcement but is objectively present: a stamped-and-sealed certificate of technical substance. That’s a level of recognition no domestic technical evaluation, media debate, or investor diligence could deliver.
The week Manus launched in March 2025, I wrote a piece arguing that Manus didn’t appear out of nowhere. It demonstrated systematic organizational capability that peer products couldn’t match across three dimensions of compounding for Agentic AI: tools, data, and intelligence. The hardest thing to copy isn’t data scale or tool count, it’s the methodology of externalizing tacit knowledge, organizing it, and depositing it into AI-friendly structures. On GAIA Level 3, Manus scored 57.7% at the time, beating OpenAI Deep Research’s 47.6%. Eight months later, the company was running at $100M ARR, $125M+ run rate, 147 trillion tokens processed, and over 80 million virtual computers created.
That earlier piece was a product observation. Today, the NDRC’s action provides reverse confirmation at the regulatory level.
Place this against the longer timeline of regulation and capital, and there’s another fact that’s not immediately obvious.
The first “first” is on the exit-path side. No Chinese AI startup had ever been acquired outright by a US tech giant before Manus. Manus is the first one to walk that path all the way to closing. Not a term sheet, not a letter of intent — a signed SPA, a public announcement, proceeds distributed to shareholders.
The second “first” is on the regulatory side. The Foreign Investment Security Review went into effect on January 18, 2021. Five years to the Manus case. In those five years, the NDRC never issued a standalone “prohibit and unwind” announcement. White & Case observed in a 2021 review that while the mechanism was institutionally in place, actual interventions happened mostly through “non-public suggestions to the applicant to withdraw or restructure” — public denials hadn’t surfaced. The Manus case is the first time in five years that this legal tool was taken all the way through the process and announced publicly, and the strictest setting was chosen.
These two firsts are the same company. The first company to walk this path is also the first company to have this path shut down on it.
Before Manus, the route of “redomicile to Singapore + acquisition by a foreign tech giant” had never been formally blocked. After Manus, it isn’t allowed. The regulator and the entrepreneur, together, drew the line through this single transaction. As the Lowy Institute put it, Manus “followed this playbook precisely”. They walked the very path that everyone, at the time, considered viable.
The NDRC’s April 27 announcement was very short. No reasoning was given; only the Foreign Investment Security Review was cited. Chinese-language analysis converged on the same reading. The clearest articulation came from Guancha:
If this reading holds, the real target of this block isn’t the Meta deal as a transaction. It’s the entire compliance playbook Chinese AI companies have used to go global over the past few years: redomicile abroad, transfer the core team and IP out of PRC-domiciled entities. The regulator has now stated publicly that no matter how thoroughly that playbook is executed, it remains within the jurisdiction of Chinese law.
This block is nearly impossible to enforce in practice. CFIUS in the past decade has done a few similar forced unwinds — for example, in 2019, Kunlun was ordered to divest Grindr (sold to San Vicente Acquisition for ~$608.5M in June 2020). In those cases, what was being divested was tangible: app, user data, servers, brand, all listable for sale to a buyer. Manus is different. The core value of an AI agent is in code, model weights, agent framework, and the know-how in engineers’ heads. Once those have been seen, used, and integrated by the acquirer, there’s no independently divestible “asset bundle” to sell. This layer is well covered in an article carried by Market News from “uncle Leslie” (since taken down across Chinese platforms).
What the regulator finally produced isn’t asset recovery, it’s a policy precedent. Putting the signal “this path — packaging up a Chinese AI company via Singapore for sale to a foreign tech giant — is not allowed” out in public, where every founder and every buyer can see it.
Worth viewing alongside this is StepFun’s move. StepFun was originally a Cayman entity. On April 2, 2026, it completed a corporate restructuring and changed its name to “Shanghai StepFun Intelligent Technology Co., Ltd.”, preparing to unwind its Cayman structure and file a domestic H-share IPO. The two events sit close together: April 2, the Cayman route starts being unwound under regulatory pressure; April 27, the Singapore route is publicly blocked. The two offshore corridors that Chinese AI companies have defaulted to over the past few years are narrowing in the same month. Manus is the most thoroughly and most publicly tightened of the two.
One clarification first. The NDRC’s judgment is at the level of the path. It targets whether “AI capability developed inside China being transferred to a foreign tech giant via a Singapore redomicile” should be permitted under national security considerations. It does not target the personal commercial motives of the Manus team. The regulator is blocking the path.
Manus is a specimen caught between two regulatory regimes. Benchmark’s Series B investment was, as of May 2025, already under review by the US Treasury under the inverse-CFIUS regime (Outbound Investment Security Program), the first publicly reported OISP case. Meta’s acquisition is the third move in Mark Zuckerberg’s Superintelligence Labs sequence: $14.3B for 49% of Scale AI (Alexandr Wang / data), then $2B for outright acquisition of Manus (Xiao Hong / agent). What Meta was buying is a globalized agent product with real ARR and a working team.
Xiao Hong and Ji Yichao made nearly every compliance call available to them. Accepting Benchmark’s lead in April 2025 was a read on the window where US capital would offer the highest valuation for a Chinese AI globalization story. Proactively redomiciling to Singapore in June 2025 was a read on the window for transacting with a Meta-class buyer not staying open forever. Walking away from the China market was a read on ARR ceilings being abroad, not domestic. These calls were a minority position in China’s AI scene in 2025. Most peers chose the gentler Cayman route — that path leads to Hong Kong market caps, but doesn’t yield outright bids from global giants. Manus chose the path that connects to top-tier buyers, and executed the compliance moves on that path the most cleanly.
The cost they’re bearing didn’t come from any wrong judgment. It came from time. The path they walked was permitted in mid-2025 and not permitted by April 2026. The time it took to walk that path passed through, exactly, the moment that rule was rewritten.
For Chinese AI startup teams still raising, preparing to go global, or making decisions about where to incorporate, this case offers a few specific judgments to use right now.
First, capital structure and incorporation choices are strategic from day one, not compliance items to handle before exit. “Build the product domestically, then redomicile” — as a single-step compliance strategy — has been shown to be insufficient by the Manus case. The regulator’s framework is piercing: registered location, contracting party, ultimate controller — none of these are immune layers. As long as the chain (core team, R&D capability, training data, IP) was transferred out of PRC-domiciled entities, the regulator considers it within jurisdiction. The only way to break this chain is to keep it from being inside PRC entities in the first place. That means moving the timing of building an overseas team and overseas revenue forward — to before fundraising, ideally before product launch.
Second, the realistic exit options have collapsed from two to one. Before Manus, both the Cayman path (Hong Kong IPO) and the Singapore path (foreign tech giant acquisition) were on the table. After Manus, the Singapore path is publicly blocked by the NDRC. The Cayman path itself is also tightening: StepFun started unwinding Cayman in early April, Moonshot is evaluating whether to follow, and the CSRC’s direction is to push red-chip structures toward direct domestic H-share filings. The remaining stable exit is a domestic shareholding company filing directly for H-shares. That path makes a clear demand on day-one structure: be a domestic entity from the start, don’t go Cayman. In 2025 this was just one option among several. From April 2026, it’s closer to default.
Third, USD funds haven’t exited China AI. Hillhouse launched a new ~$7B fund in late 2025. Sequoia, IDG, and Hillhouse continued to deploy actively in domestic AI front-line rounds through 2025-2026 — Robot Era, Tashi Auto, NewLanguage among the cases that closed. But the exit channel for USD funds has converged from “foreign tech giant acquisition” to “Hong Kong IPO + domestic M&A.” The implications for LP composition and post-investment management are significant, but the direct effect on overall fundraising activity in the next few years is probably smaller than people imagine.
Fourth, what the regulator actually produces is a policy precedent, not asset recovery. The AI agent form factor has no independently divestible asset bundle. The CFIUS-style 4-to-14-month windows are, in this case, effectively unenforceable. What to watch next is not “how Manus gets unwound” — that’s going to be complicated, drawn out, and very likely never fully completed in execution — but “at what moment the next similar deal gets stopped.” If the regulator really does start treating “act at the moment of redomicile” as the new enforcement default, then Chinese AI startup teams after Manus aren’t facing a 4-month review window, they’re facing a day-one structural decision.
My personal read is that this block will be cited repeatedly but won’t immediately become enforcement routine. Sample size is still 1. The regulator’s real intent is to put the signal out publicly and let the market self-adjust, not to throw enforcement resources at every cross-border AI transaction’s pre-clearance. Over the next 12-18 months, a wave of Chinese AI startup teams will adjust capital structure and incorporation paths on their own. The trigger for that adjustment isn’t new enforcement, it’s that investor diligence checklists and risk pricing have already absorbed the Manus case.
For builders, the real question isn’t “can I go to Singapore.” It’s “at what level of success will the regulator stop letting me go to Singapore.”
Back to the company itself. Eight months from product launch to $100M ARR. Producing what the Lowy Institute called the specimen for “the first major acquisition of a Chinese AI company by a US tech giant.” The difficulty and quality of this work is something no other Chinese AI startup team has matched in the past three years. What was blocked is the compliance assumption beneath the path. The product, the team, the capability — those remain.
Part of the cost the Manus team is bearing right now, they’re bearing for this generation of Chinese AI founders. They walked the whole path so the people behind them know where the path is, and where it isn’t.
There’s a Chinese saying: those who came before planted the trees, those who came after enjoy the shade. The Manus team planted this tree the hard way. The shade is what the next generation gets to sit in.