China Yanks Manus Back From Meta. The Singapore Exit Just Got Harder.

China just told Meta to give Manus back.

The order came down Sunday from the National Development and Reform Commission, the same office that signs off on foreign investment reviews. Meta closed its $2 billion-plus acquisition of Manus in December. China opened a security review days later. Four months on, the NDRC said it would "prohibit foreign investment in Manus in accordance with laws and regulations, and requires the parties involved to withdraw the acquisition transaction."

If you think this is normal, it isn't. Reuters notes that China rarely orders completed deals reversed. Block them ahead of close, sure. Drag them out until everyone gives up, sometimes. Forcing a buyer to actually undo a closed transaction is a different category, and the message it sends is the point.

A few things make this story worth pulling apart, because the surface read (China gets prickly about a US acquisition) misses what's actually moving.

Manus isn't even a foundation model

This is the part I keep circling back to. Manus, made by parent company Butterfly Effect, is an AI agent framework. It runs on top of existing Western large language models. They didn't train their own GPT-class model. They built a layer that orchestrates other people's models to do useful things autonomously.

That China cares enough to claw back a foreign acquisition of an agent framework, not a frontier lab, says something about how Beijing thinks about the AI stack now. The agent layer is where workflows live and where customer data accumulates. It's the actual product surface. If you're playing for long-term technology sovereignty, losing the agent layer to a US buyer is arguably worse than losing a model trainer, because the model trainer can be replicated and the operational know-how cannot.

Per CNBC, Manus claimed to cross $100M ARR in December, eight months after its general AI agent launched. That's not a research project. That's a real business with real customers, and Beijing has now decided it should stay Chinese in some meaningful sense.

The Singapore washing problem

For about a year there has been a quiet pattern: Chinese AI startups raise money, then move headquarters to Singapore. Re-incorporate, move staff, get a Singapore registration, and you exist in a regulatory zone that lets US capital invest and lets you use US-export-restricted models without the same friction. Manus did exactly this. After a $75M round led by Benchmark in May 2025, they shut their China offices in July and moved operations to Singapore. Butterfly Effect re-incorporated there. By the time Meta's bid landed, Manus was, on paper, a Singapore company.

This worked because both sides preferred not to look too hard. US investors got access to interesting AI talent coming out of China without tripping the new outbound investment restrictions. Chinese founders got Western capital and Western models without explicitly asking Beijing for permission. Singapore got the office leases.

Beijing just made it clear that the paper move isn't enough. The Manus founders, Xiao Hong and Ji Yichao, were summoned to Beijing in March. According to Reuters sources, they have been barred from leaving China since. The staff is in Singapore. The founders are in Beijing. The deal is being unwound. That is a very specific message.

Ben Chester Cheong put it well in the Reuters piece: "I would not say this ends Chinese companies moving to Singapore. Rather, it raises the compliance threshold." To actually be Singaporean for purposes of avoiding Beijing's reach, he said, you need a real operational shift. Where management sits. Where the IP is owned. Where R&D is conducted. Where data is stored. Re-incorporating in Singapore while your founders fly back to Beijing is no longer enough cover.

For founders weighing this move now, the calculation got harder. You can still leave. You might not be able to leave halfway.

What Meta was actually buying

Meta's interest in Manus was not subtle. The agent layer is the next competitive surface in AI, and Meta does not have a leading product there. Buying Manus was a way to skip a couple of years of building. Spend $2 billion, get a working AI agent framework with paying customers, fold it into Meta's stack. Reuters and CNBC both report that staff had already moved into Meta's Singapore offices and projects were continuing despite the exit bans on the founders.

That bet now blows up. Meta has previously said the acquisition complied fully with applicable law. That probably won't matter. The question is what happens next: does Meta fight this, write off $2 billion, or quietly unwind and pretend the whole thing didn't happen? My guess is the third option, because whatever value Manus had inside Meta's roadmap depended on the deal closing cleanly, and that ship has sailed.

The harder question is what happens to Manus itself. The product has paying customers and revenue. The founders are stuck in China. The staff is in Singapore. If Beijing wants Manus to stay Chinese, someone has to actually run it, and the people in the best position to do that are the ones who can't leave the country.

What this means for cross-border AI deals

A few things I think are worth holding onto:

• US firms acquiring AI startups with any China provenance now have to assume Beijing gets a vote, even if the target is "Singapore-based." That assumption applies during diligence, not at close. The Meta deal closed cleanly and Beijing reached in afterward.
• The category of "AI agent companies" is now treated by China the way semiconductor companies were treated five years ago. Anything in the AI value chain that touches data, workflows, or customer integration is potentially in scope.
• Singapore is still useful, but as Cheong said, the threshold went up. Genuine operational independence has to look genuine. Where do the engineers live, whose name is on the IP filings, where do the servers sit. "Singapore on paper" no longer holds.

For the Chinese AI ecosystem itself, this is going to chill some founders who were considering the same move. Watching Manus go through this in real time, the prospect of being summoned back to Beijing while your staff sits in Singapore is going to make people think very hard about whether the Western capital is worth the personal exposure.

The part that surprises me

I would have expected China to block this deal at the review stage in January, not let it run for four months and then unwind it. The unwind sends a stronger signal but it also creates real operational chaos for a company with paying customers. Beijing was willing to pay that cost to make the example.

Alfredo Montufar-Helu, quoted in the Reuters story, summed up the policy logic: "China is saying we will prevent foreign acquisition of assets we consider important for national security, and AI is now clearly one of them."

What I'd add is that "national security" used to mean models. Now it includes the layer that wraps the models and turns them into products. That is a meaningfully wider net than the one that existed even a year ago, and US firms shopping for Chinese AI assets are going to be operating under it whether or not they realize the assets are Chinese on the day of signing.

Manus, of all companies, was the test case. An agent framework, headquartered in Singapore, sold to an American buyer. Beijing pulled it back. If you build cross-border deal models for a living, this is the data point you redraw the curve around.