Manus was a Singapore company until it wasn't

The National Development and Reform Commission published a notice on Friday ordering Meta Platforms and Butterfly Effect Pte. Ltd., the Singapore holding company that owns the Manus agent product, to unwind their previously announced acquisition. The notice runs to four pages in the Chinese version, two in the English translation appended to the same PDF on the NDRC site, and contains a sentence I want to read carefully:

"Notwithstanding the foreign domicile of the target entity, the Commission finds that the substantive research, development, and model-training activities constituting the value of the acquired assets were conducted within the territory of the People's Republic of China, and accordingly fall within the scope of [the 2024 outbound technology transfer measures]."¹

Manus is, on paper, a Singaporean company. It was incorporated in Singapore in 2023, raised its Series A from Benchmark out of a Singapore entity in 2024, and has consistently described itself in press materials as Singapore-headquartered. Its founders are Chinese nationals; its engineering team, by every account I can find, sits in Beijing and Wuhan; its model weights, and this is the part the NDRC notice turns on, were trained on infrastructure physically located in mainland China.²

The deal, announced 6 February at a reported $3.4bn (cash and stock, with $900m in retention), was structured as a clean offshore acquisition: Meta acquires Butterfly Effect Pte. Ltd., Butterfly Effect continues to own the Manus IP, the China-resident engineering team continues to operate under a services agreement with the Singapore parent. This is the standard VIE-adjacent structure that has been used to move Chinese AI talent and IP across the border for the better part of two years. The bet, Benchmark's bet, Meta's bet, and the bet of roughly a dozen similar deals sitting in various stages of completion, was that Singapore incorporation plus offshore weight storage plus a services-agreement engineering relationship would put the asset outside the reach of Chinese outbound review.

The NDRC notice says: no.

The frame here is model weight lineage, which I have been writing about for eighteen months as the IP question that doesn't fit cleanly into any existing regulatory category. Weights are not patents. They are not trade secrets in the conventional sense, they cannot really be reverse-engineered from products, but they can be copied wholesale if you have access to the file. They are not contracts. They are a thing of value created by a specific training run on specific data with specific compute, and the question of who owns them and where they live has been, until now, genuinely unsettled in regulatory terms.

The NDRC has now answered the question in one direction. The notice's logic is straightforward: if the training run happened on PRC infrastructure, with PRC-resident researchers, on data that touched PRC networks, the resulting weights are a PRC asset regardless of which holding company's server they sit on. Butterfly Effect Pte. Ltd. owns the weights as a matter of corporate law; the NDRC is asserting that the weights are a matter of technology transfer regulation, which is a different body of law that doesn't care about corporate law.

This is not, in itself, a surprising regulatory move. It is the obvious move, and the surprise is that it took until April 2026. What is more interesting is the second sentence, which appears in a separate "guidance" document published on the same day:³

"Domestic enterprises engaged in artificial intelligence research, model development, or related infrastructure shall, prior to accepting investment, financing, acquisition offers, or strategic cooperation arrangements from United States persons or entities, submit the proposed transaction to the Commission for review."

Read this carefully. It is not "shall obtain approval before completing." It is "shall submit for review prior to accepting." The chilling effect, and it is plainly intended as a chilling effect, runs to the term sheet rather than the close. Chinese AI companies cannot now sign a term sheet with a US investor without first clearing it with Beijing, which means the term sheet doesn't get signed, which means the round doesn't happen, which means a quite large amount of capital that has been routed from US LPs through Singapore and Cayman vehicles into Chinese AI companies stops being routable.

The numbers here are not trivial. By PitchBook's count, US-affiliated capital into Chinese AI companies in 2025 ran to roughly $4.1bn across 73 disclosed rounds, with another estimated $1.5–2bn undisclosed via offshore structures.⁴ Benchmark, Sequoia (the post-split US side), General Catalyst, Lightspeed, and a long tail of US growth funds have all been active. None of them have publicly commented on the NDRC notice as of Friday evening London time, which is itself a piece of information.

The AI safety as market position frame applies here in a way that is slightly counter-intuitive. The US side of this, CFIUS, the Bureau of Industry and Security, the various outbound investment review regimes, has been moving in this direction for two years, and has produced a market in which "we don't take Chinese capital" is a sellable enterprise narrative for US frontier labs. What the NDRC has now done is produce the symmetric version: "we don't take US capital" becomes a regulatory requirement for Chinese AI companies, which forecloses a set of exit paths and forces a different capital structure. The two regimes are now, functionally, partitioning the global AI capital market along the same line, from opposite directions, which is the intelligence explosion signals frame doing exactly what it predicts: the race dynamic produces symmetric capital controls before it produces symmetric export controls on the underlying technology.

What this is a case of: the closing of the Singapore arbitrage. For roughly two years, Singapore-domiciled holdcos with PRC-resident engineering teams have been the primary mechanism for routing Chinese AI talent and IP into structures that could accept US capital and contemplate US exits. That window is now substantially narrower. Manus is the visible casualty; the more interesting question is the dozen-or-so similar structures sitting in due diligence right now.

What to watch:

Whether Meta and Butterfly Effect attempt to comply by spinning out the China-resident engineering team into a separate PRC entity and licensing weights back, the obvious workaround, and one the NDRC notice appears to anticipate and foreclose, but which someone will try.

Whether other Singapore-domiciled Chinese AI holdcos disclose NDRC review proceedings in the next 30 days. The notice's logic applies to every one of them.

Whether US-affiliated funds with active Chinese AI positions begin marking those positions down. Benchmark's Manus stake was last marked at $480m on a $3.4bn round; it is now an asset of disputed ownership in a deal that has been ordered unwound.

The structural story is that the weights have a nationality. Everyone suspected this. The NDRC has now said it out loud.

Footnotes

¶