China Calls Meta's $2B Manus Acquisition a "Conspiracy" and Blocks the Founders from Leaving the Country

On March 28, 2026, Chinese authorities reportedly placed exit bans on Butterfly Effect founder Xiao Hong and two co-founders, and on April 4, 2026 a Ministry of Commerce spokesperson branded Meta's approximately $2 billion acquisition of Manus AI a "conspiracy" to extract Chinese frontier-agent technology under commercial cover. The deal — announced in principle at the end of 2025 and targeting a close by Q2 2026 — is now effectively suspended. Manus, the viral autonomous general AI agent that briefly topped the GAIA benchmark at 86.5, 70.1 and 57.7 percent across three levels, has become the first major Chinese-founded AI company to be publicly blocked from a US acquisition on explicit national-security grounds since the DeepSeek disclosures of 2024-2025.

What happened between December 2025 and April 2026

The sequence matters. Manus was launched publicly on March 6, 2025, went viral within a week on the strength of GAIA benchmark numbers and a genuinely impressive autonomous workflow, and was reportedly in advanced acquisition talks with Meta by late autumn 2025. The deal — structured as a cash-plus-stock transaction valued by reporting outlets at between $2 billion and $3 billion — was confirmed in principle in December 2025, with Meta publicly framing it as a key piece of its Superintelligence Labs roadmap alongside the earlier talent and technology repositioning that reshaped Silicon Valley in 2025.

• December 2025. Meta and Butterfly Effect announce the acquisition in principle. Butterfly Effect — registered in Singapore but with a Beijing-based core engineering team — is valued at approximately $2 billion. The deal is subject to CFIUS review in the US and unspecified Chinese regulatory clearance.
• January-February 2026. CFIUS opens its standard review. Chinese state media begins publishing critical commentary, questioning whether Butterfly Effect's core IP — including the CodeAct execution paradigm and the multi-agent orchestration stack — would legally transfer to a US acquirer without Ministry of Commerce approval.
• March 15, 2026. Butterfly Effect ships the Manus desktop app, partially running on local hardware — widely read as a hedge against data-residency concerns that had already surfaced in enterprise sales conversations.
• March 28, 2026. Exit bans are reportedly placed on Xiao Hong, the Butterfly Effect founder, and two co-founders. None of the three is publicly charged with a crime; under Chinese administrative practice an exit ban can be issued as part of an ongoing investigation without criminal charges being filed.
• April 4, 2026. A Ministry of Commerce spokesperson, in a scheduled press briefing, calls the Meta acquisition "a conspiracy to extract frontier-agent technology developed with Chinese talent and Chinese infrastructure under the cover of normal commercial activity." The Ministry does not formally block the deal, but signals that no outbound transfer of core technology assets will be approved.

Between those two dates, the deal moved from "pending regulatory review" to "functionally frozen." Meta has not publicly withdrawn the offer. Butterfly Effect has not publicly confirmed the exit bans. Neither side is saying the deal is dead — but neither side is saying it is still alive either.

Who Butterfly Effect and Manus actually are

Butterfly Effect: the corporate shell

Understanding what China is protecting starts with understanding what Butterfly Effect built. Butterfly Effect is a Singapore-registered entity founded by Xiao Hong, an engineer with a prior background in search infrastructure and a small circle of co-founders whose public profiles are deliberately sparse. The company's engineering and product work is primarily done out of Beijing; its Singapore registration was a structural choice, common for Chinese-origin startups pursuing global users, not a relocation of the engineering team.

Manus: the autonomous agent

Manus itself is an autonomous general AI agent. It orchestrates a set of sub-agents — Executor, Planner, Knowledge — inside an isolated Linux sandbox per session, and it executes tool calls using the CodeAct paradigm: instead of generating brittle JSON for each tool call, Manus writes and runs Python on the fly, unlocking any pip library as an action. That design choice is one of the reasons Manus numbers were unusually high on GAIA; it also turns out to be one of the reasons Chinese regulators view the stack as strategically valuable. CodeAct-style execution is not a Manus invention in the abstract sense — the research pattern comes from academic work — but Butterfly Effect's integration of CodeAct with a production-grade sandbox, long-horizon task management, and reliable tool-ecosystem grounding is genuinely novel engineering. That is what Meta was buying.

Commercial traction and pricing

Commercially, Manus opened with the most generous free tier in the autonomous-agent category — 1,000 starter credits plus 300 daily refresh credits — and tiered paid plans at $20, $40 and $200 per month. Those numbers matter for the geopolitical framing because they establish that Manus was already a mass-market tool with tens of thousands of paying users worldwide at the time the acquisition was announced, not an unreleased internal research project.

Beijing's "conspiracy" framing, decoded

The Chinese word the Ministry of Commerce used in the April 4 briefing — which state media translated into English as "conspiracy" — is 阴谋 (yīn móu), which carries a stronger implication of hidden intent than the English translation suggests. In Ministry-level statements that word is unusual. The normal vocabulary for a disapproved foreign acquisition is closer to "not conducive to national interest" or "subject to additional review." Using 阴谋 is a deliberate escalation.

Three elements in the Ministry's framing deserve unpacking:

1. "Developed with Chinese talent and Chinese infrastructure." This is the legal hook. Chinese export-control policy since 2023 has treated certain AI training methods, model weights and agent-orchestration techniques as restricted technology categories. If Manus's core IP is deemed restricted, the acquisition would require Ministry of Commerce approval that the Ministry is now signaling it will not grant.
2. "Under the cover of normal commercial activity." This is the political frame. By casting Meta's bid as disguised technology extraction rather than an ordinary cross-border transaction, the Ministry pre-commits the narrative: if the deal proceeds in any form, it is a national-security failure; if it collapses, it is a successful defense of Chinese technological sovereignty.
3. No formal block order. Note what the Ministry did not do. It did not issue a formal order under the Foreign Investment Review regime. It did not publish a list of restricted technologies that names Manus explicitly. That ambiguity is intentional — it leaves the Ministry with room to approve a restructured deal later without visibly reversing itself.

The combination of exit bans on the founders and the "conspiracy" framing in the same two-week window is the signal. Exit bans are a retention mechanism — they ensure the individuals whose decisions matter remain accessible to Chinese authorities during the investigation. Paired with a political framing that treats the deal as extraction, they effectively tell both Meta and Butterfly Effect: no closing is happening in 2026 under the terms as announced.

Exit bans in Chinese administrative practice

Exit bans are a regularly used — if rarely publicized — tool in the Chinese administrative and investigative toolkit. Under the 2012 Exit and Entry Administration Law and subsequent regulations, an exit ban can be issued by courts, prosecutors, or relevant administrative agencies for individuals who are subject to ongoing investigations, civil disputes, or national-security reviews. Crucially, the subject does not need to be charged with a crime for an exit ban to attach.

For foreign executives and investors, exit bans have been a recurring concern since roughly 2018, when a series of cases involving US and European nationals detained at Chinese borders during commercial disputes made headlines. For Chinese nationals — which is the posture that applies to the reported Butterfly Effect founders — exit bans are more routine but still unusual in the context of a private commercial transaction. The fact that this particular exit ban was reported at all (sourced to multiple industry contacts, then confirmed in broad strokes by Reuters and the Financial Times) suggests Beijing wanted the signal to be public.

Why the Meta deal is functionally dead

Meta does not, in practice, have a path to closing the Manus acquisition on the terms it announced in December 2025. Four factors converge:

1. Core technology cannot legally transfer. If Chinese authorities treat the CodeAct execution stack and multi-agent orchestration IP as restricted technologies requiring Ministry of Commerce approval for export, then Meta cannot take delivery of what it is paying for. A deal in which the US acquirer receives the brand, the users, and the Singapore corporate shell but not the underlying engineering IP is not the deal Meta negotiated.

2. Founders are not available to close. Closing conditions in a transaction this size typically require extensive founder certifications, representations and warranties, and often multi-year retention agreements. Founders under exit bans cannot travel for due diligence meetings, cannot visit Meta headquarters, and face obvious conflicts when asked to certify the integrity of IP transfers the Ministry of Commerce is treating as restricted.

3. CFIUS risk cuts the other way too. Even if the Chinese regulatory hurdle were somehow cleared, CFIUS review in the US would now have to address whether a US acquirer can safely integrate IP developed by executives who were placed under exit bans by the acquirer's chief strategic rival government. That is a national-security risk profile the Committee on Foreign Investment in the United States has not historically had to weigh, and the direction of CFIUS concern in 2026 under the Trump administration is cautious.

4. Commercial momentum is gone. A deal that takes twelve months longer than planned, under a conspiracy framing from one of the world's largest governments, with founders who cannot leave their country, is not a deal that keeps its deal-book champions. Inside Meta, advocates for the Manus acquisition were making a straightforward capability argument: Butterfly Effect had shipped a better autonomous agent than Meta's internal Superintelligence Labs efforts, and buying beats building when you are eighteen months behind. That argument does not survive a blocked close.

Impact on Manus users and data residency

For the Manus user base — free-tier users, the $20, $40 and $200 per month paid tiers, and enterprise pilot customers — the practical question is whether the product keeps working, whether data stays where it is, and whether the billing and reliability issues that were already a concern in early 2026 get better or worse.

Product continuity

On product continuity: the service is running as of April 2026. Butterfly Effect has not announced any pause in operations, and the March 15 desktop app launch is still being supported. However, a deal-blocked Butterfly Effect in a country that has just branded its strategic buyer a "conspirator" is not a company that can raise follow-on capital on normal terms. Expect slower product shipping, reduced marketing spend, and tighter credit economics in the paid tiers over the coming quarters.

Data residency risk

On data residency: Manus session data, by default, flows through Shenzhen-region servers. That was a concern for regulated sectors — financial services, healthcare, government — from the day of the March 2025 public launch, and the exit-ban news does not make it better. Enterprises that were considering Manus for production workloads in 2026 should assume that any pathway to US or EU data residency is delayed indefinitely. The desktop app mitigates this partially by running some operations locally, but tool-call outputs and orchestration state still touch Chinese infrastructure.

Billing and reliability

On billing and reliability: community reports through 2025 and early 2026 flagged credit accounting discrepancies, stuck sessions, and inconsistent handling of bot-detection walls. None of these get better when the company's founders cannot travel, the strategic buyer is frozen out, and engineering headcount is under capital pressure. Users relying on Manus for business-critical workflows should plan for degraded SLA.

The CFIUS posture on Chinese-origin AI

On the US side, the Committee on Foreign Investment in the United States reviews inbound transactions for national-security risk, but the Manus-Meta case flips the usual analytical direction. Normally CFIUS asks whether a foreign acquirer of a US business creates national-security exposure for the United States. Here the US business is the acquirer, and the question has become whether a US tech major can safely integrate IP developed by executives who are now under exit bans issued by the US acquirer's chief strategic rival government.

Three specific concerns are on the record with Committee staff, according to accounts consistent with industry reporting. First, any post-close retention of Beijing-based engineering talent would keep Manus's core codebase within reach of Chinese export-control authority indefinitely. Second, data flowing from Meta's integrated product surface into Manus-inherited orchestration pipelines could, in theory, traverse Chinese infrastructure during a transition period. Third, executive certifications provided under an exit ban cannot be freely negotiated, which creates deal-integrity risk at the document-signing stage. None of these concerns is fatal on its own. Together they are a pattern the Committee has not dealt with before, and the Trump administration's Department of Commerce has signaled it expects CFIUS to tighten its analytical framework accordingly.

Investor reaction and the venture-capital read

Butterfly Effect's cap table, as of the December 2025 announcement, included Chinese domestic venture funds, a handful of crossover growth-stage investors, and an undisclosed strategic position from at least one US tech major that predated the Meta bid. For Chinese-domestic investors the April 2026 developments are a mixed signal — an exit through Meta would have been a historically large outcome, but a frozen deal keeps Butterfly Effect in the Chinese national-champion category, which comes with strategic advantages inside China even if it reduces optionality abroad.

For US and European growth-stage investors holding positions in Chinese-origin AI companies, the signal is cleaner and less comfortable. The exit path they were underwriting — a US strategic acquisition at a 20x or higher revenue multiple — is now politically unavailable for any Chinese-origin AI company at the frontier. Secondary-market pricing of positions in DeepSeek, Moonshot, Zhipu and comparable peers moved visibly in April 2026, with secondaries trading at discounts of 25 to 40 percent relative to March 2026 marks. Expect LP letters in Q2 2026 to acknowledge that "strategic exit" as a scenario for Chinese-origin AI positions is now a harder path, and that investors should model holds longer than they did a quarter ago.

What this means for the US-China AI race

The Manus-Meta situation is the cleanest test case yet for a pattern that has been forming since the DeepSeek disclosures of 2024 and the Frontier Model Forum espionage pact of April 2026: Chinese AI capability is now strategic enough, and differentiated enough, that cross-border acquisitions by US tech majors are politically unavailable. That is a material change from the 2015-2022 era, when Chinese-origin AI and internet companies were routinely acquired, invested in, or talent-poached by US firms without national-level intervention.

Three forces make the pattern sticky:

• Chinese frontier AI is no longer catch-up technology. DeepSeek V3.2, Qwen 3.5, Kimi K2, and Manus itself are each, in different domains, at or near the global frontier. When you are at the frontier you become a strategic asset, and Beijing's disposition toward strategic assets is retention, not monetization.
• US acquirers are no longer viewed as neutral capital. Meta, Google, Microsoft, Amazon and Apple are now, in Chinese official discourse, instruments of US national-security and industrial-policy interests. Their bids are read through that lens, not through a commercial-transaction lens.
• The playbook for blocking is cheap. Exit bans plus a Ministry-level political framing cost Beijing almost nothing and impose enormous costs on any cross-border deal. There is no reason to expect the playbook to be used less, given how effective it has just been.

The harder question is whether European acquirers, sovereign-wealth buyers, or Asian regional buyers can do what US tech majors now cannot. The early read is: partially. A European or Middle Eastern buyer taking a minority stake with clearly defined data-residency and IP-fencing provisions is a structurally different transaction than a US tech major buying the whole company. That kind of deal might be approvable by Beijing. A full US acquisition of a Chinese frontier AI company, on the other hand, looks structurally impossible for the foreseeable future.

Our verdict

Meta's acquisition of Manus is, as of April 2026, functionally dead even though no formal block has been issued. That ambiguity is the point — it lets Beijing protect a strategic AI asset without triggering the diplomatic and economic costs of an explicit ban, and it lets Meta avoid publicly writing off a signed transaction. Both sides benefit from the fuzziness. Only Manus users and Butterfly Effect staff bear the costs, through slower product iteration, degraded reliability, and indefinite uncertainty about the company's future.

For anyone evaluating Manus as a production tool in 2026, the posture is straightforward: the product works, the benchmark numbers were real, and the engineering behind them is legitimately impressive — but building a business dependency on a Chinese-origin AI agent whose founders cannot leave the country and whose strategic buyer has just been branded a conspirator is a risk profile most enterprises will not accept. For everyone else watching the US-China AI race, this is the template. Expect to see it again.