The $2 billion Meta acquisition of Manus has been halted behind the scenes

April 27, the National Development and Reform Commission issued a statement on its official website: making a decision to prohibit foreign investment in the Manus project in accordance with laws and regulations, requiring the parties involved to revoke the acquisition transaction.

This is the first time since 2021 that a regulatory agency has publicly announced a prohibition on investment review conclusions in the AI field.

The halted deal was Meta’s acquisition of AI startup Manus for about $2 billion, one of Meta’s largest mergers and acquisitions in history.

What did Meta pay $2 billion for

In the past two years, the competition focus in the AI industry has shifted from whose model is smarter to who can make AI truly work. The former is an upgraded version of information retrieval, while the latter is an alternative to workflow automation.

OpenAI, Google, and Anthropic are rapidly advancing their agent deployment, while Meta is noticeably lagging behind in this area.

Manus fills exactly this gap.

In March 2025, Manus released the world’s first general AI agent, with a demo video that sparked excitement both domestically and internationally: given a high-level instruction, it can autonomously complete market research, coding, data analysis, and website building.

After launch, the invitation code was once sold for several ten thousand yuan, with a waiting list exceeding 2 million people. In just eight months, the annualized revenue surpassed $125 million, and the company claims to have set the fastest record from zero to this milestone globally.

This acquisition took about ten days of negotiation. For Meta, it’s essentially time-for-money.

Developing a mature agent system internally would require many years and numerous failed iterations; Manus has already consumed these costs in advance.

What was bought was not just a product, but a market-validated agent capability framework: tool invocation system, multi-step reasoning chains, accumulated real user data, and a complete core team.

Xiao Hong was originally set to become Meta’s Vice President, reporting directly to the COO. This organizational arrangement itself indicates that Meta’s goal is not just an application, but the people who can create this application.

How the deal was blocked

On December 29, 2025, Meta announced the acquisition officially. Regulatory intervention followed.

In January 2026, the Ministry of Commerce publicly stated that it would evaluate whether the transaction complies with laws and regulations related to export controls, technology import/export, and foreign investment.

By March, the National Development and Reform Commission convened a meeting with Meta and Manus executives, and subsequently, co-founder Xiao Hong and chief scientist Ji Yichao, among others, remained in China during the investigation.

On April 27, the review result was finalized, less than four months from intervention to final decision.

Manus’s operational model is somewhat unique in the industry.

In 2022, the founding team established the parent company Butterfly Effect in China, completing technology R&D and early commercialization;

In June 2025, the operating entity was relocated to Singapore, with only about 40 core R&D staff remaining in Beijing’s 120-person team.

By the end of 2025, there was almost no trace of the company left in China.

However, the core of the review by the NDRC focused on where the technology was developed, where the data was accumulated, and who built the engineering capabilities under what environment.

Manus’s agent framework, with a core R&D cycle from 2022 to early 2025, was entirely completed in China; the product exploded in March, but the company only moved out in June; within eight months of launch, it handled requests from users involving 147 trillion tokens, and the generated content cannot be covered by changes in the registered location.

Xiao Hong graduated from Huazhong University of Science and Technology, Ji Yichao from Beijing Information Science and Technology University. Both have over ten years of product and technical experience in China. The team’s cognition, judgment, and engineering intuition are products of China’s education and entrepreneurial environment.

Why the deal was not approved

Throughout Manus’s migration and sale process, it never reported export controls or technology export reviews to relevant authorities, which is the most direct issue.

But this review pointed to more than just Manus; it revealed the operational path behind it: the team completed core capability accumulation domestically, introduced external capital through structural adjustments, and ultimately sold to overseas tech companies.

The decision by the NDRC is a clear statement on this model.

The timing of Manus’s rise coincided with the moment when the AI track began to be truly recognized by the global market.

The emergence of DeepSeek prompted a reassessment of domestic AI teams’ engineering capabilities, and Manus followed closely, further validating the practical abilities of domestic teams in the agent direction.

The formation of this capability relies on long-term support from China’s data environment, engineering talent, and entrepreneurial ecosystem. Selling it as a whole to an overseas tech company involves more than just commercial M&A in terms of regulatory considerations.

As a core technology field, AI is increasingly subject to cross-border investment reviews, a trend accelerating worldwide.

On April 24, just days before the NDRC’s decision, some leading domestic AI companies also received compliance alerts, requiring approval procedures for accepting external capital.

These signals collectively point to issues beyond individual cases.

Unreturnable deals

Regulatory requirements to revert to the state before investment target the tangible elements like equity and assets, assuming withdrawal can be operationally executed.

But Manus’s situation is different: some employees have already integrated into Meta’s internal team, funds have been transferred, and reports indicate that technical integration has already begun.

Equity can be transferred, funds can be refunded, but the collaborative structures between the teams and the tacit understanding formed among engineers working together cannot simply be undone.

According to media reports, Meta is preparing a withdrawal plan, with deadlines set by regulators, but the specific implementation path has not been publicly disclosed. Manus has made no statements.

For participants in the AI industry, this is a clear signal. The cross-border transfer of core technological capabilities, regardless of the structural arrangement, must be completed with prior compliance assessment, not after the transaction.

Where the technology comes from determines where it can go. Cross-border M&A is shifting from a path of technology diffusion to a new consideration of technology flow.