Beijing Forces Meta to Unwind $2 Billion Manus Acquisition as Engineers Lose System Access

Meta has finished walling off its operations from Manus, the Chinese-founded agentic AI startup it acquired for about $2 billion in December, according to a Bloomberg report citing people familiar with the matter. Manus employees have been locked out of Meta's internal data systems since the start of June, Meta staff are now barred from using Manus tools for internal work, and an internal memo says Meta is "sunsetting" the platform. Existing Manus projects will be migrated onto Meta's own infrastructure.

The separation is the first concrete step toward complying with an order from China's National Development and Reform Commission (NDRC) to reverse a completed deal. The timing is awkward: the unwinding is happening just as the startup's three founders try to raise roughly $1 billion to buy their company back at a valuation matching the $2 billion Meta paid.

A regulatory first, and why it matters

The NDRC ordered the acquisition undone in April under its foreign investment security review mechanism, China's rough analog to the Committee on Foreign Investment in the United States (CFIUS). What makes this notable is not the existence of the review but the outcome. This is the first time Beijing has forcibly reversed a completed cross-border AI acquisition rather than blocking one before close.

The commission asserted jurisdiction even though Manus had relocated its headquarters and core engineering team from Beijing to Singapore in mid-2025, a move that on paper should have weakened China's claim. It did not. The review escalated in March, when authorities barred co-founders Xiao Hong and Ji Yichao from leaving the mainland, and the order reportedly required Manus's Chinese assets to be restored to their pre-acquisition state within weeks.

Manus had been compared to DeepSeek in Chinese state media as a symbol of homegrown AI capability. That framing turned its sale to a U.S. hyperscaler into a test case, and Beijing decided it could not let the transfer stand.

The problem with reversing a software deal

This is where the case diverges from the playbook regulators are used to. When a fab sale or a factory acquisition gets blocked, the remedy is mechanical. You return the equity, you return the equipment, you return the intellectual property, and the parties walk away roughly where they started. Physical and contractual assets can be handed back.

Manus's value does not sit in hardware. It sits in model weights and engineering know-how, both of which have been flowing into Meta for six months. No firewall, no system lockout, and no internal memo can recall what Meta's engineers have already absorbed from working alongside that team and that codebase. Meta has not yet explained how it intends to demonstrate to the NDRC that Manus's technology is genuinely out of its stack, and there may be no clean way to prove a negative of that kind.

That asymmetry is the real story for anyone tracking AI supply chains. The unit of value in frontier AI is increasingly intangible. Once weights and training methodology cross an organizational boundary, the transfer is effectively irreversible regardless of what the paperwork later says.

Part of a broader tightening

The Manus order extends to AI companies and their engineers the same posture Beijing has been applying to silicon throughout the year. Chinese regulators have held up Nvidia's H200 shipments even after Washington cleared them for export, signaling that import approvals from one capital mean little if the other capital objects. On the domestic side, DeepSeek launched its 1.6 trillion parameter V4 model running on Huawei silicon, reinforcing the message that China intends to build its top-tier AI on homegrown accelerators rather than imported ones.

Viewed together, these moves describe a coordinated effort to keep both the chips and the model-building talent inside the national perimeter. The H200 holds restrict what flows in, the V4-on-Huawei launch demonstrates self-sufficiency, and the Manus reversal blocks what flows out. AI capability is being treated as a strategic asset subject to the same export-control logic that has governed advanced lithography and leading-edge process nodes.

The buyback question

Founders Xiao Hong, Ji Yichao, and Zhang Tao have discussed raising about $1 billion from outside investors to fund a buyback, with the goal of repurchasing the company at a valuation at least matching what Meta paid. How far those talks have advanced is unclear. Early backers, including Tencent, ZhenFund, and HSG, have already collected their proceeds from the original sale, which means a buyback would have to assemble fresh capital rather than simply reverse the cap table.

The founders therefore face a layered problem. They need to raise roughly half the original purchase price from new investors, restore the company's Chinese assets to satisfy the NDRC, and rebuild a standalone operation after six months of integration into Meta. Meanwhile Meta has to satisfy a regulator that its engineers can somehow un-learn what they have learned. Neither side has an obvious path to a clean resolution, and the outcome will set a precedent for how Beijing treats every future foreign bid for a Chinese AI lab.

For the wider market, the lesson is that cross-border AI deals now carry a tail risk that did not exist a year ago. A completed acquisition can be ordered undone after close, and the most valuable part of the asset cannot be returned. That changes the math on valuation, diligence, and integration speed for any hyperscaler eyeing Chinese-founded AI talent.