China's National Development and Reform Commission (NDRC) has formally blocked Meta's acquisition of the AI agent startup Manus, ordering the parties to unwind the deal. The decision, made through the foreign investment security review office, is more than just a failed merger — it's a clear signal that AI companies with deep model capabilities, agent frameworks, data loops, and engineering talent are no longer treated as ordinary commercial assets.
Over the past two years, many AI startups have told a globalization story: teams in China, products for overseas markets, funding in dollars, parent entities offshore, and an eventual exit to a big tech acquirer or a foreign stock listing. It seemed logical. But after Manus, that path looks a lot less certain.
It’s not just Manus — it’s a whole playbook
What makes Manus different? It’s not a simple chatbot. The real sensitivity lies in what Manus represents: an AI agent that doesn’t just answer questions but understands objectives, breaks down tasks, calls tools, executes workflows, and keeps going across multiple rounds. On the surface it’s an app, but underneath it’s closer to an operating system that makes models act.
A mature agent product isn’t a few pages or a set of prompts. It’s a whole engineering stack: task planning, tool calling, browser control, code execution, file handling, data reading, result verification, and the system that holds it all together. In the old internet era, an acquisition transferred users, brands, and teams. In the AI agent era, it can transfer task execution systems, model interaction paradigms, data loops, and engineering know-how. Those aren’t easy to classify as ordinary business assets anymore.
In other words, the Manus block isn’t just about one company being bought by foreign capital. It’s about a deal that may have crossed a line on the cross-border transfer of core AI capabilities.
Regulators look at control, not corporate shells
This episode also makes one thing clear: an offshore corporate structure won’t necessarily dodge substantive review. Many startups have been comfortable wrapping themselves in a global narrative: registered abroad, products abroad, funding from dollar funds, and teams split between China and elsewhere. That worked in the internet age. It helped with fundraising, options, overseas growth, and future exits.
But AI is different. The real question isn’t where the company is registered, but where the core technology is built, where the key team works, where the data comes from, and who controls the models and systems. If a company’s core R&D, product design, engineering systems, and talent are primarily Chinese, regulators can look through the structure and decide if it’s a foreign acquisition of key Chinese technology.
For AI startups, this is a direct reminder: offshore entities are not a magic shield. In fact, Chinese regulators have begun telling some AI startups to reject US capital outright unless they get government approval first.
In the past, founders set up offshore structures mainly for fundraising and exit efficiency. Now they have to answer another question: can this structure survive a security review? That will force some companies to redraw their boundaries. What technology stays inside China? What IP goes offshore? Can the offshore entity operate independently? What’s the relationship between the domestic team and the offshore parent? Does user data cross borders? Does model capability transfer? These used to feel like legal and finance homework. They’re becoming strategic questions that founders have to own.