There are decisions that reveal power, and others that betray its limits. Beijing’s order to unwind Meta’s acquisition of Manus belongs, rather uncomfortably, to both categories at once. On paper, the gesture is absolute. The Chinese state planner has demanded that Meta Platforms abandon its $2 billion purchase of the AI start-up, restore its Chinese assets to their original state, and comply within weeks—under the implicit threat of sanctions. The signal is unmistakable: artificial intelligence is no longer merely an economic sector; it is sovereign territory. Capital, talent, code—once they bear the imprint of China—remain, in the eyes of the Party, indefinitely reclaimable.

And yet, reality is less obedient than decree. The transaction, concluded four months ago, has already dissolved into the bloodstream of global AI. Engineers have relocated. Executives have been absorbed into Meta’s upper echelons. More critically, the code—the only asset that truly matters—has already been transferred, integrated, perhaps irreversibly diffused across Meta’s systems. One does not “undo” the circulation of intelligence any more than one rewinds a thought once expressed. The attempt, therefore, is less a reversal than a performance of sovereignty.

This is where the illusion begins. For years, China has exercised formidable control over its domestic champions—humbling Alibaba Group Holding Ltd., suspending Ant Group Co.’s IPO at the edge of triumph, forcing Didi Global Inc. to retreat from New York. These were acts of internal discipline, the reassertion of hierarchy within a closed system. Manus, however, is different. It is a creature of globalisation: Chinese in origin, Singaporean in structure, American in destination. It sits precisely at the fault line where sovereignty dissolves into networks. And it is at this fault line that Beijing now draws a line it may no longer fully control. The deeper significance of this episode does not lie in the fate of Manus itself—likely already decided—but in what it reveals about the structural transformation of AI. The true battleground is no longer the ownership of start-ups, but the capture of compute, the monopolisation of infrastructure, and the foreclosure of access.

A striking anomaly defines the current AI economy: the very firms absorbing the largest capital expenditure cycle in technological history are operating their infrastructure at extraordinarily low utilisation—GPU usage as low as 5% in some clusters—while simultaneously hoarding supply and extending asset lifespans through accounting manoeuvres. This is not inefficiency. It is a strategy. In such a system, control no longer requires ownership of the firm. It requires control of the bottleneck. Beijing understands this—perhaps better than most. Its intervention in the Manus deal is less about reclaiming a start-up than about preventing leakage from a system it is struggling to seal. The fear is not merely that technology is transferred, but that it escapes into a global architecture already dominated by a handful of players—Meta Platforms Inc., Alphabet Inc., OpenAI—whose power no longer derives from geography but from compute concentration. AI has entered a phase of “structural foreclosure”, where a small number of firms control both the hardware layer and the frontier capabilities, effectively locking out competitors and redirecting value across entire industries. In this context, Manus was never just an acquisition—it was an attempt by Meta to accelerate its position within this closed circle.

Beijing’s reaction, therefore, is not surprising. It is late. Because the paradox is now evident. China can prevent future deals. It can intimidate founders, restrict capital flows, and monitor talent. It can even, in extreme cases, coerce individuals. But it cannot retroactively contain what has already been distributed through a system whose logic is inherently global and whose core asset—intelligence—moves at the speed of integration, not regulation. The result is a form of strategic dissonance. On one side, a state seeking to reassert territorial sovereignty over intangible assets. On the other hand, an industry whose value is increasingly detached from territory and anchored instead in infrastructure concentration, capital cycles, and network effects. The Manus affair sits precisely at this intersection—and exposes its instability.

For investors, the lesson is not legal but structural. The real risk is no longer regulatory blockage of individual transactions, but the progressive fragmentation of the AI ecosystem into competing spheres of control—each attempting to retain talent, capital, and compute within its own perimeter. Chinese AI start-ups are becoming de facto, “trapped within the domestic ecosystem”, as cross-border financing channels narrow and exit options disappear. In other words, the global AI market is not integrating. It is hardening. The Manus decision is not an isolated event. It is an early signal of a broader reconfiguration—one in which sovereignty reasserts itself not by controlling companies, but by constraining the flows that sustain them. Yet, paradoxically, it arrives at a moment when those flows have already exceeded the reach of any single jurisdiction. Hence the quiet irony. Beijing has demonstrated that it can still block the door. It is far less clear that it can retrieve what has already walked through it.