Geopolitics & Sovereign Positioning

China's National Development and Reform Commission issued a one-sentence order on Monday prohibiting Meta's acquisition of Manus, an AI agent startup that originated in China and had been valued at $2.5 billion in the deal. The blocking order gave no specific justification, which is itself strategically significant: Beijing is asserting extraterritorial authority over the transfer of AI talent and technology without needing to articulate a legal basis, mirroring the opacity with which the US has applied export controls. According to South China Morning Post, the NDRC's action came roughly four months after the acquisition was announced, during which time Manus had already provided Meta employees with unlimited-usage account access — meaning technical transfer had partially occurred before the block was imposed.

The unwinding problem is now the operative challenge. Analysts cited by South China Morning Post describe the reversal as 'time-consuming, complex and difficult' given how far integration had progressed. Foreign Policy frames the reversal as reflecting Beijing's shifting national security calculus — a recognition that Chinese-origin AI agent technology represents a category of capability that should not be allowed to migrate into US frontier lab ecosystems. The deal's collapse also sends a chilling signal to other Chinese AI founders who have structured their companies with offshore holding entities: Beijing retains the authority and apparent willingness to intervene regardless of corporate structure when it judges the technology strategic.

The effective closure of the Strait of Hormuz since early March — a consequence of the Iran conflict — is producing measurable shortages of naphtha, a petrochemical feedstock critical to the production of photoresist, the light-sensitive material used in semiconductor lithography. South China Morning Post and Foreign Policy both document the supply chain cascade: Middle East petrochemical producers supply significant volumes of naphtha to Asian specialty chemical firms, which in turn supply chipmakers across Taiwan, South Korea, and Japan. The Foreign Policy piece extends the analysis to helium, another Hormuz-dependent input used in chip fabrication and AI data centre cooling systems.

This is a qualitatively different category of semiconductor supply risk from the one addressed by US-China export controls. Export controls target specific equipment and components at the advanced-node level; the Hormuz disruption is hitting commodity-level chemical inputs that underpin the entire semiconductor production stack, including legacy nodes. Nations that have invested heavily in domestic fab capacity — including the US CHIPS Act beneficiaries, Taiwan, and South Korea — face upstream exposure that no amount of onshoring of fab facilities resolves if the chemical supply chain remains concentrated and geographically vulnerable. For China, which is already constrained at the advanced-node equipment level, a photoresist shortage compounds pressure at precisely the moment it is trying to scale domestic chip production.

Two concurrent developments this week confirm that the governance of access to frontier AI models is becoming a national security instrument, not merely a commercial or ethical decision. First, CSET at Georgetown documents the emerging norm among leading US AI labs — including Anthropic and OpenAI — of restricting or fully withholding access to their most capable models on dual-use grounds, specifically citing cybersecurity and biological research risks. GPT-Rosalind and Claude Mythos are named as examples of models being held back from general release. This represents a structural shift: the frontier of AI capability is no longer a public or commercial good but a controlled asset, with access mediated by the lab's own assessment of risk — not by government mandate, but in anticipation of it.

Second, South China Morning Post reports that China's cybersecurity establishment is treating Claude Mythos as an external threat vector — an AI system capable of powering sophisticated cyberattacks that could target Chinese infrastructure. Chinese experts cited in the piece note that Beijing's strict regulatory scrutiny of its domestic AI industry may provide some buffer, but the concern is real. The juxtaposition is strategically revealing: US labs are restricting model access partly to prevent adversaries from exploiting dual-use capabilities, while China is simultaneously treating those same models as inbound threats. Both responses reflect the same underlying reality — frontier AI models are now understood as offensive instruments of national power by both sides.

The Diplomat reports that Kazakhstan and Uzbekistan are committing billions of dollars to data centres, AI infrastructure, nuclear power, and mining simultaneously — while the region's water supply is becoming less abundant, less predictable, and more legally constrained under transboundary water agreements. Data centres require substantial water for cooling; AI training and inference workloads intensify that demand. The Aral Sea basin's ongoing depletion and the legal frameworks governing Amu Darya and Syr Darya water allocation impose hard physical and political constraints on the scale of digital infrastructure these states can realistically host.

The strategic context matters: both Kazakhstan and Uzbekistan have been positioning themselves as swing states in the US-China AI competition, attracting investment from Chinese cloud providers, US hyperscalers, and Gulf sovereign wealth funds seeking alternative data centre geography outside the main US-China rivalry. If water constraints bite, these states' ability to serve as credible neutral AI infrastructure hubs is materially limited — which in turn affects the viability of the 'third pole' digital infrastructure strategy that several Global South states are attempting.