Over the past year, two rival architectures for artificial intelligence have begun to take shape.

In December 2025 the United States launched Pax Silica, a State Department-led coalition meant to secure the supply chains — critical minerals, semiconductors, AI infrastructure — that the technology rests on.

Seven months later, in July 2026, China gathered 29 countries in Shanghai to establish the World Artificial Intelligence Cooperation Organization (WAICO), an intergovernmental body pitched at the Global South.

Neither is yet a settled bloc; both are still in the making. Even so, this month Washington reportedly warned the signatories of its AI Opportunity Statement that they cannot belong to both — “to be part of everything is to be part of nothing.”

The instinctive reaction is to ask whether AI is about to break in two. The more useful question is where a break could actually hold, because the answer is not uniform across the technology.

To address it, a way forward is to compare AI with previous technologies when it comes to bifurcation/fragmentation versus globalization.

The internet and the telephone system fragmented at the surface, especially in China with the Great Firewall but also more generally through data-localization rules and national controls — but never at the base, because a network is worth more the more people it reaches, and no rival could afford a separate substrate.

It is not clear that AI obeys the same rule. Whether AI splits, and how deeply, depends on which layers can be controlled by a single actor, and we can already see that some key parts of the stack are bifurcating, i.e., separating between the U.S. and China.

The hardware can be split

The hardware layer is one of the core points of the stack and, as such, the U.S. — which has long dominated this layer — has tried to ring-fence it through export controls. Advanced chips, the lithography that makes them and the high-bandwidth memory have developed into essential chokepoints in the U.S. quest for AI dominance.

One of the key aspects of Pax Silica is the coordination of such chokepoints — among those members that hold such power — but also access to chips outside export controls and to shared compute infrastructure.

The software cannot

The model layer is another matter, and this is where the ambition to bifurcate meets reality.

Open-weight models — DeepSeek, Qwen, Kimi — are already downloaded and run worldwide, and once weights are released they cannot be recalled. Restricting how models are used would mean policing every developer and enterprise that runs one; sanctioning distillation, whereby a cheaper model is trained on a stronger one’s outputs, would mean detecting a process that leaves almost no trace.

Even cutting off access to proprietary frontier systems — as one leading U.S. developer briefly did this year for users outside the United States — does not seal the border, because capability leaks around it through open weights and distillation. Chips can be counted and stopped. Model capabilities diffuse, and no export regime yet devised can hold them.

The realistic outcome, then, is not a sealed border but an asymmetric split: a genuine fracture at the level of hardware and infrastructure beneath a software layer that stays porous.

That is the worst of both configurations. It fragments the market — two chip ecosystems, two sets of standards, duplicated infrastructure — without delivering the security either side is chasing, because the capabilities each is trying to contain slip across the divide regardless.

And it is being pushed from both directions: Washington by turning a supply-chain coalition into a loyalty test, Beijing by building WAICO as a permanent body speaking for the Global South, with Xi Jinping's call for an order in which “no country should dominate” serving as its own bid to lead.

The lesson of Singapore

Nowhere is the cost of this clearer than in Singapore, and Europe should watch it closely.

A founding member of Pax Silica, Singapore was assured in May by the coalition’s architect, namely the Trump administration, that joining implied no obligation to choose between China and the U.S. By August, the Singaporean government found itself among the recipients of a letter instructing it to choose after all.

The reality is that Singapore cannot choose. A ban on Chinese models would hollow out its own economy — and, given how porous the software layer is, would not even work. Singapore is a trusted partner discovering both that the terms changed after it signed and that the choice being demanded is one the technology makes unenforceable anyway.

Europe is in the same position, with more at stake. Its recent sovereignty legislation — the Cloud and AI Development Act and the wider June 2026 technology package — was conceived before the U.S. and China hardened their respective positions.

In fact, European legislation was built precisely to avoid choosing: a strategic-autonomy bet that assumes room to hedge. Yet in the same month, the EU signed the Pax Silica Declaration, planting itself in the U.S. camp. Its laws seek a third option; its signature already took the U.S. route.

The Singapore lesson is that autonomy and bloc membership coexist only until someone forces the question. And yet, Europe is different from Singapore because it is fully indispensable for at least one aspect of the AI stack, namely advanced lithography, without which neither side can build advanced chips.

That leverage lasts only as long as Europe is prepared to use it as an EU-level tool and not merely a Dutch one, the Netherlands being the home of ASML’s lithography machines. For the moment, Europe is relinquishing its leverage to the U.S. virtually for free.

The other risk is that China manages to crack the code of advanced lithography and the current chokepoint disappears. That would only deepen the bifurcation of AI hardware.