Digital sovereignty in AI ceased to be theory on June 12, 2026. That night, three continents received the same signal when the U.S. government ordered the shutdown of Fable 5: any AI model, however powerful, can disappear from global access by decision of a single government. The response, though incipient, is already being orchestrated.
For countries outside the U.S.-China duopoly — which controls 90% of global computing — dependence is not a strategic choice: it is a material constraint. And as Fable 5 demonstrated, that access can be cut without warning.
The catalyst
The international AI community had been watching Washington’s moves closely since January 2025, when the Biden administration introduced the AI Diffusion Rule and, with it, the Export Control Classification Number ECCN 4E091 — a mechanism that brought frontier AI model weights under U.S. export regulation jurisdiction. The Trump administration rescinded the rule in May 2025, but retained the underlying controls on model weights.
What changed with Fable 5 was not the law, but the precedent: for the first time, the U.S. government directly ordered a private company to block access to its product for foreign users. It was not a generic restriction on chips or export licenses — it was the off switch of a frontier model activated from a national capital. The message to the rest of the world was unequivocal.
The computing duopoly
The Center for a New American Security’s Sovereign AI Index (CNAS, 2026) quantifies the dimension of the challenge: the United States and China control approximately 90% of global computing capacity for frontier AI development and possess the world’s top 50 ranked foundation models. No other country appears on the list. This structural concentration — not ideological, not political, but physical — is the engine driving digital sovereignty strategies around the planet.
For countries that do not belong to either sphere of influence, dependence is not a strategic choice: it is a material constraint. It does not matter how many talents a European university trains or how many startups sprout in Bangalore: without access to clusters of thousands of accelerators, frontier model development is unviable. And as Fable 5 demonstrated, that access can be cut without warning.
The sovereign constellation
The response is taking shape on at least five simultaneous fronts.
India launched the IndiaAI Mission in March 2024 with an allocation of $1.25 billion (₹10,372 crore) to build sovereign computing capacity, a national data repository (IndiaAIKosh), and own foundation models through the BharatGen project. The Indian approach combines public investment with a national data framework that seeks to avoid dependence on external sources for model training.
Japan nearly quadrupled its budget for AI and semiconductor support to approximately ¥1.23 trillion ($8 billion+) in the 2026 budget, including an additional ¥631.5 billion ($3.97 billion) in subsidies for Rapidus, the startup aiming to produce 2nm chips by 2027. The Japanese strategy recognizes that sovereignty in AI begins at the semiconductor manufacturing layer.
The European Union deploys multiple initiatives: the EURO-3C project (€75 million) for federated telco-edge-cloud computing, the bet on Mistral AI as the reference European foundation model, and the Gaia-X sovereign cloud initiative. The bloc is also advancing the EU AI and Cloud Development Act, which seeks to create a regulatory and investment framework that allows European actors to compete without depending on American or Chinese infrastructure.
Saudi Arabia has signed AI investment agreements worth $9.1 billion in 2025 and announced partnerships of up to $600 billion with NVIDIA, AMD, and AWS. Its national strategy, channeled through SDAIA and the HUMAIN program, aspires to convert the kingdom into an alternative AI computing pole, leveraged on its energy advantage.
South Korea selected five teams in 2025 for its Sovereign AI Foundation Model project (AI Foundation Model180), a coordinated effort that seeks to develop frontier model capabilities with public funding.
The sovereign AI infrastructure market was valued at $15 billion in 2025 and is projected to reach $118.5 billion by 2034, according to MarketIntelo — a compound annual growth rate of approximately 26% that reflects the urgency with which governments are addressing this dependence.
Incoherence as policy
The U.S. export control strategy has been, in the words of Chris McGuire, Senior Fellow at the Council on Foreign Relations and former senior deputy director of the National Security Council, “completely incoherent and self-sabotaging.” In an analysis published in January 2026, McGuire argued that chip policies sent advanced hardware to China while blocking American companies from launching models. The paradox intensified when the Trump administration rescinded the Diffusion Rule — which restricted chip access — but then ordered the Fable 5 block — a restriction on models — reversing policy direction without resolving the underlying contradiction.
The consequence of this incoherence is not only strategic: it is geopolitical. Every regulatory turn in Washington reinforces foreign governments’ determination to build their own capacity. If the rules change constantly, the only reliable safeguard is independence.
Is sovereignty possible?
Not all analysts are optimistic. A Brookings study (February 2026) poses an uncomfortable question: “Is sovereignty in AI possible?” The conclusion is nuanced: most nations face deep structural dependencies — from chip manufacturing (TSMC, Samsung) to software ecosystems (CUDA, PyTorch) and the energy supply chain — that make total independence illusory.
The sovereign infrastructure that India, Japan, and European countries are building still depends largely on American and Chinese technology. A server bought by the Indian government still carries NVIDIA chips. A European data center still runs American software. Sovereignty, in this context, is not binary: it is a spectrum ranging from operational autonomy to total independence, and most countries are at the more modest end.
What is at stake, then, is not completely replacing the U.S. and China, but building reserve capacity — the certainty that, if access is cut, there is an alternative. Even if it is slower, more expensive, and less capable.
A two-speed world
The geopolitical fragmentation of AI is, in fact, creating a tiered access system. Biden’s Diffusion Rule formalized three levels — trusted allies (Tier 1), neutral nations with quotas (Tier 2), blocked adversaries (Tier 3) — and although the rule was rescinded, the logic of stratification persists. The Fable 5 shutdown added a new dimension: it is not just about what hardware each country can buy, but about what models each person can use according to their nationality.
The result is a world map of AI that increasingly resembles the post-Cold War order, but with computing as the currency. Countries that are neither first-tier allies of Washington nor of Beijing — which is the majority — face a strategic decision: align with a bloc, build their own capacity, or both.
What seems unlikely is that anyone will sit still.
Primary sources: CNAS Sovereign AI Index (2026) | Brookings — Is AI Sovereignty Possible? (Feb 2026) | Chatham House — How Middle Powers Can Weather US and Chinese AI Dominance (Feb 2026) | CFR — The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable (Jan 2026) | MarketIntelo — Sovereign AI Infrastructure Market Report 2034 (May 2026) | Reuters — Anthropic Disables Top-Tier AI Models After US Order (Jun 2026)