The Case for Sovereignty

For mission-critical businesses, sovereign AI no longer needs a business case. The new world order already made it. What remains is defining it, and the honest definition is not independence. It is managed interdependence.

The business case for sovereign AI is already settled. For any organisation whose operations cannot be interrupted, it was not settled by technologists but by policy makers, on a specific timeline, in full public view. So the useful question is no longer whether. It is what, and how you build for it.

Sovereignty is not independence. The clearest statement of this comes from Srikanth Velamakanni, co-founder of Fractal and now chair of India's NASSCOM, who argues that for any country outside the two superpowers, self-sufficiency is the wrong target. The right one is managed interdependence: controlling what you cannot afford to lose, and depending, deliberately and hedged, on the rest. His test is the one to carry through everything that follows.

The test is not whether you own the layer. It is whether someone else can switch it off, and whether that would end you.

After Srikanth Velamakanni · Fractal / NASSCOM · Bloomberg, Jul 2026

Hold that test. The rest of this piece is the case for why it is now the only sensible way to think about sovereign AI, and what it means for a regulated business.

The case is closed because the order it depended on came apart. For thirty years a few assumptions held: trade was rules-based, interdependence was a stabiliser, the United States underwrote security, and supply chains were sourced wherever was cheapest. Sovereignty was something you pooled or traded away for prosperity. The bet was that dependence was safe.

Between early 2025 and early 2026 that bet was called, and failed:

• Jan 2025The US Diffusion Rule tiered global chip access, placing neutral, allied Switzerland in the capped second tier. It was rescinded before it bit, but access had been shown to be revocable at another government's discretion.
• Apr 2025"Liberation Day" tariffs hit a near-century high; China answered with its first rare-earth export controls, affecting all countries.
• Oct 2025China imposed a rare-earth regime claiming, for the first time, extraterritorial authority over products made anywhere using Chinese material. The ECB estimates over 80% of large European firms sit within three intermediaries of a Chinese supplier.
• Nov 2025The US National Security Strategy: "the days of the United States propping up the entire world order… are over." A one-year mutual truce followed, but the machinery was left intact.

That is the shift: not one event, but the discovery that dependence, the quiet assumption underneath globalisation, is a weapon that can be pointed at you by governments whose disputes are not even your own. Access became conditional. Efficiency stopped being the only thing worth optimising for.

So the case makes itself. Everyone now invokes the term. Almost no one defines it, and a firm that cannot define sovereignty precisely will spend real money pursuing the wrong version of it. To define it, decompose it, one layer at a time.

First, the layers

The most useful decomposition comes from the man who sells most of the picks and shovels. Jensen Huang describes AI not as software but as infrastructure, a stack of physical layers, each resting on the one below. The chatbot everyone sees is only the icing. Beneath it sits a cake: energy (the floor), chips (the scarcest layer), infrastructure (data centres), models (the intelligence), and applications (where value is created). Cutting across all five sits the layer no European regulator would leave implicit: data.

Sovereignty means something different at each layer. Apply the test from the top: not do we own it? but could someone else switch this off, and would that be catastrophic?

Sovereignty, then, is not one property you have or lack. It is a portfolio: owned at some layers, controlled at others, absent at one. Hold that thought, because it is the whole answer in miniature. The goal is never to own the cake. It is to manage the portfolio.

No one owns the whole cake

Here is the fact that reframes the debate. Not a single country owns every layer, not even the United States. The US designs the world's best chips and controls the models, infrastructure, capital and energy around them. But it does not make those chips at home. It depends on a single Taiwanese foundry, TSMC, and on Taiwan itself. Cut that link and the most powerful AI nation on earth stalls at the layer it cannot substitute.

China proves the same point from the other side. Denied the frontier chip by export controls, it did not pretend the dependency away. It built around it. Sovereign at the energy, data and application layers by scale and law; sovereign-by-construction on infrastructure; and racing to close the one gap that matters, fabricating its own silicon through Ascend hardware and the CANN software stack. It cannot yet match the frontier chip. But it has built a parallel stack good enough that it can never be switched off from outside. That, not parity, is the point of the exercise.

Two countries, the US and China, hold roughly 90% of the world's leading AI compute, and between them produce nearly every top-ranked foundation model. Only one owns the full stack. Only one is deliberately building a parallel to it. Everyone else owns some layers and depends on others. That leaves three postures, and only three. The third is the one that applies to almost every reader of this piece.

No country owns every layer, so there are only three postures. The third applies to almost every organisation:

Back to the test

Now the definition from the opening pays off. That third posture is Velamakanni's managed interdependence, and the layer walk is what makes his test operational: for each layer, could someone else switch it off, and would that end you? Sovereignty is the portfolio of answers, not a single yes or no.

The expensive mistake, made in more than one European capital, is to aspire to own the cake, fail, and call that failure. It is not. Full-stack autarky is what Brookings calls "structurally infeasible for almost any country." The third posture, done well, is sovereignty. The goal was only ever to ensure nothing you cannot afford to lose sits in someone else's hands.

Switzerland, and your board

Switzerland is what managed interdependence looks like when it is done on purpose: a fully open-weight national model, sovereign compute under Swiss law, a sovereign clinical model in a hospital, world-leading talent and domestic energy. It secured the layers it could not afford to lose, and depends on the chip openly, hedged through alliances, rather than pretending otherwise. It did not try to own the cake. It managed the portfolio.

And it exposes the trap that catches every newcomer: hosting is not sovereignty. A data centre in Zurich feels sovereign; if the operator is US-incorporated, it stays reachable under US law wherever the servers sit. Residency asks where is my data? Sovereignty asks whose law can compel it? For roughly 90% of Swiss firms the binding constraint is not the chip at all. It is the CLOUD Act: jurisdiction, not silicon.

Sovereign AI, done honestly, is not independence. It is knowing exactly which dependencies you can afford, and which would end you.

For a nation this is a strategic priority. For a bank, an insurer, a hospital, it is fast becoming a board-level one, and it will not be answered by a strategy deck. Managed interdependence is not a lower ambition than sovereignty. It is what sovereignty actually is, once you stop treating it as a slogan. And it is a delivery problem: knowing, layer by layer, which dependencies you can afford, and building so the rest are never in someone else's hands.

Which layer is your organisation's real dependency, and would you know if it were switched off? I would genuinely like to know.