Technological Sovereignty

Priya Sharma has a number she likes to quote, and it is not a large one. Seventeen billion.

That is the parameter count of BharatGen Param2, the multilingual foundation model her team at India's IndiaAI Mission unveiled at the India AI Impact Summit in February 2026. Seventeen billion parameters is small. The frontier models built by American and Chinese labs run to hundreds of billions, perhaps trillions. Ask BharatGen and GPT-4 the same hard question and GPT-4 will usually give the better answer.

Sharma knows this. She will tell you so herself. And then she will tell you why it does not matter as much as you think.

BharatGen was built in India, by Indian engineers, on Indian data, across twenty-two Indian languages. It does not route a citizen's query through a server farm in Oregon. It does not depend on a licensing agreement that a foreign company can revise, or a foreign government can revoke. If relations between New Delhi and Washington sour tomorrow, BharatGen still runs. It is not the most powerful model in the world. It is India's.

That distinction — between the most powerful and one's own — is the entire subject of this chapter. Around the world, governments have concluded that consuming artificial intelligence built by someone else is not the same as possessing it, and that the difference has consequences for their economies, their security, and their cultures. They have started spending accordingly. The trouble is that the thing they are trying to buy may not be for sale at any price.

Everyone Wants It. Almost No One Can Have It.

Start with the money, because the money is what makes this real rather than rhetorical.

By 2026, the running estimate for global AI infrastructure investment through 2030 sits at roughly $1.3 trillion, much of it flying the flag of "sovereign AI" — the principle that a nation should own and control its own AI capabilities rather than rent them. McKinsey has sized the sovereign AI market specifically at around $600 billion by 2030 (McKinsey, The Sovereign AI Agenda). These are enormous numbers, and we will come back to how much weight they can bear. For now, take them as a signal: sovereignty has moved from a talking point to a line item.

And yet the thing being purchased is close to impossible to fully obtain. Artificial intelligence is one of the most globalized technologies ever built. The chips are designed in one country, fabricated in another, packaged in a third, and installed in a fourth. The research circulates in open papers read on every continent within hours of publication. The talent moves — an Indian engineer trained in Bangalore, hired in California, poached to Abu Dhabi. Even a model proudly labeled "domestic" is, on inspection, a lattice of foreign components, foreign ideas, and foreign standards.

True independence — building AI end to end from purely indigenous resources, needing no one — is available to perhaps two countries on Earth, the United States and China, and even they are not fully self-sufficient. Both remain hostage, at the deepest layer, to a supply chain that runs through Taiwan.

So sovereignty cannot sensibly mean independence. It has to mean something more modest and more useful: the ability to act deliberately inside an interdependent system rather than having your choices made for you. Who you depend on. Where you build. Which rules you follow, and which you help write. Understood this way, sovereignty is not a wall around a country. It is a country's freedom of movement within a shared space. The nations that succeed will be the ones that treat it as participation plus leverage, not separation. That is a harder and more honest goal than autarky, and it is the one this chapter takes seriously.

Five Layers, Five Different Kinds of Trouble

To see why full sovereignty is so elusive, it helps to take the AI stack apart. Building it domestically means gaining control — or at least meaningful autonomy — across five layers, and each one fails you in a different way.

The table below sketches the terrain; the paragraphs that follow walk through it.

Layer	What it takes	How hard to control domestically	The binding constraint
Compute infrastructure	Data centers, power, cooling	Moderate	Reliable, affordable electricity
Semiconductors	Advanced chip fabrication	Extremely hard	Years, capital, and Taiwan
Foundational models	Training large models	Achievable, at a capability cost	Compute and talent
Data	Local-language training corpora	Uneven	Population scale and openness
Governance	Rules for development and use	Achievable, but fragmenting	Market size and standards power

Compute is the foundation, and it is fundamentally a story about energy. Data centers are electricity refineries with servers attached. Deloitte projected that over $100 billion would be committed to sovereign AI compute in 2026, with global data center capacity heading toward 130 gigawatts by 2030 — and it also flagged that for every $1 billion spent on the centers themselves, roughly another $125 million is needed to build out the power networks to feed them (Deloitte, Technology Sovereignty, 2026). This is why energy geography becomes AI geography. France, running on nuclear baseload, can power vast data centers with low-carbon electricity it already controls. The UAE and Saudi Arabia are pouring petrodollars into centers backed by cheap fossil energy. India and much of the Global South start from unreliable grids and expensive power — a handicap that no policy document can legislate away.

Semiconductors are the layer where sovereignty goes to die, and they deserve the most attention because they are the hardest. Only a handful of firms — Nvidia, AMD, a few others — design cutting-edge AI accelerators, and nearly all of them are fabricated in a single place: Taiwan. No country outside Taiwan, South Korea, and partly the United States can manufacture the most advanced chips at scale. The concentration is not an accident of policy that a subsidy can reverse. A leading-edge fab costs upward of twenty billion dollars, takes years to build, and depends on an ecosystem — extreme-ultraviolet lithography machines from a single Dutch company, ultra-pure chemicals, specialist labor — that itself took decades to assemble. You cannot buy your way past that timeline with money alone; the equipment, the yield-learning, and the tacit expertise are the actual bottleneck. This is what makes the semiconductor layer so much more intractable than the others: everything above it can be attempted in months or a few years, but a competitive chip industry is the work of a decade or more. It is also why export controls and the possibility of conflict in the Taiwan Strait are treated as existential in every serious sovereignty plan — a single disruption at this layer severs every layer above it overnight.

Foundational models are where the picture brightens. Training a large model is expensive but genuinely within reach for a well-resourced state, as India's BharatGen, France's efforts, and dozens of national projects show. The models are smaller and less capable than the American and Chinese frontier — sovereignty bought at a performance discount. The unavoidable question at this layer is not can we but how much capability are we willing to trade for control, a question with no universally right answer.

Data is, in some respects, the most tractable layer — and the most double-edged. Good models need training data that reflects local languages and contexts, and here scale is destiny. India's billion-plus users generate enormous volumes of local-language text and speech across every domain. Iceland, Estonia, or a Swahili-speaking nation simply has less raw material to work with. But data sovereignty carries a trap worth naming plainly: keeping data inside national borders protects citizens from foreign surveillance while exposing them to domestic surveillance. "Keep the data home" can mean "shield it from the NSA" or "hand it to our own intelligence services," and often means both at once. Whether data localization protects people or merely reassigns who watches them depends entirely on the quality of the domestic institutions doing the watching — which is precisely the variable that data-sovereignty rhetoric tends to skip over. Sovereignty from whom, and for whom, are the questions that decide whether it is a protection or a rebranding.

Governance — who writes the rules — is achievable for almost anyone, but it fragments the market. The EU's AI Act, India's emerging framework, and national policies everywhere assert that AI must obey local law and reflect local values. Fair enough. But every additional regulatory regime is another compliance cost, and companies rationally optimize for the biggest markets: the US, the EU, China. For a small country, regulatory sovereignty can curdle into its opposite — adopting Brussels's rulebook wholesale rather than risk being locked out of the ecosystem entirely. You can have your own rules, or you can have the same rules as the market you need access to. Often you cannot have both.

The layers are also chained together, which is the deeper point. Strength in models means little without compute; compute means little without power; and all of it rests on chips you probably cannot make. Weakness at any layer leaks upward through the whole stack.

graph TD
  A[Semiconductors<br/>hardest to control] --> B[Compute infrastructure<br/>gated by energy]
  B --> C[Foundational models<br/>achievable, at a capability cost]
  C --> D[Data<br/>gated by scale and openness]
  D --> E[Governance<br/>achievable, but fragments the market]
  A -.a single disruption here.-> E

Five Countries, Five Playbooks

Because the constraints differ by resource base and geography, so do the strategies. The contrasts are more instructive than the commonalities.

India plays to scale and talent. The IndiaAI Mission, approved in 2024 with a budget of ₹10,371.92 crore — roughly $1.25 billion — targets indigenous foundation models, shared public datasets, compute, and governance (IndiaAI Mission). India's edge is its people: some of the world's best AI engineers are Indian, and a meaningful number are returning from Silicon Valley to build at home. Its weaknesses are energy reliability and chip access. So India runs a hybrid: partner selectively with foreign firms, keep strategic control of the core, and make sure the domestic system can stand alone if foreign access is ever cut. BharatGen is the proof of concept — not the best model, but a model that cannot be switched off from abroad.

France plays to institutional credibility, nuclear power, and regulatory heft, and it has made peace with hybrid architectures rather than chasing purity. Bleu — a joint venture of Orange and Capgemini running Google technology under strict legal safeguards, launched with €107 million — is the emblem: foreign technology operated inside domestic legal and physical control. France is also a founding member of Gaia-X, the European push for a federated, interoperable cloud, and has committed up to €50 billion in partnership with the UAE to expand data center capacity. The logic is explicit: no single European state can match the US or China at the frontier, so France pools. Collective sovereignty across borders beats national independence pursued alone.

Germany plays to industry. Rather than contest the whole stack, Europe's largest manufacturing economy has aimed its sovereignty effort at industrial AI — smart factories, autonomous production, precision robotics — where capability converts directly into competitive advantage and where foreign dependency is a genuine security risk. On 18 November 2025, France and Germany convened a Summit on European Digital Sovereignty, launching a joint task force due to report in 2026. Germany's bet is domain-specific sovereignty: lead decisively at the application layer even while depending on others for models and chips. It is a serious answer to the impossibility of stack-wide independence — own the part of the stack where you already have an edge.

Smaller states play niche. Full-spectrum sovereignty is beyond their means, so they buy strategic depth where they hold an advantage and trade it for influence. Singapore leads on AI governance and fintech. Israel concentrates on defense AI and cybersecurity. The UAE builds data-center infrastructure and AI-driven public administration. For these countries the goal was never independence; it is indispensability — becoming the partner that larger powers cannot easily route around in the domains that matter most to them.

The United States, oddly, is also chasing sovereignty — a telling sign of how far the concept has traveled. In January 2026 the White House released an AI Action Plan built on three pillars: accelerate innovation, build domestic infrastructure, and lead in international AI diplomacy and security. Its infrastructure incentives target large sovereign data centers — projects of at least $500 million in capital, 100 megawatts or more, oriented around national security (White House AI Action Plan, 2026). Why would the world's AI leader need sovereign infrastructure? Because American AI is corporate and global — its supply chains, talent, and markets all cross borders — and if national interest ever diverges from corporate interest, Washington wants domestic capacity it can command. The plan leans hard on export controls, and there the American dilemma is sharp. Restrict too little and rivals close the gap; restrict too much and rivals simply build their own — as China's DeepSeek showed by reaching competitive performance on a fraction of the expected compute. Even the hegemon is navigating the sovereignty trade-off, and it has not found a stable answer either.

Why Governments Are Paying to Be Less Efficient

For thirty years the organizing principle of the global economy was simple: source each input wherever it is cheapest and fastest. Sovereignty inverts that. It accepts inefficiency on purpose, as the price of resilience and control. That is a genuine reversal, and it is worth being precise about what risks governments think justify it.

Three, mainly. The first is supply-chain disruption: a pandemic, a blockade, a natural disaster, or a conflict in the Taiwan Strait that halts chip flows and freezes every dependent industry. The second is foreign surveillance: data and inference running on foreign infrastructure is data and inference a foreign government may, under its own laws, be able to reach. The third is geopolitical leverage: a supplier who controls a critical input can extract concessions, and a nation dependent for its AI on a single foreign patron has quietly outsourced part of its foreign policy. Weighed against those, the efficiency premium of doing some things domestically starts to look less like waste and more like insurance.

Whether it is good insurance is the harder question, and the honest answer is: it depends on who you would otherwise depend on. For a nation whose principal AI supplier is a stable, treaty-bound ally, deep integration may be the smarter bet — cheaper, faster, and more capable than a homegrown alternative that lags the frontier by years. For a nation whose supplier is a strategic rival, even an expensive, inferior domestic stack can be worth it. Sovereignty is not virtuous in itself. It is a hedge, and the value of a hedge depends on what you are hedging against. A country that spends billions duplicating capabilities it could safely rent from a trusted partner has not achieved security; it has bought an expensive symbol.

What Eighty Percent Buys You

Suppose a country pulls it off — not the frontier, but a sovereign stack running at, say, eighty percent of frontier capability, under full domestic control. What does that actually get them, and what does it cost them?

It gets a surprising amount. Eighty percent is more than enough to run most of what a state and its economy actually need: translation and public services in local languages, document processing, fraud detection, logistics, medical triage, tutoring, defense and intelligence applications that must never touch foreign infrastructure. It gets cultural fidelity — a model that understands the country's languages, references, and norms rather than approximating them through an American lens. It gets negotiating leverage: a country that can build its own AI, even a lesser one, bargains from a stronger position than one that cannot. And it gets continuity — the assurance that the lights stay on if geopolitics turns.

What eighty percent forecloses is the frontier itself: the cutting-edge scientific discovery, the most advanced reasoning, the capabilities that may confer decisive economic or military advantage and that live, by definition, at one hundred percent. For most national purposes that gap is tolerable. For the narrow set of applications where being second-best means losing, it is not. The strategic judgment every government must make is whether its critical needs live in the eighty percent it can own or the twenty percent it cannot — and that is a judgment about capability gaps that no one can make with full confidence.

If the Blocs Actually Form

Now widen the lens. Suppose the major economies all succeed, and the world fragments into sovereign AI blocs — an American sphere, a Chinese sphere, a European federation, perhaps an Indian pole, with smaller states orbiting one or another. Is that good or bad for AI itself?

Mostly bad, with an asterisk. AI progress has run on openness: shared papers, shared benchmarks, shared standards, models and datasets that cross borders freely. Fragmentation taxes all of that. Incompatible standards mean work gets duplicated across blocs instead of built upon. Talent circulates less freely. Interoperability erodes; a tool built in one bloc may not run, or may not be legal, in another. The overall pace of advance slows, because the field's greatest accelerant has always been that a breakthrough anywhere quickly becomes a foundation everywhere.

The asterisk is competition. Rival blocs racing each other can, for a time, spur investment and urgency in ways a comfortable monopoly would not. But the deepest losers are not the big blocs. They are everyone outside them — the small and developing states that, in an open ecosystem, could adopt the best tools from anywhere, and that in a fragmented one must pick a patron and inherit its standards, its restrictions, and its blind spots. Fragmentation is a tax paid most heavily by those who never had the resources to build a bloc of their own.

The Divergence, and What Is Owed Across It

Which brings us to the structural fault line running under this entire subject. Building sovereign AI at scale requires deep capital markets, reliable energy, a large talent pool, and the ability to sustain multi-billion-dollar, multi-year projects — the multi-gigawatt campuses rising across Europe are each measured in billions and years, and each demands continuous reinvestment as the technology moves. A handful of large economies can do this. Most countries cannot, and no amount of ambition closes that gap.

For everyone else, the path to meaningful autonomy does not run through independence. It runs through partnership, specialization, and integration into regional ecosystems — Gaia-X, South Asian cooperation frameworks, bilateral deals that trade infrastructure access for alignment on values and standards. This is autonomy of a real but qualified kind: influence over the terms of one's dependence rather than escape from it.

And it raises a genuine obligation for the capable nations, one the sovereignty conversation mostly ignores. There is a difference between helping a smaller country build genuine domestic capacity — training its engineers, sharing tools, financing infrastructure it will actually own — and drawing it into your sphere as a captive customer dressed up as a partner. The first expands the number of countries with real agency over their technological futures. The second is digital patronage with better branding. A sovereignty offer that leaves the recipient more dependent than before, on terms it cannot renegotiate, is not a gift; it is a longer leash. Which kind of help the AI powers choose to extend will do a great deal to determine whether the coming decade widens the intelligence divide or narrows it.

How Much of This Should We Believe?

A word of caution about the numbers, because this chapter has leaned on some large ones. The $600 billion sovereign AI market and the $1.3 trillion in planned infrastructure are forecasts, not facts. They rest on assumptions — that announced projects get built, that current investment intent holds, that geopolitical conditions stay roughly as they are — and every one of those assumptions is fragile. A single major conflict, a chip embargo, a financial contraction, or a change of government in a big spender could move these figures substantially in either direction. Announced spending is also notoriously softer than spent money; headline commitments routinely exceed what materializes. Treat these numbers as a measure of the scale of intent — real, serious, and directionally telling — rather than as a prediction you could bank. The direction is far more certain than the magnitude.

The deeper uncertainty is not about the money but about the strategy's core bet. Sovereign models built today at eighty percent of frontier capability are viable precisely because eighty percent is useful. But no one knows whether that gap stays constant, narrows, or widens as frontier AI advances. If it holds or shrinks, sovereignty at a modest capability discount is a durable strategy, and every country that built one made a sound investment. If it widens — if the frontier pulls away fast enough that eighty percent becomes forty — then today's sovereign models age into expensive museum pieces, and the countries that poured billions into them will have bought yesterday's autonomy at tomorrow's prices. Which of those futures arrives is genuinely unknown, and it is the single variable that most determines whether the great sovereignty push of the 2020s is remembered as foresight or as folly.

Why 2026

Deloitte flagged 2026 as the year the shift toward technology sovereignty would accelerate, and the timing is not arbitrary. Several pressures converged. Supply-chain shocks made abstract dependencies concrete. Geopolitical tension — over Taiwan, over export controls, over who gets access to the best chips — made the risks of dependence impossible to ignore. And the visible capability gap between those who build frontier AI and those who merely consume it made the stakes of staying a consumer clear to every government paying attention.

What tipped in 2026 was not the invention of sovereignty as an idea — that had been discussed for years — but its migration from aspiration to operating principle. Before, sovereignty was something ministers gave speeches about. Now it is a constraint baked into procurement decisions, infrastructure plans, and regulatory design. A country that treats AI as something to buy is, whether it admits it or not, outsourcing decisions about its own future. In 2026 enough governments recognized this at once that the question stopped being whether to build and became how fast, at what cost, and with whom.

There is a hard temporal fact underneath all of it, though, and it constrains everything above. The countries now breaking ground on domestic chip fabrication will not have competitive volume production for the better part of a decade. Fabs take years; competitive fabs take longer; catching a moving frontier is harder still. So for the next ten years or so, "sovereignty" for almost every nation on Earth means sovereignty with a permanent asterisk at the silicon layer — real control over models, data, governance, and increasingly compute, sitting atop chips they cannot yet make and must still import from a supply chain running through a handful of vulnerable places. That is not a reason to abandon the pursuit. It is a reason to be clear-eyed that, for the foreseeable future, every sovereign AI stack in the world rests on a foundation its owner does not control.

Summary

Sovereignty is a matter of degree, not independence. True AI self-sufficiency is available to at most two countries, and even they depend on Taiwan for chips. For everyone else, sovereignty means the power to choose one's dependencies deliberately rather than have them imposed — participation with leverage, not separation.
The stack has five layers, each failing differently. Compute is gated by energy; semiconductors are the near-immovable bottleneck; foundational models are achievable at a capability cost; data is gated by scale and cuts both ways on surveillance; governance is achievable but fragments the market. Weakness at any layer leaks upward, and the silicon layer is the one no amount of money quickly fixes.
National strategies track resources and geography. India leverages scale and talent (IndiaAI Mission, ~$1.25 billion; BharatGen). France pools sovereignty through hybrids, nuclear power, and Gaia-X. Germany owns the industrial application layer. Smaller states specialize for indispensability. Even the US pursues sovereign infrastructure via its 2026 AI Action Plan.
Governments accept inefficiency to hedge three risks — supply-chain disruption, foreign surveillance, and geopolitical leverage. Whether the hedge is worth it depends on who the alternative supplier is: deep integration with a trusted ally can beat an inferior homegrown stack, while dependence on a rival justifies real cost.
Partial sovereignty delivers most of the value. Eighty percent of frontier capability under full domestic control covers most public, economic, and security needs, plus cultural fidelity and negotiating leverage — foreclosing only the frontier itself. The open risk is whether that gap stays small or widens as the frontier advances.
Fragmentation into blocs would slow AI overall and fall hardest on states too small to build a bloc, sharpening the divide between AI builders and AI subjects. This creates a real obligation on capable nations to build genuine capacity in smaller states rather than dress dependence up as partnership.
The headline numbers signal intent, not destiny. The $600 billion market and $1.3 trillion in planned infrastructure are geopolitically sensitive forecasts of intent, not bankable predictions. And for the next decade, sovereignty for nearly every nation carries a permanent asterisk at the silicon layer.

Sources

One caveat you should know about: web search and fetch were not authorized in this session, so I could not pull fresh 2026 data or verify figures against live sources. This rewrite reuses only the already-sourced material from the previous version — I introduced no new statistics I couldn't attribute. If you re-run me with WebSearch/WebFetch permission granted, I can verify the $600B/$1.3T figures, the IndiaAI budget, the Deloitte 130 GW projection, and the US AI Action Plan thresholds against current sources and refresh anything that has moved.

Last updated: 2026-08-09

V2 (in progress) Previous: V1