Let us start with South Korea, because it is where the fault lines of AI mania are first becoming evident. Its Kospi fell by almost a third from its June peak as investors began to doubt that the world’s AI build-out could be paid for. Then, on 31 July, it rose 17.9% in a single session—the largest one-day gain in the index’s history—after the largest American cloud providers re-affirmed that they would keep spending. Hundreds of billions in market value vanished and returned within a fortnight.
Note what the Kospi actually is. Two firms, Samsung and SK Hynix, account for more than half its value and the bulk of its listed earnings. They make the memory that trains and runs advanced models. So the index did not merely react to the AI cycle. It is the AI cycle.
An entire index swings wildly in accordance with global AI dynamics. That is the threshold this technology has crossed. It is no longer a matter for technology policy. It has become a foundational economic force and so an object of economic and geopolitical strategy.
The scale is now hard to wave away. AI-related firms account for more than 80% of the S&P 500’s gains this year; strip them out, and the index has barely moved. Global AI spending is projected to pass $2.6 trillion in 2026, and data-centre infrastructure alone may approach $2.9 trillion by 2028. That capital is already lifting US growth and pulling semiconductors, power generation, cooling systems, land and specialized services along.
The risks scale with the spending. The Kospi episode showed that AI volatility can move entire national markets and the real economies behind them. Worse, much of the spending is circular. Hyperscalers fund the chipmakers who supply the same hyperscalers; model firms and their backers sit on both sides of the same contracts.
Circular financing flatters the boom and conceals fragility. When confidence turns, the loop could run backwards with equal force. A system this concentrated and self-referential is unlikely to correct gently.
On their part, governments have drawn one clear conclusion: AI can be a source of power. The contest between the US and China has expanded from who creates the cleverest model to who controls the compute, chips and supply chains that decide who can build and run intelligence at scale.
Export controls on advanced chips, weights and equipment have tightened. China has pushed its chip fabrication forward while releasing capable open-weight models at a fraction of the cost. Each side wants less dependence and more leverage. The landscape is breaking into blocs and this has given us a new vocabulary of AI sovereignty.
Here, the word does more harm than good, so it is worth being exact. Sovereignty is not autarky. No country can build the whole stack. The capital needed is too large, talent too scarce and current lead times too long. A nation that tries to make everything would end up poorer and no freer. Independence at every layer would be an expensive weakness rather than a strength.
Sovereignty is strategic autonomy: the capacity to decide for oneself without being coerced through dependence on someone else’s chips, clouds, code or courts. It is the freedom to choose how intelligence is built, deployed and governed within one’s borders—and the confidence that this freedom cannot be switched off from outside.
That confidence has to be bought, and at a price that sovereignty talk often skips.
First, we must turn what we have into leverage that others need. Clean power, technical talent, trusted data, particular manufacturing skills and critical minerals are bargaining assets only if a country makes itself relatively indispensable, globally. Leverage, not isolation, is the working currency of autonomy. A country with nothing others need is not sovereign, as it can be ignored.
Second, we must honestly choose the few layers where homegrown capability is non-negotiable, and fund them with patience. The chokepoints are known: lithography, leading-edge fabrication, high-bandwidth memory, the software that locks in dominant chips and the refining of the minerals beneath it all. Each takes five to ten years to build, or longer. The money and resolve must be committed now and defended through political cycles.
Third, and this deserves the most candour, while that capability matures, a slow process, supply has to be secured through partnerships. Coalitions that pool investment, spread supply sources and fend off coercion across chips, energy, minerals and compute offer a practical route to security.
Interdependence, however, is never neutral. A coalition assembled by one great power to cut reliance on another establishes an order, with a centre. Sheltering under it would trade the coercion feared from one country for structural dependence accepted on a second. That can be a shrewd bargain. It is not the same as autonomy.
So, an honest position holds two things at once. No nation will command every layer of the stack, and none needs to; the aim is real influence over the few layers that decide whether a country can innovate, compete and govern on its own terms. But every nation reaching for that by building its own compute, data centres and models also adds to the glut of capacity that made Korea’s market swing by a fifth in a day.
Autonomy pursued by every country at once inflates the very bubble whose fragility made autonomy feel urgent. The cure and the disease thus share a supply chain. The madness of crowds is no longer confined to stock markets.
There is no clean version of this. Even success buys a qualified freedom—held inside somebody else’s order, resting on an edifice that can shake.
That is what sovereignty in the age of artificial intelligence really amounts to. Not self-sufficiency, which is a fantasy; nor the shelter of a patron, which is only dependence under another name; but the narrower and harder discipline of knowing precisely which dependencies one can live with and which one cannot.
These are the authors’ personal views.
The authors are, respectively, chief economic advisor to the Government of India and distinguished fellow, Niti Aayog.