On 12 June, the United States Department of Commerce gave Anthropic, one of America’s leading artificial intelligence companies, 90 minutes to restrict its two most powerful models to US citizens. Rather than building nationality checks at that notice, the company withdrew the models — Claude Fable 5 and Mythos 5 — from the market altogether. Overnight, researchers, doctors, educators, and developers in more than a hundred countries, India among them, lost access to tools many had built their work around.
For everyone outside the United States, it was a live demonstration of what dependence on foreign AI infrastructure now means: access can be switched off by a government you do not elect, in roughly the time it takes to read this article.
India should study that demonstration closely, because it is more exposed than its recent progress suggests. The IndiaAI Mission has assembled a shared compute pool of more than 34,000 graphics processing units, and home-grown models such as Sarvams, trained on Indian hardware across 22 Indian languages, are genuine achievements. But the structural picture is unchanged. Only two countries — the US and China — are building full-stack AI sovereignty: the chips, models, clouds, energy, talent, and legal frameworks that together amount to real independence. Everyone else, India included, lives in managed dependence. Every Indian GPU is imported; the frontier systems Indian developers benchmark against, and the clouds that serve them, are controlled abroad. When access is withdrawn, as it was this month, there is no domestic frontier model to fall back on — only the hope that diplomacy restores what a directive removed.
A risky dependence
The closest precedent is the “exorbitant privilege” the US has enjoyed for half a century as issuer of the world’s reserve currency — a dependence that has shaped everything from borrowing costs to exposure to sanctions. AI dependence reaches deeper still, into public health, the courts, agricultural planning, and national security. And it is harder to hedge. The semiconductor export controls that Washington has tightened since 2022 bite over months, leaving room for supply chains to adjust. A model-access directive bites instantly. Software has no supply chain and no ready substitute. Access to AI models is, in that sense, a riskier dependence than access to chips — and India’s strategy still rests heavily on rented frontier access.
The instinct will be to answer with self-reliance — to build a sovereign frontier laboratory of India’s own. That instinct is right in spirit and unaffordable in fact: the capital, compute, talent, and time a frontier lab demands put it beyond all but a couple of states. The realistic answer is collective: a coalition of complementary economies that together hold enough of the AI stack to constitute a distributed third option, neither American nor Chinese. Call it collective programmable sovereignty. No single member owns the whole stack; between them, they own enough of it to matter.
India would not be a junior partner in such an arrangement; it would be one of its anchors. The coalition’s compute could sit in Taiwan and South Korea, its energy in the Gulf, its regulation and foundational research in Europe, its largest consumer markets in Brazil and Indonesia — and its data, talent, and applied dynamism, in very large part, in India. India’s digital public infrastructure, already operating at population scale, is precisely the kind of asset that turns raw capability into AI that is actually deployed and used. Its engineering talent, at home and across a global diaspora, is the resource most other members of such a coalition lack.
The arithmetic is favourable. The combined output of such a group would exceed that of the US, giving it the weight to negotiate with global AI providers on its own terms — much as the European Union’s single market has turned its data-protection rules into a standard that American firms observe, whether they like it or not. The building blocks already exist. What is missing is the political decision to assemble them before the next directive arrives, rather than after it.
Also read: We need an AI pause button. It’s time to stop and think
Taking the lead
Encouragingly, the first moves in this direction are already visible. At this month’s G7 summit in France, Prime Minister Narendra Modi used the artificial-intelligence session to press for common standards and to speak for a Global South that has no wish to be locked into either bloc. Days earlier, India and France — responding directly to the Anthropic curbs — agreed to establish a joint working group on AI governance. These are modest beginnings, but they are beginnings of the right kind: complementary economies choosing to coordinate rather than each going alone.
This is not a call for autarky, nor for decoupling from American or Chinese technology, which would be neither possible nor wise. It is a call for resilience. Sovereignty in the AI era does not require owning every chip or training every model. It requires the capacity to create alternatives that can genuinely be switched on — to reduce systemic dependence while remaining integrated in the global system. Just as India diversifies its energy suppliers and its foreign-exchange reserves rather than relying on any single source, it must now diversify its access to artificial intelligence itself.
As the world’s largest democracy and an increasingly confident voice for the emerging economies, India is unusually well–placed to convene such a coalition rather than merely join one. The question raised by the June directive is not whether something like it will happen again; it will. The question is whether India and its partners will have built a plural, resilient AI architecture by the time it does. That architecture will not assemble itself. India should begin building it now.
Alicia García Herrero is Chief Economist for Asia Pacific and the Middle East at Natixis, and Senior Fellow at Bruegel. She tweets @Aligarciaherrer. Soumitra Dutta is Director of the Portulans Institute. Views are personal.
(Edited by Aamaan Alam Khan)