Should India Build Its Own Sovereign AI Stack? The Costs and Trade-Offs

In June 2026, Anthropic suspended access to its newly released Fable 5 and Mythos 5 models after the U.S. Department of Commerce issued export control directives restricting access for foreign nationals. Access was restored three weeks later once the controls were lifted. For most of the world, this was a brief, technical news item. In India, it became the clearest possible illustration of an argument that had been building in policy circles for over a year: a country where 41% of workers reportedly use AI nearly every day — a higher share than in the US or China — had just watched, in real time, that its AI access ultimately runs through a foreign government's export control authority. As one widely circulated line put it: India called its AI sovereign. The U.S. government could still reach into it.

That's the emotional core of the sovereign AI debate. The economic core is harder and less satisfying: building genuine independence from that risk would cost India an amount of money it almost certainly doesn't have to spend, on a timeline that would leave it permanently behind the technological frontier it's trying to catch up to.

What "sovereign AI stack" actually means, and why the definition matters

There's no legal or technical standard for the term — it's a political and strategic concept, and different countries mean different things by it. In India's usage, it typically bundles together four distinct capabilities: owning or controlling the compute infrastructure rather than renting it entirely from foreign clouds; domestic semiconductor capacity; indigenously trained foundation models; and data and applications governed under domestic regulatory authority rather than a foreign jurisdiction's terms of service. A more precise framing, from recent Indian policy analysis, breaks the stack into five layers: energy infrastructure, chips, data centers, models, and applications — each with its own separate cost structure, timeline, and strategic urgency.

This distinction matters enormously for the cost question, because "does India need sovereign AI" and "does India need sovereignty at every layer of the stack" are completely different questions with completely different price tags.

The case for building it

The strategic logic isn't paranoid. AI is increasingly treated as infrastructure as consequential as electricity or telecommunications, and a country that controls none of the layers underneath it — chips, compute, models — is making a long-term bet that the countries that do control those layers will always act in its interest. India's current AI usage runs overwhelmingly through foreign foundational models accessed via Indian-built applications, which is precisely the arrangement the Anthropic export-control episode exposed as fragile. India's own AI Impact Summit in early 2026 signaled a shift in official thinking — policymakers no longer treating AI as a downstream consumer technology, but as a strategic capability worth building domestic capacity in, the same way earlier generations of Indian policy treated steel, telecommunications, or nuclear power.

Progress on the parts of the stack India has actually chosen to prioritize has been real. India's shared compute capacity crossed 45,000 GPUs by mid-2026, with 237 projects receiving subsidized compute support covering more than 93 lakh GPU-hours, and the government has selected 20 indigenous foundation model proposals — including 12 large multimodal models and 8 smaller language models — out of 506 applications received. Sarvam AI's recent funding round, backed partly by domestic capital, values the company at $1.5 billion. This isn't nothing. It's a genuine, government-coordinated attempt to build model and compute capacity domestically rather than simply consuming whatever foreign labs ship.

The case against trying to build all of it

The honest cost accounting is where the sovereignty argument runs into hard limits. Building a truly full-stack sovereign AI capability — competitive semiconductor fabrication, hyperscale compute, frontier-grade foundation models, and the applications layer on top — has been estimated to run into the hundreds of billions of dollars, a scale that outstrips what India's private capital markets and current public budgets can realistically sustain. For comparison, India's IndiaAI Mission carries a total outlay of roughly ₹10,372 crore (a little over $1 billion), while the newer Semicon 2.0 semiconductor program has been approved at nearly ₹1.28 lakh crore — more than twelve times larger, and still nowhere near what full domestic chip self-sufficiency would require. One striking data point on private capital's appetite here: HCL Technologies' recent $151 million investment in Sarvam AI, treated as a significant vote of confidence in sovereign AI, amounted to less than 10% of what the company paid out to its own shareholders as dividends in the same financial year. That's not a company under-investing out of malice — it's a fairly accurate signal of how the risk-adjusted returns on frontier AI infrastructure currently look to Indian capital allocators compared to safer, faster-return uses of the same money.

And the chip problem sits underneath all of it regardless of how much India spends on models and applications. India currently has no domestic capability to fabricate the advanced GPUs its AI compute runs on — the India Semiconductor Mission and Tata's Dholera fab remain restricted to assembly and packaging rather than the leading-edge fabrication that produces the chips actually running AI workloads. Every GPU in India's growing "sovereign" compute pool is, today, an American-designed and largely foreign-fabricated import. A February 2026 interim trade framework between the US and India includes language specifically protecting India's access to advanced AI chips — reassuring, but a negotiated trade provision, not a guarantee, and precisely the kind of arrangement that can be revisited the next time export control policy shifts, as it already did once in mid-2026.

The pragmatic middle ground the debate is actually converging on

The more sophisticated Indian policy analysis emerging through 2026 has largely abandoned the all-or-nothing framing in favor of a selective sovereignty approach: rather than trying to control every layer of the stack — an economically inefficient goal given how concentrated frontier chip and compute capability already is in the US and China — India should identify the specific layers where sovereignty delivers the most strategic and economic value relative to its actual capabilities, and accept continued dependency at the layers where trying to compete would be a poor use of scarce capital. That likely means continuing to build genuine capability in foundation models and applications, where India's talent base and market scale give it a real chance at competitiveness, while treating chip fabrication sovereignty as a much longer-horizon, differently-resourced problem that won't be solved by the same funding envelope or timeline as the model and compute layers.

What this means in plain terms

Should India build a sovereign AI stack? Yes, in the sense that reducing acute dependency at the layers where it can realistically compete — models, applications, and enough domestic compute to avoid total reliance on foreign clouds — is a defensible and increasingly urgent use of public and private capital, one the Anthropic episode made concretely visible rather than theoretical. No, in the sense that a full-stack sovereignty push, chasing semiconductor and hyperscale compute independence on the same timeline, would cost more than India can realistically deploy and would still likely leave it behind the frontier by the time it got there. The trade-off isn't between sovereignty and dependency. It's between the specific, affordable forms of sovereignty India can actually build in this decade, and the much larger, much more expensive form of sovereignty that stays out of reach regardless of how urgent it starts to feel every time a foreign export control decision reminds India how much of its AI stack still runs on someone else's terms.