Two numbers describe the same country, nine days apart from Independence Day, and they point in opposite directions.

The first: India's public "sovereign" GPU capacity is on track to cross 100,000 units by December 2026, up from roughly 34,000 earlier this year, under the IndiaAI Mission. Maharashtra alone has approved an AI policy targeting over ₹10,000 crore in investment and 150,000 jobs by 2031, anchored by a ₹500 crore AI Startup Venture Fund. Officials at February's India AI Impact Summit talked, correctly, about India building AI capability on its own terms.

The second: the combined market capitalisation of TCS, Infosys, Wipro, HCL Technologies and Tech Mahindra has fallen more than 46% — from a record ₹33.71 lakh crore in August 2024 to around ₹18.15 lakh crore in July 2026 — as AI-driven disruption guts the outsourcing model that built the modern Indian IT industry. Infosys alone has nearly halved from its December 2024 peak. Wipro is down around 54% from its 2021 high.

Same technology, same year, same country. One number is a policy pitch. The other is a stock chart. Neither cancels the other out.

The word "sovereign" is doing more work than the infrastructure can back up

Every GPU inside the IndiaAI Mission's compute stack is imported. Not most — every one. There is no domestic chip manufacturing capacity feeding any part of India's AI build-out; the entire base sits on Nvidia, and increasingly AMD and Intel, hardware bought or leased from abroad. A "compute access framework" now lets the Ministry of Electronics and IT negotiate bulk procurement on behalf of Indian institutions — which is a real and useful piece of policy — but it's bulk buying, not manufacturing.

That's not a minor asterisk on the sovereignty story. It's the whole story. What India has actually built is:

  • Software and governance layers that move fast — 34,000-plus GPUs allocated, open-sourced language models across 22 Indian languages, an AI governance framework, and state-level policy like Maharashtra's moving in months rather than years.
  • A hardware layer that moves at the speed of an import order — every unit of compute underneath that software stack is subject to US export rules, the India-US trade relationship, and whatever Nvidia's allocation priorities happen to be that quarter.

Private capacity is arriving too — a Nvidia-Yotta Data Services partnership is bringing over 20,000 liquid-cooled Blackwell Ultra GPUs online, part of a $2 billion supercluster meant to cut India's dependence on overseas infrastructure. It is still built entirely from imported silicon, run inside India rather than manufactured there.

The other half of the picture: the industry AI is actively disassembling

While the state builds AI capacity, AI is dismantling the business model that made Indian IT a global export story in the first place. The mechanism is straightforward and already visible in quarterly numbers: clients now expect the productivity gains from AI-assisted coding and support work to show up as lower bills, not fatter margins for the vendor. Muted US technology budgets and hawkish Fed commentary have compounded it, but the structural piece — AI compressing the billable-hours model outsourcing was built on — isn't cyclical. It doesn't come back when rates fall.

Nifty IT has underperformed the broader market by roughly 20% since the start of 2026. That's not a sector having a bad year. It's a sector whose core product is being repriced by the same technology India is simultaneously trying to build sovereign capacity in.

Why both things are true without contradiction

A government building compute infrastructure and startups raising money doesn't require the IT services majors to be doing well — they're different parts of the AI economy, one consuming compute and capital, the other losing pricing power to AI-assisted competitors including, ironically, companies that could eventually run on that same sovereign infrastructure.

But "AI sovereignty" as a phrase implies control over the full stack, and right now India controls two of the three layers — the models and the policy — while renting the third at whatever price and priority the GPU export market allows. That's a real capability. It is not sovereignty in the sense the word usually carries, and conflating the two makes it harder to see what actually still needs building: not another incubator, but a chip.