India's Sovereign AI Race Has a Compute Problem Nobody's Solved
Krutrim, Sarvam, and Bhashini prove India can build AI models. None of them prove it runs on Indian hardware, and that gap is the whole story.

India has spent the last three years building its answer to ChatGPT, and by the metric everyone actually measures, it worked. Krutrim launched in April 2023 and became the country's first AI unicorn before most people had finished arguing about whether India needed its own model at all. Sarvam AI, founded by two IIT Madras researchers who'd spent years on Indic-language research at AI4Bharat, followed with a 105-billion-parameter model trained specifically for the country's 22 official languages. Bhashini, the government's own translation and speech layer, now runs underneath MyGov, CoWIN, and state portals serving over 140 million users, in languages no foreign frontier lab has bothered to support with any real accuracy: Gondi, Bodo, Maithili, Santali.
That's the win. Here's the part that doesn't get said out loud nearly as often: every one of those models runs on imported hardware. Krutrim Cloud, the compute layer behind India's flagship "sovereign" AI company, runs on Nvidia GPUs. Sarvam's 105-billion-parameter model was trained using IndiaAI Mission compute, which itself means renting time on a few thousand GPUs, none of them designed or built in India, for a few months. India's domestic chip design and manufacturing capability, the thing that would actually make any of this sovereign in a hardware sense, doesn't exist yet at meaningful scale. It's a decade-plus project even under an optimistic timeline, and the government's own roadmap agrees: the India Semiconductor Mission's public targets put the country at 70-75% domestic capability for chips covering existing applications by 2029, with the more advanced 3-nanometre and 2-nanometre process nodes, the ones modern AI accelerators actually need, pushed out to a "Semicon 2.0" phase with no firm date attached.
That gap isn't theoretical. It's under construction right now, and you can see exactly where it stops short. Tata Electronics, in partnership with Taiwan's Powerchip Semiconductor Manufacturing Corporation, is building India's first front-end wafer fabrication plant at Dholera in Gujarat, backed by roughly ₹91,000 crore, north of $10 billion, in committed investment. It's a genuinely serious facility: 50,000 wafer starts a month once it's running. It's also being built on a 28-nanometre process node. For comparison, the Nvidia GPUs actually training Sarvam's model, and every other frontier AI system on the planet, are built on 4-nanometre and 5-nanometre nodes, a full generation and a half more advanced than anything India's flagship fab is currently tooled to produce. Smaller nodes pack more transistors into the same die space, which is most of what separates a chip that can train a large model from one that can't. India isn't just missing advanced-node fabrication capacity today. Its single biggest, most-funded bet on fixing that doesn't reach advanced-node territory even once it's finished.
What "sovereign" actually means here
Words matter in this beat more than most. "Sovereign AI" has become shorthand for "a model trained on Indian data, tuned for Indian languages, hosted with data residency inside India's borders." All of that is real and all of it matters, particularly the data-residency piece, which solves a genuine problem: every query sent to a foreign model today leaves the country, gets processed under a foreign legal jurisdiction, and comes back priced in dollars. Krutrim and Sarvam fix that specific problem. Neither fixes the deeper one.
Training something competitive with a genuine frontier model, the kind OpenAI or Anthropic ships, requires cluster-months of upward of 10,000 H100-equivalent GPUs, at a hardware cost running into the hundreds of millions of dollars even before you account for the fact that India doesn't currently have access to that scale of compute at all, imported or otherwise. Sarvam's training run, a few thousand GPUs for a few months, is a real and serious engineering achievement. It is also nowhere close to that number.
Sovereign AI without domestic compute is a sovereign coat of paint on someone else's engine.
What actually gets built with ₹76,000 crore
None of this means the semiconductor push is theater. The India Semiconductor Mission has real money behind it, close to ₹76,000 crore, and real facilities breaking ground: Micron's assembly and testing plant in Sanand is already shipping made-in-India memory modules, and a handful of other assembly, testing, and packaging projects across Gujarat and Assam are moving from announcement to construction faster than most government-led industrial pushes manage. The Design Linked Incentive scheme has quietly funded a wave of chip-design startups too, with close to ₹430 crore in venture funding flowing to companies working on the design side rather than manufacturing. A fully indigenous 64-bit microprocessor, DHRUV64, already exists, built by a government-backed research center for use in 5G infrastructure and industrial automation.
What all of that adds up to is a real, growing capability in chip assembly, packaging, and mature-node manufacturing. None of it adds up to the specific thing "sovereign AI" needs, which is the ability to design and fabricate the advanced-node accelerator chips that large model training actually runs on. Those are two different industries wearing the same headline. India is building real strength in one of them. The AI story keeps borrowing the word "sovereign" from a project that, for now, is solving an adjacent problem.
The momentum question
There's a reason this story has quietly cooled since its loudest moment. The India AI Impact Summit in Delhi drew genuine energy, and Union Minister Amit Shah told the crowd Sarvam exemplifies why "the future belongs to India." Six months later, Inc42's own reporting on the sector noted the race to build a truly Indian LLM had lost some of its steam. Both things are true at once, and neither is a contradiction. The founders didn't stop building. The hype cycle simply ran its natural course, the way it does everywhere, and what's left standing afterward is the actual hard problem: models are the layer that's easiest to build fast and impressive to demo. Chips are the layer that takes a decade and doesn't demo well at all.
What to actually watch
The real signal isn't the next model release, and it isn't the next unicorn valuation. It's whether India's policy apparatus starts treating domestic semiconductor capability as seriously as it's treated the model layer, since one of those problems can be solved by a strong engineering team with a few years and a training budget, and the other one genuinely can't. Every rupee and every headline currently goes to the layer that's already working. The layer that would make "sovereign" mean what it says isn't getting the same attention, and won't build itself while everyone's watching the leaderboard instead.



