India's LLM Roadmap: What It Actually Takes to Reach the Frontier

The foundation is real. Under the IndiaAI Mission—approved in March 2024 with an outlay of about ₹10,371 crore—the government selected 20 indigenous foundation models for support, sanctioned over 93 lakh GPU hours, and pooled tens of thousands of GPUs into shared compute.

The flagship results are concrete: Sarvam AI open-sourced 30B and 105B models trained entirely on mission compute, while BharatGen (IIT Bombay-led), Krutrim, and Tech Mahindra's Project Indus are all in the field.

The honest scorecard: India leads globally on Indian-language benchmarks—Sarvam wins roughly 90% of comparisons there—but sits well behind frontier GPT and Claude systems on English-centric reasoning and coding. The rest of this roadmap is about the four pillars that close that gap.

Frontier training is brutally compute-hungry—tens of thousands of top-tier GPUs running for months. This is India's hardest constraint.

Progress is genuine: the shared compute facility, 93+ lakh sanctioned GPU hours, and semiconductor manufacturing approvals worth ₹1.64 lakh crore all point the right way. But two gaps remain:

The priority is sustained expansion of sovereign compute—so Indian labs can train frontier-scale models at home without waiting in line for foreign capacity.

If compute is India's weakness, data is its superpower.

No other country has India's linguistic wealth: 22 official languages and hundreds of dialects, spoken by over a billion people. Western frontier models train overwhelmingly on English and cannot match this depth—which is exactly why Sarvam wins the overwhelming majority of Indian-language comparisons.

To turn that into a durable advantage, India should invest in:

This is the highest-leverage investment India can make. Compute can be rented; this data moat cannot be copied.

Many sovereign efforts worldwide adapt an existing open model—fine-tuning a Western or Chinese base rather than training from scratch. It is faster and cheaper, but it caps how far you can go and leaves you dependent on someone else's foundation.

Sarvam's from-scratch training matters precisely because it proves India can build foundations, not just decorate them. The next leap is training frontier-class models domestically—larger, more capable systems that compete on global reasoning, not only on Indian-language tasks.

That requires accepting cost and failure: frontier training runs are expensive and often do not work the first time. Countries that reach the frontier budget for that reality; those that only fine-tune stay a step behind by design.

Models are built by people, funded by money that can wait.

The IndiaAI Mission's structure—keeping IP with the applicants and subsidizing compute—is a good base. Sustaining it through political and market cycles is the real test.

India will not out-spend the United States or out-open China in the next few years, and pretending otherwise wastes energy. A realistic, ambitious target looks different:

Measured against citizens' lives—health, agriculture, justice, and education delivered in their own languages—this is the version of 'winning' that matters most. India does not need to be first at everything. It needs to be indispensable at the things only it can do.