India’s Sovereign AI Push: Inside the IndiaAI Mission’s Race for Chips and Homegrown Models

India has long been celebrated as a software superpower, yet in the age of artificial intelligence the country has grown determined to own not just the code but the entire stack beneath it, from the chips that crunch the numbers to the large models trained on Indian languages and data.

That ambition is embodied in the IndiaAI Mission, a programme sanctioned by the cabinet in March 2024 with a budget of roughly ten thousand three hundred and seventy-two crore rupees, or about one and a quarter billion dollars, built around seven pillars ranging from compute and foundation models to safety and skills.

The hunt for compute

At the heart of the mission lies the unglamorous but decisive question of computing power, and here India has moved with surprising speed, deploying by the middle of 2026 around thirty-four thousand advanced graphics processing units across data centres and making them available at a heavily subsidised rate.

That rate, reported at roughly sixty-five rupees per GPU-hour, is a small fraction of what comparable computing costs on Western cloud platforms, a deliberate strategy to put serious firepower within reach of startups, academic researchers and government agencies that could never otherwise afford it.

The government has made clear this is only the beginning, announcing plans to add another twenty thousand units to reach fifty-four thousand in the near term, with a stated target of one hundred thousand public GPUs by December 2026 and private investment from giants like Reliance and Tata expected to push national capacity beyond two hundred thousand.

Homegrown models take shape

Raw computing power, however, is only useful if it produces something, and here too the mission is bearing fruit, with the government backing twenty indigenous sovereign model proposals from companies such as Sarvam, Gnani and BharatGen, of which several have already been released to the public.

The most closely watched debut came from Sarvam, which in February 2026 launched two open-source models, a thirty-billion-parameter system and a larger one of a hundred and five billion parameters that activates around nine billion parameters per token and offers an expansive context window.

What makes these models significant is not that they immediately rival the very largest Western systems, but that they are built in India, for Indian needs, with a deep focus on the country’s many languages, a capability that foreign models have historically served poorly or not at all.

Ambition against the odds

Yet formidable obstacles remain, because true sovereignty in AI ultimately depends on the ability to design and manufacture advanced chips, an arena still dominated by a handful of foreign firms, leaving India reliant on imports for the very hardware its mission runs on.

There are also the perennial challenges of talent, energy and the sheer cost of keeping pace in a field where the frontier moves every few months, meaning the state-backed effort must constantly run just to stay in place, let alone catch up with the leaders.

Even so, the scale and speed of the IndiaAI Mission signal a clear intent, that the world’s most populous nation refuses to be a mere consumer of artificial intelligence built elsewhere, and is instead determined to shape a future in which it is also a genuine creator.