Artificial Intelligence – India uses AI at huge scale but builds far less of it, and is now catching up.
- 100m weekly ChatGPT users (second-largest market after the US)
- 38,000+ GPUs on the state portal (from about ₹134 an hour for an H100)
- $676m AI startup funding, H1 2026 (four times H1 2025, across 57 deals)
- 3.0× AI skills against the world (the highest penetration of any country)
India uses more AI than almost anywhere on earth and builds less of it than its size suggests. Those are not two facts. They are the same one: the users arrived first, and everything since has been an attempt to catch up with them.
A hundred million Indians open ChatGPT in a given week. That is the second-largest user base on earth and the largest population of student users anywhere. None of it waited for an Indian model, an Indian data centre or an Indian rule book. It arrived on phones that were already there, over data that was already cheap.
The building is younger and far smaller. Indian AI startups raised $676m in the first half of 2026, four times the year before but tiny next to the single rounds two American labs closed in the same months. So the state is buying compute rather than backing capital: more than 38,000 GPUs on a common portal at roughly ₹134 an hour, and public money behind four attempts at a foundation model.
The signal to watch: Whether an Indian model ends up inside something people pay for. Parameters and GPU counts are inputs; production is the only output that counts.
Where we would take your group
- Sarvam AI, Bengaluru. Shipped 30- and 105-billion-parameter models in February. India’s second AI unicorn, on $234m.
- BharatGen, IIT Bombay. Param2: 17 billion parameters across 22 Indian languages, built with public money and Nvidia’s stack.
- Gnani.ai, Bengaluru. Vachana clones a voice from under ten seconds of speech, in twelve Indian languages.
Questions for the debrief
- India leads the world both in AI skill penetration and in AI talent leaving. Which of those two will still be true in 2030, and what would change it?
- The state is subsidising compute rather than funding models directly. Is cheap GPU time industrial policy, or a transfer to whoever was going to buy them anyway?
- Indian labs build for 22 languages and the low-end phone, constraints no frontier lab has to deal with. Is that a defensible moat, or a smaller market dressed up as a strategy?
