India's AI Moment: Built Here, Built for Here

For most of the last decade, India's story in technology was about talent and scale: the engineers who powered the world's software, the payments rails that moved billions of transactions, the apps that reached a population larger than any other. In AI, the story is changing. India is no longer content to be the world's biggest user and back-office for other people's models. It is building its own.

Here is a look at the technologies and trends that define Indian AI right now.

1. Sovereign language models, trained at home

The clearest sign of the shift came in February, when New Delhi hosted the India AI Impact Summit, the fourth in the series of global AI gatherings that began at Bletchley Park and moved through Seoul and Paris. The summit was where India showed its first batch of homegrown foundation models.

Sarvam AI, a Bengaluru startup, introduced two large language models, one with 30 billion parameters and one with 105 billion, both trained in India. In March the company released both as open-weight models under the Apache 2.0 licence, so developers can download and build on them freely. Sarvam also opened a consumer assistant called Indus in limited beta during summit week, and it has been working with state governments, including a planned sovereign AI capacity hub in Odisha, on practical use cases like industrial safety and local-language skilling.

BharatGen, a consortium led by IIT Bombay, launched Param2, a 17-billion-parameter mixture-of-experts model built to work across Indian languages. It is a publicly funded effort, and the team has also been working on speech models and document-vision models aimed at governance, healthcare, education and finance.

The philosophy behind these projects is worth noting. Sarvam's co-founder Vivek Raghavan has argued that sovereignty matters more than building the biggest model. India's bet is not to win a parameter-count race but to build models that are efficient, open, and good at the things Indian users actually need.

2. Voice first, not text first

Most frontier AI has been trained mainly on text, and mostly English text. India has a different starting point: hundreds of millions of people who are more comfortable speaking than typing, and who speak in a language other than English.

That is why voice has become a signature theme. The IndiaAI Mission has openly talked about voice-first models as a way to differentiate Indian AI from Big Tech. Gnani.ai unveiled its Vachana voice stack at the summit, and Sarvam's models are being tied into voice-based citizen services, including work with UIDAI on Aadhaar-related interactions.

Underpinning much of this is Bhashini, the government's language platform, which offers translation and speech services across India's 22 scheduled languages and has already powered multilingual chatbots at large public events. For a farmer, a shopkeeper or a first-time smartphone user, an AI that listens and answers in their own language is the difference between a technology they watch and one they use.

3. Public compute and public data

Training models takes chips and data, and both are expensive. India's answer has been to treat them as shared infrastructure.

The IndiaAI Mission, approved in 2024 with an outlay of roughly ₹10,300 crore, has pooled a national compute facility that crossed 38,000 GPUs by late 2025. Startups, researchers and academics can rent that capacity at heavily subsidised rates, reported at around ₹65 per GPU-hour. The Mission has chosen a dozen teams to build foundation models, including Sarvam, Gnani.ai, BharatGen, Soket AI, Fractal Analytics and Tech Mahindra.

On the data side, AI Kosha is a government platform that makes non-personal datasets, such as agricultural, weather and public-sector data, available to AI builders. The Mission is also setting up IndiaAI Data Labs in Tier-2 and Tier-3 cities, spreading AI training and data work beyond the big metros.

4. AI for farms, clinics and classrooms

The most interesting Indian AI is often the least glamorous. The Union Budget for 2026-27 proposed Bharat-VISTAAR, a multilingual AI tool meant to bring AgriStack and ICAR's agricultural advice together so farmers can ask questions and get guidance in their own language. Similar thinking runs through education tools for personalised tutoring in regional languages, and healthcare pilots built on sector-ready models.

Industry voices have been saying the same thing: the phase of flashy experiments is ending, and the focus is moving to AI that is reliable, context-aware, and embedded in core business and public processes.

5. A light-touch approach to governance

India has so far chosen not to rush into a single sweeping AI law. Its AI governance guidelines take a risk-based, proportionate approach and do not permit unrestricted deployment of high-risk systems. The aim is to keep room for startups and open-source work while still setting guardrails where harm is most likely.

The honest caveats

It would be easy to cheer without qualification, so here are the open questions.

• "Sovereign" has limits. Sarvam's models are sovereign where it counts most, in keeping public data inside the country, but India still depends on imported chips and global supply chains.
• Bias is local too. Western AI debates tend to centre on race. In India, caste is a live concern, and developers will need to test for and reduce bias that generic benchmarks don't catch.
• Access must stay broad. Subsidised compute and open weights only deliver on their promise if smaller players, not just the biggest firms, can actually use them.
• Benchmarks are not adoption. Strong demos at a summit are a start. The real test is whether these models run quietly inside government offices, banks, clinics and farms a year from now.

What to watch next

Three things will show whether this momentum is real: how many developers pick up the open Sarvam and BharatGen models, how fast voice-based services reach people who have never used a chatbot, and whether the next generation of Indian models, which the Mission has said it wants to push toward much larger sizes, arrives on schedule.

India's advantage was never going to be spending the most. It is the scale of its users, the depth of its language diversity, and a habit of building public digital infrastructure that others can build on. If AI follows the path that UPI took, the most important Indian AI of the next few years may be the kind you hardly notice, because it simply works, in your language, on your phone.