By the numbers
Rs 10,372 crore total outlay, mission launched March 2024
38,000+ GPUs onboarded through 14 AI service providers, with 20,000 more under process
190 projects approved - 78 government, 46 startups and MSMEs, 30 early-stage startups, 27 academic
12 teams shortlisted in phase one to build indigenous foundation models
4 models launched at the IndiaAI Impact Summit 2026 - Sarvam AI, BharatGen, Gnani and Soket
Rs 988.6 crore to BharatGen, the largest single share of a Rs 1,500 crore allocation
India's national AI programme is now large enough to be compared with private funding rounds, and most coverage of it repeats announcements rather than figures. The numbers below come from Press Information Bureau releases, which is to say from the government's own record rather than from reporting about it.
Where the money is
| Figure | Amount | Source and date |
|---|
| Total mission outlay | Rs 10,372 crore | PIB, 25 March 2026 |
| Foundation model allocation | Rs 1,500 crore | Reported on the MeitY disbursement |
| BharatGen's share | Rs 988.6 crore | The largest single share of that allocation |
| Remaining foundation model funding | Not published per company | No per-team breakdown has been released |
| AI Impact Summit 2026 commitments | About $250 billion in AI infrastructure | PIB, 13 March 2026. Private commitments, not public spending |
The gap in the public record. BharatGen's Rs 988.6 crore is the only per-company figure that has been published. The remaining Rs 511.4 crore of the Rs 1,500 crore foundation model allocation is split among the other selected teams in proportions nobody has disclosed. Anyone telling you what Sarvam or Gnani received is inferring it.
The compute pillar, which is the real subsidy
The money gets the headlines but the compute is what actually changes the economics for an Indian startup. Training or serving a model at commercial GPU rates is the single largest cost in the business, and the mission's answer is a common compute facility offered below market.
| Compute | Figure | Status |
|---|
| GPUs onboarded for the common compute facility | More than 38,000 | Live, through 14 AI service providers |
| Additional GPUs under the mission | 20,000 | Under process as of 13 March 2026 |
| GPU hours sanctioned for research | 93.18 lakh | Across 237 projects, reported 23 July 2026 |
| Who can use it | Indian startups and academia | Offered at a subsidised rate |
Thirty-eight thousand GPUs is a serious number in absolute terms, though the comparison that matters is not to other national programmes. A single large private cluster in the United States runs to six figures of accelerators. What this buys India is not frontier-scale training capacity but a floor under its startup ecosystem: enough compute that a team with a good idea and no capital can get a model trained.
Who got the 190 projects
| Recipient | Projects |
|---|
| Government entities | 78 |
| Startups and MSMEs | 46 |
| Early-stage startups | 30 |
| Researchers and academia | 27 |
| Students | 5 |
| Early-stage researchers | 4 |
| Total | 190 |
The distribution is the interesting part. Government entities took 78 of 190, more than startups and MSMEs at 46 and early-stage startups at 30 combined. Whether you read that as the state seeding its own adoption or as crowding out depends on your priors, but it is a materially different profile from a programme designed purely to fund private innovation.
The twelve teams, and the four that shipped
Twelve teams were shortlisted in phase one to build indigenous foundation models and large language models. Four had models launched at the IndiaAI Impact Summit 2026: Sarvam AI, BharatGen, Gnani.ai and Soket AI Labs, which the PIB release spells Socket. Separate reporting also names Avataar as a selection for a video generation model.
Shortlisted and shipped are different things, and the distinction is worth holding onto. Twelve teams were selected; four models have been launched publicly. The remaining eight may be in progress, may have been quiet releases, or may not have produced anything yet - the releases do not say, and the full list of twelve has not been published.
What you can use today. Of the four, only Sarvam has published both open weights and pricing. Sarvam-30B and Sarvam-105B went out under Apache 2.0 in early March 2026 with weights on Hugging Face and AIKosh, and its API lists rates from Rs 15 per million input tokens. For the rest, access terms are not published in a form we could verify. That is the honest state of play and it is the gap between a model existing and a model being useful.
What the Summit number is and is not
PIB puts investment commitments announced around the IndiaAI Impact Summit 2026 at roughly $250 billion in AI infrastructure. That figure has been widely quoted in a way that blurs it into the mission's own spending, and it should not be. The Rs 10,372 crore is public money committed and being disbursed. The $250 billion is private commitment announced, which is a different kind of number with a different relationship to reality.
The Union Budget for 2026-27 added a long-term tax holiday for data centre and cloud infrastructure investment, which is the lever most likely to decide whether any meaningful share of that $250 billion actually lands.
How this compares
Rs 10,372 crore is a little over a billion US dollars at current rates. Set against private AI funding it is modest - several individual rounds raised in 2026 were larger, and India's own best-funded AI startup has raised $524 million on its own. Set against what it buys, it reads differently: a national compute facility, 190 funded projects and four shipped foundation models for roughly the cost of one large private round.
For a fuller picture of the companies involved, see our ranking of India's top AI companies by what they actually ship.
FAQ
How much is the IndiaAI Mission worth?
Rs 10,372 crore, per a PIB release of 25 March 2026. The mission launched in March 2024. At current rates that is a little over a billion US dollars, which is less than several single private AI rounds raised in 2026 and more than any other Indian public technology programme of its kind.
Which companies are building India's foundation models?
Twelve teams were shortlisted in phase one. PIB names four whose models launched at the IndiaAI Impact Summit 2026: Sarvam AI, BharatGen, Gnani and Soket, the last spelled Socket in the release. Separate reporting also names Avataar for a video generation model. The full list of twelve has not been published in the releases we read.
How much did BharatGen get?
Rs 988.6 crore, reported as the largest single share of MeitY's Rs 1,500 crore foundation model allocation. No comparable per-company figure has been published for the other teams, so BharatGen's share is the only one that can be stated.
Can startups actually get the GPUs?
That is the stated purpose. More than 38,000 GPUs have been onboarded through 14 AI service providers for a common compute facility offered to Indian startups and academia at a subsidised rate, with 20,000 more under process as of March 2026. Separately, 93.18 lakh GPU hours were sanctioned across 237 research projects as of July 2026.
Are any of these models actually usable?
Sarvam's are. Sarvam-30B and Sarvam-105B were open-sourced under Apache 2.0 in early March 2026 with weights on Hugging Face and AIKosh, and Sarvam publishes per-token API pricing. For the other teams, model access and terms are not published in a form we could verify.
Why do the project counts differ between sources?
Because they count different things at different times. PIB reported 190 approved projects as of 25 March 2026, broken down by recipient. Separate reporting in July 2026 referred to 237 research projects receiving GPU-hour allocations. Those are not the same measure and we have not tried to reconcile them.
Sources
PIB, 25 March 2026 - outlay, GPUs and the 190 approved projects ·
PIB, 13 March 2026 - Rajya Sabha reply on foundation models and GPUs ·
PTI via The Wire on BharatGen's allocation ·
BharatGen ·
Sarvam AI API pricing