1 November 2026 • 6 min read
Training or fine-tuning a model on rented GPUs is the biggest bill many Indian AI startups see. The IndiaAI Mission runs a government portal that lets approved startups, researchers and students rent GPUs from empanelled cloud providers at a subsidised rate. This page sets out what the government says it offers, who qualifies, how an application moves, and what to check before you budget around it.
The government launched the IndiaAI Mission with an outlay of ₹10,372 crore. A PIB release of 25 March 2026, a Lok Sabha reply, says more than 38 thousand GPUs for a common compute facility have been onboarded through the AI compute portal and are being provided to Indian start-ups and academia at an affordable rate.
On price, a PIB backgrounder of 10 February 2026 says the 38,000+ GPUs "are available at ₹ 65 per hour, which is nearly one third of the global average cost". It also reports 1,050 TPUs onboarded. A later PIB factsheet (13 August 2026) says shared compute capacity reached 45,000+ GPUs as of June 2026, and that by August 2026 237 projects had used subsidised computing, covering 93.18 lakh GPU hours.
Treat ₹65 as a headline average, not your price. The portal lists different rates for different chips and cluster sizes, shown below.
IndiaAI publishes a "Ready Reckoner" that maps use cases to GPU configurations. Prices are on-demand, per hour, as of 13 March 2026 (the AMD MI325X list is dated 27 February 2026). The document does not say whether these are before or after the subsidy, so confirm this on the portal.
| Instance | Typical use named by IndiaAI | Price per hour (₹) |
|---|---|---|
| L4, 1 GPU | Small LLM inference, translation transformers | 44.86 |
| L40S, 1 GPU | Medium LLM inference, face verification | 67.5 |
| A100 80GB, 1 GPU | Light GenAI (Llama 8B), X-ray inferencing | 135.9 |
| H200 SXM, 1 GPU | Online document summarisation, ASR fine-tuning | 140.0 |
| MI300X, 1 GPU | Memory-intensive LLMs, RAG, agents | 168.22 |
| B200 SXM, 1 GPU | VLM training, high-concurrency inference | 323.0 |
| H100 SXM, 8 GPUs | Foundation model training, 70B to 100B | 1224.0 |
Two readings follow. A single-GPU instance for inference or light fine-tuning costs a few hundred rupees an hour at most. A full eight-GPU node for pre-training costs more than ₹1,200 an hour, so a multi-day run adds up even at these rates.
The portal's eligibility page groups applicants. For startups and MSMEs it lists: DPIIT recognition, experience in AI/ML, and "Annual rev 50 Lacs or Funding of INR 100 Lacs". Its description of DPIIT recognition states that the entity should be incorporated as a private limited company, a registered partnership firm or an LLP, be not older than 10 years, and have annual turnover not above INR 100 crore in any previous financial year.
There is also a track for early-stage startups. The page lists "4 Years experience in AI/ML (including research)" and a letter from an IndiaAI CEO or officer of joint-secretary rank or above. The wording on the page is terse, so read the live page for the exact document you need. Researchers, students, IndiaAI fellows and government entities have their own tracks.
PIB's March 2026 reply shows who has been approved so far: of 190 projects, 46 are startups and MSMEs and 30 are early-stage startups. 78 are government entities.
According to the portal's FAQ:
On cost: IndiaAI subsidises approved projects, but users must bear any cost beyond the subsidy. You pay the cloud provider directly, on that provider's billing cycle.
Compare three options on the same run before you apply: the portal rate you would be approved for, a commercial Indian GPU cloud, and buying hardware. Put the portal's monthly cycle and 30-day window into the schedule, because they set when you can start. If your run cannot wait for the next approval list, the cycle may cost you more than the subsidy saves.
Eligibility rules and rates change. Check with the portal or an adviser before you commit a budget.
PIB's backgrounder of February 2026 gives ₹65 an hour for 38,000+ GPUs, described as nearly one third of the global average cost. The portal's own rate list varies by GPU type and cluster size, so use that list for a budget.
For the startup and MSME track the portal's eligibility page lists DPIIT recognition, AI/ML experience, and either annual revenue or funding above a stated floor. Check the live page for the current criteria.
Requests go in between the 1st and 25th of a month, and PMEC publishes the approved list on the 10th of the next month or the next working day, according to the portal FAQ.
No. It is subsidised access to rented compute. Users pay costs beyond the subsidy to the cloud provider.
Related on this site: LLM costs at scale and building AI features for Indian apps.
Nine stories, every morning at 09:15 IST.
Get the brief free →