IndiaAI rents its GPUs by the hour. The numbers show who pays for them

If you have booked a cab on an app, you already know the business model. The platform owns no cars. It signs up people who do, sets a fare, and takes a cut. Since January 2025, the Government of India has run its AI compute the same way, except that it pays part of the fare.

Here is the ratio that frames everything. By August 2026, the IndiaAI Mission had sanctioned 93.18 lakh GPU-hours to 237 projects, per MeitY's reply in the Lok Sabha. Spread across the 45,000-plus GPUs PIB says sit in the pool, that is about 207 hours per GPU. Eight and a half days of work, for a pool that runs 8,760 hours a year.

The pool, counted three ways

The Cabinet approved the mission on 7 March 2024 with ₹10,371.92 crore over five years and a target of "10,000 or more" GPUs, PIB recorded. The state did not buy them. It ran an empanelment: private operators offer GPUs on their own clouds at a bid price, and IndiaAI resells access through a compute portal.

The count then moved fast. 18,693 GPUs after the first round in January 2025. 34,333 in May 2025, a figure PIB built on a restated base of 18,417. 38,231 "onboarded from 14 empaneled service providers" by March 2026. And "45,000+ … as of June 2026" in PIB's August round-up.

Read the labels, not just the bars. The first is a target. The rest are GPUs operators have committed to the pool, not GPUs anyone has seen running. The portal's own table of empanelled hardware sums to 45,081 units, of which 17,408, or 39%, are offered by one firm, Yotta Data Services.

The gap between committed and delivered is now on the record. On 2 October 2026 a senior official told BusinessLine the mission "currently has access to about 30,000 GPUs against commitments of 45,000," with some providers unable to deliver as equipment costs rose. That is an unnamed official, not a gazette. If it holds, a third of the pool is paper, and the opening ratio shifts: 93.18 lakh hours over 30,000 GPUs is about 311 hours each. Thirteen days, not eight and a half.

How the discount is built

PIB's headline rate is "approximately Rs 65 per GPU per hour, except for select high-end GPUs." The arithmetic behind that number shows who actually pays.

Take one line from the January 2025 bid annexure: a single NVIDIA H100 SXM, on demand. The lowest bidder's published market price was ₹480 an hour. Its bid was ₹170, a 64.58% discount. The state then pays up to 40% of the user's bill, per the ministry's release. So the user pays ₹102, the state pays ₹68, and the operator has already given up ₹310 against its own list price.

For every rupee of public subsidy on that line, the bidder conceded about ₹4.56. List prices are what list prices are; the bid is the real price. Still, most of the "below market" in this market came from competitive bidding, not from the treasury.

Where the money sits

Compute is 44% of the outlay, on the ministry's own breakdown. The money arrives slower than the GPUs. Budget allocations for the first three years total ₹1,973 crore (₹173, ₹800 and ₹1,000 crore), or 19% of the five-year outlay, per the August 2026 reply. Actual receipts were ₹21.79 crore in 2024-25 and ₹379.15 crore in 2025-26 as of 9 February 2026, according to a Rajya Sabha answer.

Who got the hours

The same February answer counted the users: 114 academic researchers, 47 startups and MSMEs, 58 government entities and smaller cohorts, 305 in all. By August the unit had changed to projects: 237. Divide the sanctioned hours by the projects and the average is about 39,000 GPU-hours each, the equivalent of one GPU running for four and a half years. Averages hide the shape. A national model-training run and a student's fine-tuning job each count as one project, and no published table splits the hours by user type.

The largest cheques sit in the foundation-model pillar. The government named 12 organisations in March 2026, including Sarvam AI, Soket AI, Gnani AI, Gan AI and the IIT Bombay-led BharatGen consortium, with assistance "to cover compute usage costs." On MediaNama's reading of the annexure, BharatGen received ₹1,058.52 crore, more than four times the next allocation. That is 54% of the ₹1,971.37 crore pillar, to one consortium. By August the shortlist had grown to 20 proposals.

The wholesaler starts buying stock

The voucher model buys speed: 45,000 committed GPUs in about 27 months, against a 10,000 target. It does not buy a domestic accelerator. The portal's hardware list is NVIDIA, AMD, Intel, Google and AWS silicon. The silicon story is told elsewhere.

It is also no longer the only model. MeitY has issued a purchase order for a roughly 1.1 exaflop AI system at the NIC data centre in Delhi. On 7 October 2026 C-DAC opened an empanelment for up to 10,000 dedicated compute units, hired for government and eligible users, MediaNama reported.

BusinessLine's official also described an 8-exaflop system, built with G42 and Cerebras, to be owned by the government as a gift from the UAE. No contract has been published.

What the numbers don't tell you

Utilisation. In February a Rajya Sabha member asked for the GPUs' "utilisation levels." The answer listed user counts and gave no utilisation figure. No release we found does.

Sanctioned is not consumed. PIB's August release says projects "accessed" 93.18 lakh GPU-hours; the Lok Sabha answer the day before says "sanctioned." Nobody publishes hours burned, per GPU type, per provider. Nor whether the foundation-model grants draw on the same pool, or how many committed GPUs sat idle while users waited for an allocation. Until those figures exist, ₹65 is a price, not a result. The cheapest GPU-hour is the one booked and never run, and that is the number missing from every table.