Cheap GPUs for AI: How to Rent an H100 in India for ₹150/Hour

If you have ever wanted to train an AI model but flinched at the price of a single high-end graphics card, India's government has quietly built an answer. Through the IndiaAI Mission, students, researchers, startups and small businesses can now rent H100-class GPUs at heavily subsidised rates — in some cases around ₹150 per GPU-hour or less. This guide breaks down what IndiaAI GPU compute actually is, who qualifies, what it costs, and the exact steps to get hours allocated to your project.

Why GPU access is India's real AI bottleneck

The public conversation around AI obsesses over models and chatbots. The unglamorous truth is that the binding constraint for most Indian builders is compute — the racks of GPUs that turn an idea into a trained model. A single NVIDIA H100 card can cost tens of lakhs to buy outright, and renting one from a global cloud often runs roughly ₹210 to ₹335 per hour once you convert the typical $2.5–$4 rate.

For a bootstrapped startup or a PhD scholar, those numbers are a wall. A serious training run can chew through hundreds or thousands of GPU-hours, and the bill arrives whether or not the experiment works. That is precisely the gap the IndiaAI Mission set out to close — not by handing out free chips, but by pooling national demand and subsidising the price.

What the IndiaAI compute facility actually offers

The headline is scale. The mission has already deployed over 38,000 GPUs across a network of private providers, and the government has signalled a target of 100,000 GPUs by the end of 2026. A meaningful slice of the current fleet is high-end H100 and H200 silicon — the same class of hardware that powers frontier models abroad.

Instead of building one giant government data centre, the mission empanelled 14-plus service providers who supply the GPUs and run them under defined service-level agreements. You don't deal with the hardware directly; you book capacity through a single national portal and the empanelled agency delivers it on the cloud.

The pricing is the part worth bookmarking:

• The lowest successful bids for top-tier GPUs landed near ₹150 per GPU-hour.
• Some reported rates for specific configurations dip well below that once subsidies apply.
• Eligible projects — especially those in areas of national importance — can claim up to 40% off the cost.

Stack those together and you get frontier-grade compute at a fraction of what a commercial cloud charges, which is the whole point.

Who can rent subsidised GPUs

The access list is deliberately broad. This is not a scheme reserved for a handful of funded unicorns. Eligibility extends to:

1. Researchers and academic faculty at recognised institutions.
2. Students and PhD scholars, including IndiaAI fellowship awardees.
3. DPIIT-registered startups building AI products.
4. MSMEs that want to add AI to their operations.
5. Government entities running public-interest projects.

Each category has its own documentation requirement to prove you are who you say you are — a startup shows its DPIIT recognition, a student shows institutional enrolment, and so on. The common thread is that the subsidy is tied to a genuine, stated purpose rather than open-ended speculation.

How to apply, step by step

The process runs entirely through the IndiaAI compute portal, and it follows a clean four-stage flow. Knowing it in advance saves you from a lapsed approval.

1. Register using Meri Pehchaan. You sign in through India's unified login — via DigiLocker, Parichay or ePramaan. This ties your account to a verified identity from the start.
2. Submit eligibility documents. Fill the registration form and upload the papers that match your category. A designated verifying official then checks them.
3. Wait for verification. Your registration is confirmed only after that official clears it, so use accurate, current documents to avoid back-and-forth.
4. Submit a project proposal. This is the decisive step. You spell out your technical approach, the expected impact, the number of GPU-hours you need, whether you are seeking a subsidy, and an estimated bill of materials.

Allocation then depends on three things: your eligibility tier, the GPU-hours you ask for, and whether you want the subsidised rate. The stronger and more specific your proposal, the easier it is to justify a larger allocation.

The catch most applicants miss

There is one rule that trips up newcomers: an approval is valid for roughly 30 days, and you must start using the service within that window or it expires. The subsidy is not a credit that sits in your account indefinitely — it is a booking against real, finite hardware that someone else could be using.

In practice that means you should treat the application as the last step, not the first. Have your training code working on a small machine, your dataset cleaned and uploaded, and your experiment plan written before you click apply. Builders who request hundreds of hours and then spend two weeks debugging their data loader can watch the clock run out on a subsidy they fought to get.

A few habits that protect your allocation:

• Dry-run locally or on a small instance so the expensive GPUs only run code that already works.
• Checkpoint frequently so a crash doesn't waste subsidised hours.
• Estimate honestly — padding your bill of materials wastes capacity and weakens future requests.

Why this matters beyond the discount

Cheap compute is not just a line-item saving; it changes who gets to build. When an H100-hour costs less than a meal, a final-year student in a tier-2 college can fine-tune a model for a regional language, and a two-person MSME can prototype a vision system for its factory floor. That widening of access is the strategic bet behind the IndiaAI Mission — turning compute from a luxury into infrastructure.

There is healthy debate about the approach. Critics warn that subsidising imported GPUs could blunt the incentive to build domestic chips, and that ultra-cheap compute can attract low-value usage. Those are real tensions worth watching as the fleet scales toward six figures.

But for a builder today, the calculus is simple. The hardware that defines the AI race is sitting in Indian data centres, rentable through a single portal, at prices that finally make experimentation affordable. The barrier was never the idea — it was the GPU-hour. For the first time, that barrier has a government-backed discount attached to it, and the only thing standing between you and a training run is a well-argued proposal.