IndiaAI Mission GPUs: a builder's guide to sub-₹100/hour subsidised compute

What this changes for builders

• The capacity is real now — India's common compute pool has crossed 34,381 GPUs onboarded from 14 empanelled service providers, not a press-release promise.
• The rate is genuinely cheap — about ₹65 per GPU-hour on average; roughly ₹92 for an H100. That undercuts global cloud list pricing by around 70%.
• The subsidy stacks — eligible projects can claim up to a 40% cost reduction, pushing approved national-importance work below ₹100 per GPU-hour.
• It is not self-serve — access runs through the IndiaAI Compute Portal as a registration-and-approval process. Plan for a proposal and a wait, not a credit-card checkout.
• The lesson is global — subsidised sovereign compute is now a pattern across India, the UK and beyond. If you build anywhere, this is worth understanding.

The numbers, verified

Let us pin down what is actually on offer, because secondary coverage has been sloppy with the figures. According to the government, India's common compute capacity has crossed 34,000 GPUs — precisely 34,381 reported — onboarded from 14 empanelled service providers under the IndiaAI Mission. That Mission is a ₹10,000 crore (roughly $1.25bn) programme, and shared subsidised compute is its most builder-relevant pillar.

On pricing, the average subsidised rate sits at approximately ₹65 per GPU per hour. The newer, training-grade silicon costs more: an H100 — the card you would reach for to train or heavily fine-tune a foundational model — runs at about ₹92 per GPU-hour. For projects that clear the approval bar as nationally important, the effective rate after subsidy lands below ₹100 per GPU-hour. Taken together, IndiaAI compute pricing undercuts global cloud providers by roughly 70%.

On top of the base rate, the government provides up to a 40% cost reduction for eligible applicants — researchers, startups, MSMEs and academic institutions — administered by a committee under the Ministry of Electronics and Information Technology (MeitY). The empanelled providers carrying this capacity include Yotta, E2E Networks, NxtGen and Jio Platforms, among others.

Who qualifies, and how to apply

This is a guide, so here is the path end to end. The single most important thing to internalise is that IndiaAI compute is project-gated. You are not buying anonymous capacity; you are getting subsidised access for a specific, approved piece of work.

Step 1 — Confirm eligibility. The scheme is aimed at researchers, startups, MSMEs and academic institutions. A solo Builder with a serious research or product proposal can qualify; a vague "we want cheap GPUs" pitch will not.

Step 2 — Register on the IndiaAI Compute Portal. Access runs through the official portal on the IndiaAI compute hub. Create an organisation account and complete the verification fields.

Step 3 — Submit a project proposal. Describe the workload concretely: the model and parameter count, the dataset and its provenance, the GPU type and count you need, the number of hours, and the timeline. A precise, costed proposal moves faster than a speculative one.

Step 4 — Await approval and your subsidy tier. The MeitY-linked committee assesses the proposal and assigns the applicable subsidy, up to 40%. National-importance work is where the full reduction sits.

Step 5 — Book hours with an empanelled provider. Once approved, you select capacity from a provider such as Yotta, E2E Networks, NxtGen or Jio Platforms and schedule your run. The subsidised rate is applied to the approved allocation.

Worked example — a fine-tuning run, costed

Numbers settle arguments. Take a realistic mid-sized job: fine-tuning a 13B-parameter model on a domain dataset, sized at 8×H100 for 100 GPU-hours each — 800 GPU-hours total. We will use the verified ₹92/GPU-hour H100 rate and compare against a representative global cloud on-demand price for an equivalent 8×H100 node.

Scenario	Rate (per GPU-hour)	GPU-hours	Total cost	Vs IndiaAI
IndiaAI H100 (base)	₹92	800	₹73,600	baseline
IndiaAI H100 + 40% subsidy	₹55.20	800	₹44,160	40% cheaper still
Global cloud on-demand (≈70% dearer)	≈₹307	800	≈₹245,000	~3.3× the base cost

The base IndiaAI run lands at ₹73,600. The same job on representative global on-demand cloud pricing — roughly 70% higher per GPU-hour — comes in near ₹245,000. That is a saving of about ₹171,000 on a single 800-hour run, before you even reach the 40% subsidy. Add the subsidy for an approved national-importance project and the bill falls to roughly ₹44,160 — under a fifth of the global-cloud figure. For a seed-stage Indian startup, the difference between ₹44k and ₹245k per fine-tuning cycle is the difference between iterating monthly and iterating once.

A note on honesty in the maths: the global-cloud figure is a representative on-demand list price for illustration, not a quote. Committed-use discounts, spot capacity and falling H100 rates narrow the gap — we have covered the broader trend in our look at Nvidia Rubin's 10× inference cut and the new unit economics. But even on a conservative reading, subsidised IndiaAI compute is in a different price band.

What UK and global builders should take from this

It would be a mistake to file this as India-only news. The deeper story is that subsidised sovereign compute has become a global pattern, and the structure of each programme tells you how to engage with it.

The UK runs a comparable instinct with a different shape. Its Sovereign AI work concentrates large allocations — on the order of a million GPU-hours — into a small number of selected firms, alongside national supercomputing capacity. India, by contrast, spreads a shared pool thinly across many smaller teams at a published per-hour rate. One model picks winners and hands them deep capacity; the other lowers the floor for everyone who can write a credible proposal. Neither is strictly better — but if you are a small UK or European team, the Indian model is the one that hints at where you might lobby for access.

The practical takeaway for any builder, anywhere: public compute programmes are now a line item worth planning around, not a curiosity. Before you size a training budget on commercial cloud alone, check whether a sovereign or national scheme covers your jurisdiction and your use case. The same discipline that makes Indian startups competitive on unit economics — exemplified by funding rounds like Sarvam AI's reported $350M raise at a $1.5B-plus valuation — increasingly rests on access to subsidised compute, not just on capital.

There is also a managed-infrastructure angle. Subsidised raw GPUs solve the price problem but not the orchestration problem; you still have to schedule, scale and tear down efficiently. That is why serverless-GPU platforms keep raising — see Modal Labs' $355M Series C and the $300M-ARR serverless-GPU moment. The winning pattern for many teams is cheap sovereign capacity for the heavy training run, plus a managed layer for bursty inference.

The honest catches

To keep this guide useful rather than promotional, here are the trade-offs to weigh before you reorganise your roadmap around IndiaAI compute.

1. Approval is a gate, not a formality. The subsidy exists because the work is vetted. If your project is purely commercial with no public-interest framing, you may get base-rate access but not the full 40%.
2. Capacity is contended. A shared national pool means peak-period queues for H100 blocks. Build slack into training timelines.
3. It is a portal, not an API-first cloud. The developer experience is registration and allocation, not instant programmatic provisioning. Treat it as scheduled capacity.
4. The rate is policy, not a contract. Subsidised pricing reflects a current government programme. It is excellent today; do not architect a business that only survives at ₹55/GPU-hour forever.

One more practical point on sequencing. Because the rate is so favourable and the pool is shared, the smart move is to do as much cheap experimentation as possible before you commit to a long approved run. Prototype your training recipe, your data pipeline and your evaluation harness on a handful of GPU-hours — or even on a single card — so that when you finally book an 800-GPU-hour H100 block, you are running a recipe you trust rather than debugging at ₹92 a card-hour. Wasted hours on subsidised compute are still wasted hours, and a contended pool punishes teams that treat allocation as scratch space. Pair this with a clear-eyed view of where self-hosting beats commercial inference once your model is trained; the break-even maths shifts quickly at these rates.

For builders who clear those caveats, the case is straightforward. Sub-₹100/hour approved compute, an average rate near ₹65, and a 40% subsidy on eligible work add up to one of the cheapest credible routes to serious model training anywhere in the world right now. Primary details are on the IndiaAI compute hub, and the capacity milestone is documented in the PIB press release on India's common compute crossing 34,000 GPUs.