What you need to know
- The pool crossed 34,000 GPUs. India's national common-compute capacity rose to roughly 34,333 — 18,417 existing units plus 15,916 newly added — with more than 38,000 onboarded through the IndiaAI compute portal, per PIB and DD News.
- The price is the headline. Empanelled providers won bids at about ₹115.85 a GPU-hour for standard GPUs and around ₹150 for high-end units — roughly 42% below prevailing market rates, according to the IndiaAI portal.
- There is a second discount on top. Projects judged to be of national importance can receive a further subsidy of up to 40%, pulling the effective rate below ₹100 a GPU-hour for approved work.
- The target is ambitious. The stated ambition is to scale to roughly 100,000 GPUs by the end of 2026, funded out of the ₹10,371 crore (about US$1.25bn) IndiaAI Mission outlay.
- It is a fine-tuning and evaluation windfall, not a pre-training one. Cheap GPU-hours change the economics of iteration far more than they change the economics of training a frontier model from scratch.
The numbers, and what holds up
It is worth being precise, because the figures move around in coverage. The capacity milestone is the firmest of them: a government release in mid-2025 stated that capacity had crossed 34,000 GPUs, made up of 18,417 existing units and 15,916 freshly added, delivered by seven empanelled providers including Yotta Data Services, Sify, Netmagic, Locuz, Cyfuture, Ishan Infotech and Vensysco. As the compute portal onboarded more, the total reported climbed past 38,000.
On price, the IndiaAI portal and subsequent reporting put the lowest winning bids at ₹115.85 per GPU-hour for standard-tier GPUs and around ₹150 for high-end units — described as roughly 42% below market. The headline "40% discount" is a separate mechanism: it is an additional subsidy of up to 40%, granted to projects of national importance and signed off by a committee under the Ministry of Electronics and Information Technology (MeitY). Stack the two and an approved high-priority project can land below ₹100 a GPU-hour. The ₹115–150 band is the rate any approved applicant can expect; sub-₹100 is the rate a favoured project can reach.
The 100,000-GPU figure is a target, not a current count — treat it as direction of travel rather than installed base. And the ₹10,371 crore Mission outlay is a five-year envelope, of which Compute Capacity is the single largest line at about ₹4,563.36 crore. Disbursement has been slower than the headline suggests; that is the honest caveat behind the optimism.
To actually access the subsidised pool: register on the IndiaAI compute portal, choose an empanelled provider (Yotta and Sify have had the most capacity online), then submit a workload request specifying GPU type, count and duration. The base ₹115–150 rate applies to most approved requests; the further up-to-40% discount is a separate application that needs MeitY committee sign-off and a national-importance justification, so write that section as if a reviewer who is not a machine-learning specialist will read it.
What cheap GPU-hours change — and what they don't
The instinct on seeing sub-₹150 GPU-hours is "now I can train my own model". For almost every builder, that is the wrong conclusion. Pre-training a serious foundation model still means thousands of GPUs running for weeks; the bill runs to crores even at subsidised rates, and the talent and data pipeline are harder problems than the compute. What genuinely shifts is the cost of everything downstream of a base model.
Fine-tuning a 7B–13B open-weight model with LoRA or QLoRA on a domain corpus is a job of GPU-hours, not GPU-weeks. Running a proper evaluation harness across a dozen candidate configurations — the unglamorous work that separates a demo from a product — is suddenly affordable enough to do properly rather than skipping it. So is building a synthetic-data pipeline, distilling a large model into a smaller one, or simply benchmarking three serving engines on your real traffic shape before you commit. These are the workloads where a 42% price cut, compounded by the national-importance discount, moves a project from "we can't justify it" to "let's run it twice".
A worked cost comparison
Put rough numbers on a representative job: fine-tuning a 13B model with QLoRA, which on a single high-end GPU might take in the order of 40 hours per experiment, run across three experiments. The figures below are illustrative — IndiaAI rates are from the portal and PIB; the hyperscaler on-demand figure is a typical published India-region rate for a comparable single high-end accelerator and will vary by provider and commitment.
| Access path | Approx. ₹/GPU-hour | 3 × 40-hr fine-tune runs (120 GPU-hr) | Notes |
|---|---|---|---|
| IndiaAI — approved national-importance project | < ₹100 | < ₹12,000 | Base rate plus up-to-40% extra subsidy |
| IndiaAI — standard subsidised rate (high-end) | ~₹150 | ~₹18,000 | Any approved applicant, empanelled provider |
| IndiaAI — standard subsidised rate (standard GPU) | ~₹115.85 | ~₹13,900 | Lowest winning bid tier |
| Hyperscaler on-demand (illustrative) | ~₹250–300 | ~₹30,000–36,000 | India region, single high-end GPU, no commitment |
The point is not the exact rupee figure — it is the ratio. On the same job, the subsidised pool runs at roughly half to a third of an unreserved hyperscaler hour, and the national-importance tier roughly a third or less. Over a quarter of real experimentation, that is the difference between running evaluations you can defend and cutting corners to stay inside a budget.
Subsidised price is not the same as instant access. Empanelled capacity is allocated, not infinitely elastic — high-end clusters can be queued, and disbursement against the Mission's outlay has lagged its headlines. Plan your runs with slack, keep checkpoints frequent so a pre-empted job is recoverable, and do not architect a product whose unit economics only work at sub-₹100 GPU-hours you have not yet been granted.
The UK contrast: grants and equity, not a price list
Britain is solving the same sovereignty problem with a different instrument. Rather than publishing a subsidised per-hour rate across many providers, the UK channels startup access through the AI Research Resource (AIRR) — a network of publicly funded supercomputers anchored by Isambard-AI in Bristol and Dawn in Cambridge. Isambard-AI runs 5,448 NVIDIA GH200 Grace Hopper superchips and delivers around 21 AI exaFLOPS, built for roughly £225m; Dawn at Cambridge adds a further large cluster.
On top of that capacity sits the £500m Sovereign AI Fund, launched in April 2026, which pairs equity investment of up to £20m per startup with around one million GPU-hours of compute access and fast-tracked visas. The whole effort nests inside a broader UK Compute Roadmap of more than £1bn for AIRR expansion and a £750m commitment to a new national supercomputer later this decade.
The philosophies diverge in a way builders should notice. India is competing on price at breadth: many providers, a published rate, open to startups, researchers, MSMEs and academia who apply through a portal. The UK is competing on concentration: a smaller number of very large grants of compute and capital, awarded to a curated set of companies, tied to equity. India's model favours the long tail of builders running modest fine-tuning and evaluation jobs; the UK's favours a few ventures chasing frontier-scale work. Neither is obviously better — they are bets on different shapes of ecosystem.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →What a builder should actually do
If you are building in India, the practical playbook is short. First, register on the IndiaAI compute portal now, before you have a workload ready — the account and provider selection take time you do not want to spend mid-sprint. Second, scope your first request as a fine-tuning or evaluation job, not a training run; that is where the price cut compounds and where approval is easiest to justify. Third, if your work plausibly serves a public-interest domain — Indian-language models, healthcare, agriculture, public services — write the national-importance case explicitly and chase the additional subsidy; the gap between ₹150 and sub-₹100 is real money across a quarter of experiments.
If you are building in the UK, the lesson runs the other way. You will not find a published per-hour rate to optimise against; you will find grant rounds and an equity-linked fund. That rewards a different motion — a credible research or commercialisation pitch, a named project, and a willingness to take investment alongside compute. UK builders watching India's price-led model should note what it unlocks at the bottom of the funnel: a far larger number of small teams able to iterate cheaply. The question for British policy is whether concentration or breadth grows more Builders in the end.
For everyone, the cost lever that beats both schemes is not begging for cheaper hardware — it is needing less of it. Tighter evaluation loops, caching, distillation and disciplined model routing routinely cut spend more than any subsidy. We dug into that in our LLM cost-optimisation guide, and the wider context of India's AI build-out in India AI hits $50B.
Sources worth reading first-hand: the DD News compute milestone report, the IndiaAI portal, and the UK government's AI Research Resource pages.