This agentic AI tutorial walks through building a tool-using agent on India's subsidised public GPU capacity, where the average rate is approximately Rs 65 per GPU per hour, as stated by MeitY in a Lok Sabha reply. The same reply confirms 38,231 GPUs onboarded from 14 empanelled service providers under the IndiaAI Compute Capacity framework, with the wider IndiaAI Mission carrying an outlay of Rs 10,372 crore. If you are a student, researcher, or early-stage startup in India, that pricing changes what you can afford to experiment with: a four-hour agent build session on one GPU works out to roughly Rs 260 at the stated average rate.
TL;DR
- Average subsidised rate is approximately Rs 65 per GPU per hour, except for select high-end GPUs (PIB).
- 38,231 GPUs from 14 empanelled providers are onboarded; 1,050 TPUs have also been added (PIB).
- Access runs through compute.indiaai.gov.in, where eligible users get up to 40% reduced cost.
- Six eligible user categories, from students to government entities; 190 projects have been approved so far (PIB).
- Sovereign models from Sarvam AI, BharatGen, and Gnani AI are available to developers on AIKosh, so you can build an agent without training a base model.
- Last verified: 28 September 2026.
What is the IndiaAI compute portal, and who is eligible?
The IndiaAI Compute Capacity pillar pools GPU capacity from empanelled private providers and sells access to approved Indian users at a subsidised rate. The allocation and request flow lives at compute.indiaai.gov.in, and the official hub page states that eligible users can access AI compute at up to 40% reduced cost.
Six user categories are listed on the portal: Startups and MSMEs, early-stage startups, researchers or academia, early-stage researchers, students, and government entities. The approved project mix gives a sense of who actually gets through. Of 190 approved projects, 78 went to government entities, 46 to startups and MSMEs, 30 to early-stage startups, 27 to researchers and academia, 5 to students, and 4 to early-stage researchers (PIB).
On the public allocation table you can see empanelled providers including E2E Networks Limited and Yotta Data Services, offering A100 40GB and 80GB, H100 SXM, L40S, and L4 cards. That spread matters for an agent project: an L4 or L40S is usually enough for inference-heavy agent work, while H100 SXM is priced and allocated for training scale.
How do you get an allocation, step by step?
- Register on the compute portal. Create an account at compute.indiaai.gov.in and pick the user category that genuinely matches your status. Category drives the subsidy you are eligible for.
- Describe the project, not the hardware. Applications are assessed per project. State the agent you intend to build, the model you will run, and the expected GPU-hours, rather than asking for the largest card available.
- Choose a provider and GPU type. Select from the empanelled list (E2E Networks, Yotta, and others) and match the card to the workload. For a first agent, an L4 or L40S instance is the sane starting point.
- Read the subsidy line separately from the bill. The allocation table shows "Subsidy Allocated" as its own column against "Bill Of Materials". One published H100 SXM allocation lists Rs 44,656 of subsidy against a Rs 1,11,640 bill of materials (IndiaAI allocation table). The subsidy is not the price; it is the discount against the provider's invoice.
- Start before the deadline. Each allocation row carries a date by which the user must begin consuming the allocation. Unused allocations do not sit idle indefinitely.
What does an agentic AI tutorial stack look like on IndiaAI compute?
The build itself is ordinary, which is the point. Rent a single GPU instance, then assemble three layers.
Model layer. Either call a sovereign model API or serve open weights on the rented card. Models from Sarvam AI, BharatGen (the IIT Bombay consortium), and Gnani AI were launched at the IndiaAI Impact Summit 2026 and are available to developers on the AIKosh platform (PIB). Twelve organisations and consortia are building sovereign foundational models under the Mission, so the catalogue keeps changing. If you would rather self-host, serve an open-weight model on the GPU with a standard inference server.
Orchestration layer. A loop that plans, calls tools, reads results, and decides what to do next. Keep the first version to a single agent with two or three tools; multi-agent graphs add failure modes faster than they add capability. Our walkthrough of building agentic AI systems covers the loop in more detail, and the framework comparison is worth reading before you commit to one.
Tool layer. Two starter agents are worth building in order: a tool-using agent that calls one external API and returns a checked answer, and a data-querying agent that translates a natural-language question into SQL against a small local database. Both are testable, both fail visibly, and neither needs more than one GPU.
Budget the run, not the month. At the stated average of approximately Rs 65 per GPU per hour (PIB), a four-hour session on one card is about Rs 260 (4 x 65) and a 20-hour week of experimentation is about Rs 1,300 (20 x 65) - arithmetic on the quoted rate, not a separately billed figure. High-end GPUs are explicitly excluded from that average, so verify the rate on your own allocation row before planning spend.
Why is agentic AI on the national agenda right now?
Two dated signals. The NASSCOM Agentic AI Confluence 2026 ran on 24 September 2026 at Radisson Blu on Outer Ring Road in Bengaluru, with a developer track that included a crash course on agentic AI and hands-on labs for building tool-using and data-querying agents (NASSCOM). Separately, Union Minister Ashwini Vaishnaw announced on 17 February 2026 that a further 20,000 GPUs would be added on top of the existing 38,000-plus (PIB). Supply is expanding while the skills gap is being addressed in public. If you are planning a learning path, our guide to learning agentic AI in India and the course shortlist are the practical next steps.
What are the honest limitations?
Allocation is not instant, and it is not guaranteed. Approval is per project and per category, and with 190 projects approved to date, the funnel is competitive rather than open-access. The Rs 65 average also hides variance: it excludes select high-end GPUs, and your actual invoice comes from the provider, with the subsidy applied against it.
There is also a design trap. Cheap compute encourages people to reach for a larger model when the real problem is a badly specified agent loop. A weak agent on an H100 is still a weak agent. Our build guide argues for fixing the loop, the tools, and the evaluation harness before spending on hardware.
On topic selection: we priced 656 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only 72 (11.0%) cleared a winnable bar of 150-6,000 monthly searches, difficulty 20 or below, a genuine technical term, and at least three words (n=656, measured 2026-09-28). That is why a specific, buildable guide beats another broad overview.
FAQ
Q: How much does IndiaAI compute cost per GPU hour?
A: The average rate is approximately Rs 65 per GPU per hour, except for select high-end GPUs, according to MeitY's Lok Sabha reply published by PIB. Your final cost is the provider's bill of materials less the subsidy allocated to your project.
Q: Can students access IndiaAI GPU compute?
A: Yes. Students are one of six eligible categories on the compute portal, and 5 of the 190 approved projects were student projects as reported by PIB. Approval is per project, so a clear proposal matters more than the size of the request.
Q: Which GPUs are available through the IndiaAI portal?
A: The public allocation table lists A100 40GB and 80GB, H100 SXM, L40S, and L4 cards from empanelled providers including E2E Networks and Yotta Data Services. For a first agent build, L4 or L40S is usually sufficient.
Q: Do I need to train my own model to build an agent?
A: No. Sovereign models from Sarvam AI, BharatGen, and Gnani AI are available to developers on AIKosh, and open-weight models can be served on a rented GPU. Most agent work is orchestration and tool design rather than model training.
Q: How long does an IndiaAI compute allocation last?
A: Each row on the allocation table carries a date by which the user must start consuming the allocation. Check that deadline when your request is approved, because an unused allocation is not held open indefinitely.
Q: Is the Rs 65 rate cheaper than commercial cloud GPUs?
A: PIB has described the Rs 65 per hour figure as nearly one third of the global average cost. Treat that as the government's own framing and compare it against a live quote from a commercial provider for the specific card you need.
Top comments (0)