IndiaAI GPUs at ₹67 an Hour: Can a Student or Small Startup Really Get One? (2026 Guide)
Short answer: yes, a student or small startup can apply, but it is not as simple as the headline suggests. The famous ₹67 an hour price is for a single NVIDIA L40S, a mid-range GPU, not an H100. On the official rate list, H100s are sold only in 4-GPU and 8-GPU blocks, which works out to about ₹153 per GPU-hour or more, and a single H200 is listed at ₹140 an hour. You need to log in with DigiLocker or a similar government ID, fit an eligibility category, apply between the 1st and 25th of the month and wait for results on the 10th of the next month.
When the IndiaAI Mission announced subsidised GPUs, the number that went viral was "GPUs at ₹67 an hour". It made Indian AI compute sound cheaper than anywhere in the world. Many students and indie builders have since searched for how to actually get one.
This guide uses the official IndiaAI compute portal, its published rate list (the Ready Reckoner) and current provider prices to answer the practical questions: who qualifies, how the process works, what each GPU really costs and when renting elsewhere or using an API is the smarter move.
What the IndiaAI compute portal is
The IndiaAI Mission has an outlay of more than ₹10,000 crore, and nearly half of it goes to compute. The government did not build one giant data centre. Instead, it ran a tender, empanelled private cloud providers at negotiated prices and put them behind a single portal, compute.indiaai.gov.in.
According to Business Standard in June 2026, more than 18,000 GPUs were empanelled in the early phases against an original target of 10,000, and more than 38,000 GPUs are now available through the wider ecosystem at subsidised rates.
The empanelled providers listed on the portal are CMS, Cyfuture, E2E Cloud, Ishan, JPL, Locuz, NxtGen, NTT, Neysa, Orient, Sify, Tata Communications, Vensysco and Yotta. You apply on the government portal, but the GPU itself runs in one of these providers' data centres, and you pay that provider for anything the subsidy does not cover.
The official rate list: what each GPU costs
The portal's Ready Reckoner lists on-demand prices per hour. Most prices are as of 13 March 2026. We have added a per-GPU column so the numbers are comparable.
| GPU instance | Price per hour | Per GPU per hour | Memory per GPU |
|---|---|---|---|
| NVIDIA L4 (1 GPU) | ₹44.86 | ₹44.86 | 24 GB |
| Intel Gaudi2 (1 GPU) | ₹57.60 | ₹57.60 | 96 GB |
| NVIDIA L40S (1 GPU) | ₹67.50 | ₹67.50 | 48 GB |
| NVIDIA A100 80GB (1 GPU) | ₹135.90 | ₹135.90 | 80 GB |
| NVIDIA H200 SXM (1 GPU) | ₹140.00 | ₹140.00 | 141 GB |
| Intel Gaudi3 (1 GPU) | ₹153.00 | ₹153.00 | 128 GB |
| NVIDIA H100 SXM (8 GPUs) | ₹1,224.00 | ₹153.00 | 80 GB |
| AMD MI300X (1 GPU) | ₹168.22 | ₹168.22 | 192 GB |
| NVIDIA H100 NVL (4 GPUs) | ₹674.96 | ₹168.74 | 94 GB |
| NVIDIA H100 PCIe (8 GPUs) | ₹2,008.00 | ₹251.00 | 80 GB |
| NVIDIA B200 SXM (1 GPU) | ₹323.00 | ₹323.00 | 180 GB |
Three things stand out:
- The ₹67 GPU is the L40S. With 48 GB of memory, it is good for inference, LoRA fine-tuning of 7B to 13B models and image generation. It is not a training GPU for large models.
- There is no single H100 on the list. The smallest H100 option is a 4-GPU H100 NVL block at ₹674.96 an hour. If you need just one big GPU, the single H200 at ₹140 an hour has more memory than an H100 and costs less per hour than any H100 option.
- These are list prices, before any subsidy. IndiaAI can subsidise part of an approved project's bill. The portal itself does not publish a percentage. Business Standard quotes a voucher system that typically covers 40% to 60% of compute costs, and other reports cite up to 40%. Treat the subsidy as possible, not guaranteed.
The memory figures in the last column are the standard specifications of each GPU, not from the portal.
How it compares with renting directly
E2E Networks is both an empanelled IndiaAI provider and a public GPU cloud, so it makes a clean comparison. Here are its own on-demand prices, excluding 18% GST:
| GPU | IndiaAI portal (per GPU-hour) | E2E on-demand | E2E spot | Portal saving vs on-demand |
|---|---|---|---|---|
| L4 | ₹44.86 | ₹49.00 | Not offered | About 8% |
| L40S | ₹67.50 | ₹102.00 | Not offered | About 34% |
| A100 80GB | ₹135.90 | ₹189.00 | ₹66.00 | About 28% |
| H100 | ₹153.00 (in 8-GPU blocks) | ₹379.05 | ₹70.00 | About 60% |
| H200 | ₹140.00 | ₹502.55 | ₹88.00 | About 72% |
| B200 | ₹323.00 | ₹759.05 | Not offered | About 57% |
For global context, a July 2026 comparison put H100s on AWS, Azure and Google Cloud at roughly $10 to $12 per GPU-hour in 8-GPU instances. At about ₹95 to the dollar, that is roughly ₹950 to ₹1,140. Even before any subsidy, the IndiaAI rate for an H100 is a fraction of that.
Two honest caveats:
- Spot prices can beat the portal. E2E's spot H100 at ₹70 an hour is cheaper than the portal's ₹153, and you can start in minutes without an application. Spot machines can be taken away when demand rises, so they suit jobs that save checkpoints often.
- The real saving is on big, steady jobs. For an H200 or a block of H100s running for weeks, the portal is dramatically cheaper. For a few hours of experiments on a small GPU, the saving may not be worth a month's wait.
Who can apply
The portal lists seven categories of users. These are the official criteria:
| Category | What you need |
|---|---|
| Students | AI/ML projects and an aligned academic profile |
| IndiaAI Fellowship students | Selection under the IndiaAI Fellowship |
| Researchers and academia | An h-index of 5 or more, ML/AI publications in journals with impact factor above 2.5, or 150+ citations |
| Early-stage researchers | An AI/ML researcher profile and experience in AI/ML or allied fields |
| Startups and MSMEs | DPIIT recognition, AI/ML experience, and annual revenue of ₹50 lakh or funding of ₹1 crore |
| Early-stage startups | 4 years of AI/ML experience, including research, and a letter from the IndiaAI CEO or an officer of Joint Secretary rank or above |
| Government entities | A letter from a Director, MD, GM or authorised officer |
For students, recent coverage by CareerIndia says the portal asks for an APAAR ID, the government's academic ID for students. Startups need DPIIT recognition, and MSMEs need Udyam registration.
The honest reading: a student with a real AI project has a defined route. A brand-new startup with no revenue or funding has the hardest path, because it needs either the revenue or funding threshold or a senior officer's letter.
How to apply, step by step
- Log in with a government ID. The portal uses DigiLocker, e-Pramaan or Jan Parichay (Meri Pehchaan) for sign-in.
- Register as an end user. Fill the profile form for your category and upload the documents it asks for, such as APAAR, DPIIT, Udyam or fellowship proof.
- Wait for verification. Your profile is checked before you can request compute.
- Build your bill. Use the portal's price calculator to pick GPUs and hours. This becomes a draft bill of materials, including any subsidy you request.
- Submit between the 1st and 25th of the month. Requests are batched monthly.
- Check results on the 10th of the next month. The approved list is published on the 10th, or the next working day.
- Start within 30 days. Approvals are valid for 30 calendar days, so have your code and data ready before you apply.
Requests under 5,000 GPU-hours are auto-approved after profile checks. Larger requests, and according to CareerIndia any request that asks for a subsidy, go to a Programme Monitoring and Evaluation Committee for review. In practice that means: ask for a sensible number of hours, and expect a subsidy request to take longer.
If you submit on the 25th, the earliest you can know is about two weeks later. If you submit on the 2nd, it is closer to six weeks. Plan around that.
What a student or small startup can actually do with it
Here is what real jobs would cost at portal list prices, before any subsidy:
| Job | GPU | Hours | Cost at list price | If a 40% subsidy is approved |
|---|---|---|---|---|
| LoRA fine-tune of an 8B model for a final-year project | L40S | 10 | ₹675 | ₹405 |
| Fine-tune a 30B-class model | H200 | 24 | ₹3,360 | ₹2,016 |
| Run an open model for a demo week, always on | L40S | 168 | ₹11,340 | ₹6,804 |
| Serve a model all month, always on | L40S | 720 | ₹48,600 | ₹29,160 |
The 40% column is illustrative. Whether you get a subsidy, and how much, depends on approval. GST and any provider charges for storage or data transfer are extra.
The first two rows are where the portal shines. The last row is where it can mislead you, because a GPU bills every hour it is on, even when nobody is using your app.
When an API is cheaper than any GPU
Business Standard reports that more than 60% of startups requesting IndiaAI GPUs want them for inference, meaning running a model for users rather than training one. For many of them, a GPU is the wrong tool.
An always-on L40S costs ₹48,600 a month at portal list price. That is a fixed cost whether you serve ten users or ten thousand. A pay-per-token API costs nothing when nobody uses your app and scales only with real traffic. For an early product with uneven usage, the API almost always wins until traffic is steady and high.
A simple rule of thumb:
- Use the IndiaAI portal for training, fine-tuning and research runs with a clear start and end, especially on H200s and H100 blocks.
- Use spot instances for quick experiments that can restart from a checkpoint.
- Use an API for serving users in an early-stage product, and switch to dedicated GPUs only when traffic justifies it.
Velona gives you GPT, Claude, Gemini, DeepSeek, Qwen and other models through one OpenAI-compatible API, paid in rupees with UPI, so you can launch on an API and move to your own fine-tuned model later without rewriting your app. Our local versus cloud AI guide covers the self-hosting side in more detail.
The catches nobody mentions
- Short leases on fine-tuning. An industry source told Business Standard that a seven-day lease cap on fine-tuning runs is an operational risk. Save checkpoints often.
- Uneven hardware. Many empanelled GPUs are better suited to inference than frontier training, according to the same report.
- Price uncertainty. Short contract cycles make it hard to plan long projects around today's rate.
- First come, first served. One expert quoted by Business Standard warned this can favour institutions already good at writing proposals.
- The 30-day window. If your data or code is not ready when approval arrives, you can lose the slot.
Frequently asked questions
Can students get GPUs from the IndiaAI compute portal?
Yes. Students are one of the portal's official user categories. You need an AI/ML project and an aligned academic profile, and recent reports say the portal asks for an APAAR ID. You log in with DigiLocker or a similar government ID.
Is the IndiaAI GPU really ₹67 per hour?
For one GPU type, yes. The NVIDIA L40S is listed at ₹67.50 an hour. Bigger GPUs cost more: a single H200 is ₹140 an hour and H100s work out to about ₹153 per GPU-hour or more, sold in 4-GPU or 8-GPU blocks.
How much subsidy does IndiaAI give?
The portal does not publish a fixed percentage. Business Standard reports a voucher system that typically covers 40% to 60% of compute costs for approved projects, and other reports cite up to 40%. Approval is not guaranteed.
How long does IndiaAI GPU approval take?
Requests are accepted from the 1st to the 25th of each month, and the approved list is published on the 10th of the following month. Approvals are valid for 30 days, so you must start within that window.
Can a startup without DPIIT recognition apply?
The standard startup route needs DPIIT recognition plus revenue of ₹50 lakh or funding of ₹1 crore. The early-stage startup route instead needs 4 years of AI/ML experience and a letter from the IndiaAI CEO or a senior officer.
Which IndiaAI GPU should I pick for fine-tuning?
For 7B to 13B models with LoRA, the L40S at ₹67.50 an hour is usually enough. For larger models or full fine-tuning, the single H200 at ₹140 an hour offers 141 GB of memory and is the best value on the list.
Is it cheaper than AWS?
Much cheaper. A July 2026 comparison put H100s on AWS, Azure and Google Cloud at roughly $10 to $12 per GPU-hour, around ₹950 to ₹1,140 at about ₹95 to the dollar, against about ₹153 on the portal before any subsidy.
Sources
- IndiaAI compute portal: eligibility, process and empanelled providers (IndiaAI, last updated 2 November 2026)
- IndiaAI compute Ready Reckoner: rate list (IndiaAI, prices as of 13 March 2026)
- Availability of 18,000+ affordable AI compute units (IndiaAI)
- India's AI compute race and subsidised GPUs (Business Standard, June 2026)
- GPUs will be available at Rs 67/hr on IndiaAI compute portal (ETV Bharat)
- IndiaAI compute GPU slots 2026 (CareerIndia, August 2026)
- E2E Networks GPU pricing (E2E Networks)
- GPU cloud pricing in India 2026 (DEV Community, July 2026)