Renting GPUs in India has become much easier, but prices are still hard to compare. Some providers publish hourly rates, others only “contact sales”, and the headline price often leaves out networking, storage or support. Here is how to read GPU pricing and what a fair price looks like in 2026.
The three ways you pay
- On-demand: pay by the hour, stop any time. Best for experiments and short jobs.
- Spot (or preemptible): spare capacity at a large discount, often 40–60% off, but the server can be taken back at short notice. Good for jobs that can restart from a checkpoint.
- Reserved: commit for 1, 6 or 12 months and pay less per hour. Best for production inference and long training runs.
What the market charges
We surveyed the public price lists of international and Indian GPU clouds in October 2026. On-demand prices per GPU per hour, for providers that publish them, fell roughly in these ranges:
| GPU | International providers | Indian providers | zaryacloud |
|---|---|---|---|
| H100 | about $3 to $6 | about ₹350 to ₹490 | |
| H200 | about $4 to $6.50 | about ₹400 to ₹530 | |
| B200 | about $6 to $9 | about ₹470 to ₹760 |
Ranges exclude the largest hyperscalers (AWS, Azure and Google Cloud), which list these GPUs well above the top of the range. Market prices change often; check each provider’s current page before deciding.
Hidden costs to watch for
- Cluster networking: some clouds charge extra per GPU for InfiniBand, which you need for multi-node training.
- Support fees: a percentage of your monthly bill for anything beyond email support.
- Storage and data transfer: fast shared storage for training data, and fees for moving data out.
- Minimum terms: “from” prices that only apply to 3-year or very large commitments.
- Currency: dollar billing means your rupee cost moves with the exchange rate.
At zaryacloud, every price is public and shown in both dollars and rupees, and the price list spells out what is included and what is billed separately.
How to compare offers fairly
- Compare per GPU per hour, not per server. An 8-GPU server at $26 per hour is $3.25 per GPU.
- Compare the same GPU version. An H100 PCIe card is slower than an H100 SXM in an HGX server.
- Ask what is included: vCPU, RAM, local NVMe, networking and support.
- Price the whole job, not the hour. A faster GPU at a higher rate can finish sooner and cost less.
- Check where your data will live if you have data-residency rules.
Ways to pay less
- Use spot instances for experiments and checkpointed training.
- Reserve capacity for steady production workloads: 1, 6 or 12 months can save up to 28% with us.
- Right-size: many inference jobs run perfectly on L40S or RTX PRO 6000 instead of H100s.
- Quantise models to 8-bit or 4-bit to fit them on fewer GPUs.