NVIDIA H100 Price in India 2026: Rental, Spot and Monthly Cost
The H100 remains the reference GPU for serious LLM training and inference, but “H100 price in India” can mean three different things: a cloud rental rate, a reserved India-hosted price, or the cost of buying hardware. This page focuses first on the number most teams can act on today: the cost to rent one H100.
The figures below are a dated snapshot from the GPUIndia feed. Open the live H100 rental page for the current provider table because spot capacity and on-demand rates change.
Check today's H100 rate
See the cheapest on-demand and spot prices across tracked providers.
Open live H100 pricesH100 rental price snapshot
| Billing mode | Lowest tracked rate | Highest tracked rate | What it means |
|---|---|---|---|
| Spot | $0.99/hr | $5.00/hr | Lower price, interruption risk |
| On-demand | $2.50/hr | $8.50/hr | Flexible rate, better predictability |
The cheapest tracked on-demand rate is roughly ₹208/hr at ₹83/USD. The cheapest spot rate is about ₹82/hr. Those are not guaranteed India-hosted quotes. A provider may place the VM in another region, charge separately for storage or egress, or have no capacity when you need it.
Monthly H100 cost at different usage levels
Use hours, not the calendar, to estimate a rental budget. The table below uses the $2.50/hr on-demand snapshot and shows how utilization changes the bill.
| Usage | Hours/month | Compute cost | Approx. INR |
|---|---|---|---|
| Light experimentation | 100 | $250 | ₹20,750 |
| Half-time service | 365 | $913 | ₹75,750 |
| Full-time service | 730 | $1,825 | ₹1,51,475 |
| Full-time spot at $0.99 | 730 | $723 | ₹60,009 |
These numbers are compute only. Add persistent disk, model storage, CPU/RAM, networking, egress, taxes and managed-service fees. For a production inference endpoint, also budget for replicas, failover and idle capacity.
Where the H100 is worth its price
70B-class inference
An H100 gives an 80GB memory class with strong Tensor Core performance and a mature CUDA software path. It is a practical baseline for teams serving a 70B-class model at useful concurrency, provided the model and quantization fit.
Fine-tuning and training
For checkpointed training or LoRA work, spot H100 can be attractive if the job can resume after preemption. On-demand is easier to budget when a run has a deadline.
When an H100 is too much
Small models often do not need an H100. For a model that fits in 32GB, an RTX 5090 can be a much cheaper experiment or inference option. For memory-heavy models, H200 can avoid sharding. Match the card to VRAM and throughput rather than buying the most familiar name.
Spot vs on-demand H100
Spot is not simply “cheap H100.” It is a different operating model:
- Use spot for: checkpointed training, evaluation, batch jobs, experiments and flexible workloads.
- Use on-demand for: user-facing inference, client deadlines, migrations and jobs that cannot be restarted.
- Design for interruption: write checkpoints to durable storage, keep environment setup reproducible and monitor reclaim events.
At the snapshot rates, the spot discount can be large, but a failed long run can erase the saving. Read the provider's interruption policy before comparing only the hourly number.
Rent or buy an H100 in India?
Renting wins when usage is uncertain, the project lasts months rather than years, or you need to change GPU generation quickly. Buying has a chance to win at high utilization over a long horizon, but the hardware quote is only the start. Include server platform, networking, cooling, electricity, maintenance, financing and resale risk.
Use the TCO calculator to test your own hours/month and horizon. It separates the advertised rental rate from hidden costs and shows a directional breakeven instead of presenting a universal answer.
IndiaAI reference prices and what they tell you
The public IndiaAI compute price list provides another benchmark with separate on-demand and reserved rates. Its instance configurations are not identical to the marketplace feed, so do not copy one rate into the other table. It is most useful as a reference for how reservation duration and instance size change the effective hourly price.
Bottom line
For September 2026 planning, use $2.50/hr as the lowest tracked on-demand H100 snapshot and $0.99/hr as a spot reference, not a promise. Start with spot if the workload is restartable. Choose on-demand when reliability matters. If the H100 does not fit your model or is underutilized, compare H200, A100 and RTX 5090 by workload cost before committing.
Sources and method: GPUIndia provider feed snapshot dated 14 September 2026; INR examples use ₹83/USD. See the IndiaAI public compute price list for its independent instance and reservation reference. NVIDIA's H100 specifications are the hardware reference. Prices change daily.