Rent NVIDIA H100 in India — ₹1,80,000/month ($2,099)
80GB HBM3 at 3.35 TB/s with FP8 — the 70B workhorse. Benchmark your model before taking it for the month.
H100 pricing
An NVIDIA H100 80GB node costs ₹1,80,000 ($2,099) per month in Mumbai as of July 2026, about ₹247 per hour effective, on a 1-month minimum. Managed ops adds ₹15,000 ($179) per node per month.
| Config | Total VRAM | Per Node / Month | ≈ Effective / hr |
|---|---|---|---|
| 1× H100 80GB | 80GB | ₹1,80,000$2,099 | ≈ ₹247/hr≈ $2.88/hr |
| 2× H100 NVLink | 160GB | On request | — |
| 4× H100 NVLink | 320GB | On request | — |
| 8× H100 NVLink | 640GB | On request | — |
* Prices checked July 2026. Monthly commitment, 1-month minimum; no hourly product. ≈ /hr = monthly ÷ 730, for comparison only. INR prices attract 18% GST, claimable as input tax credit. Managed ops add-on: ₹15,000 ($179) per node/month. Node CPU, RAM, and NVMe sized at scoping; multi-GPU and NVLink pricing confirmed at scoping.
Will your model fit on one H100?
Weight sizes at the stated precision; KV cache needs headroom on top. Unsure — a benchmark run settles it.
| Model | Params | Precision | Fits on 1× H100? | Notes |
|---|---|---|---|---|
| Mistral 7B | 7B | FP16 | Yes | ~14GB weights; full KV headroom |
| Llama 3.1 8B | 8B | FP16 | Yes | ~16GB weights |
| Qwen2.5 32B | 32B | FP16 | Tight | ~64GB weights; limited KV cache |
| Llama 3.1 70B | 70B | INT4 (AWQ/GPTQ) | Yes | ~40GB weights quantized |
| Llama 3.1 70B | 70B | FP16 | No | ~140GB weights; 2× H100 or 1× H200 |
| Mixtral 8x7B | 47B | INT8 | Yes | ~47GB weights |
| DeepSeek V3 | 671B | FP8 | No | Multi-node H200/B200 territory |
H100 — or something else?
Buy vs rent: the H100 breakeven
Buying an H100 card costs about ₹25–30 lakh ($25,000–$30,000) before the server, power, cooling, and import duty. At ₹1,80,000 per month, 14 to 17 months of rental equals the card price alone, with zero capex, a 1-month minimum, and the option to move to newer classes as they ship.
H100 vs A100
The H100 trains transformer models up to 3x faster than the A100 on NVIDIA’s published benchmarks, adds FP8 through the Transformer Engine, and moves 3.35 TB/s of memory vs 2 TB/s. Here it costs ₹1,80,000 vs ₹97,000 per month. For training and high-throughput serving, the H100’s cost per job usually lands lower despite the higher monthly rate. Both cards carry 80GB, so model fit is identical.
H100 vs H200
Same Hopper compute, different memory. The H200 carries 141GB HBM3e vs 80GB (76% more) and 4.8 TB/s vs 3.35 TB/s bandwidth (43% more). Pick the H200 at ₹2,50,000 per month when weights plus KV cache exceed 80GB, such as Llama 3.1 70B at FP16. Otherwise the H100 does the same work for ₹70,000 less each month.
H100 vs AWS p5
AWS p5.48xlarge on-demand works out to about $6.88 per H100 per hour as of July 2026. Our monthly node is roughly ₹247 ($2,099 ÷ 730 ≈ $2.88) per hour effective, with INR billing and a Mumbai datacenter for DPDP workloads.
NVIDIA H100 80GB — chip reference
| Architecture | Hopper |
| Form factor | SXM5 |
| VRAM | 80GB HBM3 |
| Memory bandwidth | 3.35 TB/s |
| CUDA cores | 16,896 |
| Tensor cores | 528 (4th gen) |
| FP32 | 67 TFLOPS |
| FP16 Tensor | 1,979 TFLOPS |
| FP8 Tensor | 3,958 TFLOPS |
| Interconnect | NVLink 4.0, 900 GB/s |
| PCIe | Gen5 x16 |
| MIG | Up to 7 instances |
| TDP | 700W |
H100 hosting questions
How much does an H100 server cost per month in India? +
₹1,80,000 ($2,099) per node per month at our Mumbai location, about ₹247/hr effective, as of July 2026. The rate is a monthly commitment with a 1-month minimum; there is no hourly product. Managed ops (setup, drivers, CUDA, monitoring, <15-min P1 response) adds ₹15,000 ($179) per node per month. 18% GST applies on INR invoices and is claimable as input tax credit.
Should I buy an H100 or rent one? +
An H100 card retails around ₹25–30 lakh before the server, power, cooling, and import duty. Renting at ₹1,80,000 per month means 14 to 17 months of rental equals the card price alone, with zero capex and a 1-month minimum. Rent unless you have multi-year, near-constant utilization and your own datacenter.
H100 vs A100: which should I choose? +
Choose the H100 for training and high-throughput inference: up to 3x faster on transformers, FP8 support, 3.35 TB/s bandwidth. Choose the A100 80GB at ₹97,000 per month when the workload is 70B INT4 inference or mid-size fine-tuning and budget decides. Both are 80GB cards, so the difference is speed per rupee, not model fit.
What models run well on one H100? +
Anything up to roughly 70B quantized: Llama 3.1 70B INT4, Qwen2.5 32B, Mixtral 8x7B INT8, and every 7B–13B model at FP16 with full KV headroom. 70B at FP16 needs 2× H100 or one H200. We confirm fit on a benchmark node before you commit.
How long does H100 provisioning take? +
A single H100 node is ready in 3–5 business days. Multi-GPU NVLink configurations (2×, 4×, 8×) take 5–7 business days. We confirm the exact lead time during the scoping call and keep you updated through provisioning.
Does H100 hosting satisfy DPDP data residency? +
Yes. The node runs at our Mumbai location under Indian jurisdiction. Inference payloads stay on your server; we collect only infrastructure metrics. We sign a Data Processing Agreement confirming no data is used for training or leaves India, which satisfies DPDP Act 2023 localization requirements.
Is there an H100 trial before the monthly commitment? +
For qualified teams, yes — we provision an H100 benchmark node so you can validate your model on the actual hardware first. Request one with your model, expected concurrency, and production timeline, and our engineering team scopes the configuration.
Ready to run on an H100?
Tell us your workload — an engineer replies with a firm monthly quote in one business day. Qualified teams can benchmark on the exact hardware before committing.