Rent the NVIDIA H200 141GB in India from ₹2,20,000 a card a month with host included, one-month minimum. Ready now.
141GB per card for large models and long context.
A short paid trial, with price and duration agreed before setup.
Businesses we’ve hosted and supported since 2009
Large models · long context · training
₹2,20,000a card a month + taxes
Trial node on H200, then ₹2,20,000 a card a month if you keep it.
Trial on H200 Chat on WhatsAppReselling GPUs to your own customers? Partner pricing on request.
Hosted in Mumbai. Each card keeps its own memory; two cards don’t pool.
Host included: CPU, RAM and storage, on one monthly INR invoice. Your quote lists the configuration and any extras.
| GPU | Memory | Per card a month | Choose it when |
|---|---|---|---|
| NVIDIA H100On request | 80GB | From ₹1,80,000 | Your model and context fit in 80GB with headroom. |
| NVIDIA H200Ready now | 141GB | From ₹2,20,000 | 80GB is the limit: long context, large KV cache, bigger models. |
| NVIDIA B200On request | 180GB | From ₹4,59,000 | You need Blackwell, or more than 141GB per GPU. |
Also ready now: L40S 48GB from ₹55,000 and RTX PRO 6000 96GB from ₹1,10,000. All GPUs on the GPU cloud page.
One monthly price per card. We confirm the variant, card count and interconnect before setup.
| H200 141GB | SXM | NVL |
|---|---|---|
| Memory | 141GB HBM3e | 141GB HBM3e |
| Memory bandwidth | 4.8 TB/s | 4.8 TB/s |
| NVLink | 900 GB/s | Bridge, 900 GB/s per GPU |
| Max power | Up to 700W | Up to 600W |
| MIG | Up to 7 at 18GB | Up to 7 at 16.5GB |
| FP16 Tensor, dense | 989 TFLOPS | 835 TFLOPS |
| Form factor | SXM module on an HGX board | Dual-slot PCIe, air-cooled |
Source: NVIDIA H200 specifications, NVIDIA.
Weights only: parameters times bytes, plus 20% headroom for the runtime. Context and users add KV cache on top.
| Model | FP16 | 8-bit | 4-bit | One H200 runs it at |
|---|---|---|---|---|
| 13B | 26GB | 13GB | 6.5GB | FP16 |
| 34B | 68GB | 34GB | 17GB | FP16 |
| 70B | 140GB | 70GB | 35GB | 8-bit |
| 405B | 810GB | 405GB | 202.5GB | 2 cards at 4-bit |
Two cards don’t pool memory; your software splits the model. Full table: GPU inference sizing.
We rent cards; we don’t sell them. SXM H200 comes on HGX boards of 4 or 8 GPUs, so buying means a whole system.
| Topic | Buy it | Rent it from us |
|---|---|---|
| Up front | A 4- or 8-GPU system, plus duty and GST. | First month, ₹2,20,000 plus taxes. Host included. |
| Running it | Power up to 700W a GPU, rack, cooling and your own staff. | Host and NVIDIA driver included. You run the software. |
| A year | Depends on the quote and how long you keep the system. | ₹26,40,000 for 12 months, before taxes (our maths). |
| If plans change | You resell the system. | One-month minimum. Step down to H100 if 80GB turns out enough. |
More GPU prices: H100 vs A100 vs H200 comparison · H100 80GB rental price · B200 180GB rental price · L40S 48GB rental price · LLM hosting · All GPU cloud prices.
An NVIDIA H200 141GB is ₹2,20,000 per GPU card per month, about ₹2.2 lakh, with a baseline host. Your quote lists its CPU, RAM, storage and network, and the card count. The minimum term is one month.
We rent cards; we don’t sell them. SXM H200 comes on HGX boards of 4 or 8 GPUs, so buying means a whole system plus duty, GST, power and cooling. Renting one card is ₹2,20,000 a month with a host.
Rent H200 when memory is the limit: 141GB of HBM3e at 4.8TB/s against 80GB on H100. If your model and context fit H100 with headroom, H100 saves ₹40,000 a card a month on the same Hopper generation.
At 8-bit, yes: about 70GB of weights, leaving room for cache. At FP16, no: about 140GB of weights alone won’t fit in 141GB with runtime overhead. The sizing table is on GPU inference.
No. It is the price of one H200 card a month. An eight-GPU HGX H200 node with NVLink is quoted as a system.
Yes. Prices exclude taxes, and Indian invoices carry applicable GST, on one monthly rupee invoice. Whether you can claim input credit depends on your tax status and the law.
Need GPUs by the hour? Run them on AWS, managed by us: managed AWS hosting.
