NVIDIA H200 GPU server in India
One NVIDIA H200 with 141 GB of HBM3e memory, in Mumbai, for ₹2,79,399/month. For serving large models to many users, long context and fine-tuning — with your data in India and a GST invoice in rupees.
- GPU
- 1× NVIDIA H200 141 GB
- vCPU
- 64
- RAM
- 256 GB
- Storage
- 100 GB NVMe
- Since 2006, Jaipur
- 3,500+ businesses served
- Mumbai data centre, Yotta NM1
- GST invoice in INR
Who the H200 is for
For you if: You serve large open models — 70 billion parameters and up — to many users at once, need long context for documents and code, or fine-tune large models; and you want the most GPU memory on one card in India, billed monthly in rupees.
Not for you if
- Your model fits in 24 GB → an L4, from ₹37,499/month
- You render or need graphics → RTX PRO 6000; the H200 is built for AI and scientific computing
- You're pre-training a model from scratch across many GPUs → that needs a cluster; tell us the job and we'll say plainly whether we can help
- You need it for a few days → ask about hourly billing (on request)
H200 plan
| Plan | GPU | vCPU | RAM | Price | Enquire |
|---|---|---|---|---|---|
H200 | 1× NVIDIA H200 141 GB | 64 | 256 GB | ₹2,79,399 a month, ex-GST | Ask |
Price a month in INR, exclusive of 18% GST. Includes 100 GB of NVMe storage, 10 TB of transfer a month and one IPv4 address. Hourly billing: on request. Renews at the same price.
What the plan includes
- One NVIDIA H200 GPU with 141 GB of HBM3e memory at 4.8 TB/s
- The whole GPU is assigned to your server, and no one else's work runs on it.
- 64 vCPU and 256 GB of RAM, for loading and preparing data
- 100 GB of NVMe storage and 10 TB of transfer a month
- One IPv4 address, and root access to the server
- GPU drivers, CUDA and PyTorch installed on request, at no extra charge
Add-ons
| More NVMe storage, for large model files | price on request |
| Additional IPv4 | price on request |
| Drivers, updates and monitoring done for you | quoted |
| Scheduled off-server backups | see cloud backup |
What we do, what you do
| Task | We do | You do |
|---|---|---|
| GPU hardware, network and power; replacing failed parts | ||
| Installing the operating system | ||
| GPU drivers, CUDA and PyTorch | on request | keeping them updated |
| Serving software (vLLM, TensorRT-LLM or your own) | ||
| Security updates and firewall | ||
| Monitoring GPU use and health | ||
| Backups of models and data |
What the H200 is good at
- 141 GB of GPU memory
- The most on one card here: a 70-billion-parameter model at 8-bit with plenty left for long context and many users at once, or models over 100 billion parameters at 4-bit.
- 4.8 TB/s memory bandwidth
- When a large model writes an answer, each new token means reading the model from memory, so memory speed largely sets how fast it answers. The H200's HBM3e is the fastest memory of the three GPUs we offer.
- Fine-tuning
- QLoRA can fine-tune a 70B model within this memory — its authors fine-tuned a 65B model on one 48 GB GPU. Full fine-tuning needs about 18 bytes per parameter, so on one H200 it suits models of a few billion parameters.
- Split into smaller GPUs
- No. The whole GPU is assigned to your server; it isn't divided into smaller GPUs.
- Storage for large models
- Model files are big: a 70B model is about 140 GB at 16-bit, more than the 100 GB included. Plan extra NVMe storage when you order.
What happens after you say yes
- same dayTell us the model, context length and users; we check it fits and that the GPU is in stock
- Usually the same working day, when the GPU is in stockInvoice sent; server set up in Mumbai with root access
- when you're readyYou load your model and go live — or we move your setup from another provider
- 24/7GPU or hardware problem: call our toll-free number in India
In Mumbai, supported from Jaipur
The H200 runs in Mumbai, so the documents, code and customer data your models read stay in India, and answers reach users across India quickly. Support comes from our team in Jaipur, in English and Hindi, in IST, by WhatsApp, phone and email, and the invoice is in rupees with GST.
Moving a model server from another GPU provider, or from AWS or Azure? We set up the H200 server, copy your models and data, and test them before you switch. How migration works
Questions people ask
What is the NVIDIA H200 good for?
NVIDIA builds the H200 for generative AI and scientific computing. In practice: serving large open models to many users, long context for documents and code, and fine-tuning — work that needs more GPU memory than smaller cards have.
What does an H200 cost per hour in India?
The plan is monthly: ₹2,79,399/month works out to about ₹383 an hour if it runs all month (730 hours). Hourly billing for short jobs: on request.
Which models fit on one H200?
A 70-billion-parameter model at 8-bit, with plenty of room for context, and models over 100 billion at 4-bit. A 70B model at 16-bit needs about 140 GB for its weights alone, which leaves no room to run — use 8-bit, or more than one GPU.
H200 or H100?
Both are NVIDIA's Hopper generation. The H200 has 141 GB of HBM3e at 4.8 TB/s; the H100 has 80 GB at 3.35 TB/s, or 94 GB at 3.9 TB/s on the NVL version. We offer the H200 as standard; ask if you need H100s specifically.
H200 or RTX PRO 6000?
The H200 has more memory and much faster memory, for large models and many users. The RTX PRO 6000 costs less than half as much (₹1,35,099/month against ₹2,79,399/month), holds a 70B model at 8-bit in its 96 GB, and also renders.
Is it the SXM or the PCIe version?
Both are available: tell us which you need when you order. Memory and bandwidth are the same on both: 141 GB of HBM3e at 4.8 TB/s.
Do you install CUDA, PyTorch and vLLM?
We install the GPU drivers, CUDA and PyTorch on request, at no extra charge. You have root access, so you run the serving software you prefer: vLLM, TensorRT-LLM, SGLang or your own.
How quickly can I get one?
Usually the same working day, when the GPU is in stock. GPUs sell out, so we confirm the card is available before we send the invoice.
Tell us the model and the users. We'll size it.
Model, context length and how many people use it at once — one recommendation, with the price.
Support 9 am – 7 pm IST. Server-down issues are answered 24/7, in English and Hindi.
Tell us what you run and we'll reply with one recommendation.
No sequence, no sales calls. One reply from an engineer, within working hours.