Rent NVIDIA H200 NVL GPUs in India

Renting one NVIDIA H200 NVL on AceCloud costs ₹343.32 per hour or ₹2,00,498 per month, for an instance with 16 vCPUs and 128 GB RAM in Noida (as of October 2026, excluding taxes). Two-GPU instances start at ₹740.88 per hour and four-GPU instances at ₹1,481.69 per hour. A 6-month term takes 5% off the monthly rate and a 12-month term takes 10% off. The full breakdown by configuration is on our H200 NVL pricing page.

Pay hourly if your H200 will run for less than about 584 hours in a month. Beyond that, AceCloud’s monthly plan is cheaper. 584 hours is roughly 24 days of continuous use, and it’s the point where ₹343.32 an hour adds up to the ₹2,00,498 monthly price.

In practice, hourly suits proofs of concept, benchmarks and short fine-tuning runs. Monthly suits production inference or anything that stays up around the clock. If you’re unsure, start hourly, check your actual usage after a week, and switch.

Single-GPU H200 instances on AceCloud are available on demand: you can launch one from the console without talking to sales. For 2-GPU and 4-GPU instances, or for capacity guaranteed over a fixed period, we recommend reserving through our sales team so the GPUs are held for you on the date you need them.

AceCloud offers the NVIDIA H200 NVL, the PCIe version of the H200. It has the same 141 GB of HBM3e memory and 4.8 TB/s of memory bandwidth as the H200 SXM.

The two differ in how they’re built into servers:

• SXM modules sit on 8-GPU HGX baseboards, connect through NVSwitch and draw up to 700 W.
• NVL cards fit standard PCIe servers, draw up to 600 W, and link over NVLink bridges in groups of up to four.

For workloads that fit on four GPUs or fewer, the NVL gives you the same memory at a lower cost. If you need 8-GPU HGX nodes, talk to our team about what we can offer.

Rent the H200 if your model, or the KV cache for your context length, doesn’t fit in 80 GB. If it does fit, the H100 is usually the better buy.

The H200 NVL gives you 141 GB of memory and 4.8 TB/s of bandwidth, against 80 GB and 3.35 TB/s on the H100. Its peak tensor throughput is actually a little lower than the H100 SXM’s: 1,671 vs 1,979 TFLOPS at FP16 with sparsity. So training a model that already fits in 80 GB can run as fast, or faster, on an H100.

On AceCloud, a single H200 costs ₹2,00,498 a month and a single H100 costs ₹1,80,000. You pay about 11% more for 76% more memory, which works out to roughly ₹1,422 per GB of GPU memory against ₹2,250 for the H100.

As a rule of thumb, model weights take about 2 bytes per parameter at FP16 or BF16 and 1 byte at FP8. You then need room on top for the KV cache and activations. On AceCloud’s H200 NVL instances, that works out roughly as follows:

Fine-tuning needs much more memory than inference because of gradients and optimizer states. LoRA fine-tuning of a 70B model is a realistic job for 2× H200.

Every H200 you rent on AceCloud is dedicated to you. The GPU is attached directly to your virtual machine through passthrough and isn’t shared with, or sliced between, other customers. The throughput you measure in a benchmark is the throughput you get in production, without noisy neighbours.

Launching an H200 on AceCloud takes four steps:

1. Create an account and complete verification.
2. In the console, go to GPU instances, choose the H200 NVL and the Noida region, and pick a configuration such as N.H200.16.128 for a single GPU.
3. Choose an image. Ready-made images include PyTorch, TensorRT and Triton Inference Server, or you can bring your own.
4. Launch and connect over SSH.

You can do the same through the AceCloud API or your infrastructure-as-code templates.

Yes. New AceCloud customers get ₹20,000 in credits once their account is approved, and the credits work on the H200. At ₹343.32 an hour, that’s about 58 hours on a single H200, enough to benchmark your model or run a short fine-tune before you commit. Credits are valid for 7 days from sign-up.

Yes. AceCloud’s H200 instances run in our Noida and Mumbai data centre in India, and your data stays in the region you deploy in unless you move it yourself. AceCloud is certified to ISO/IEC 27001:2022, ISO/IEC 27017 and ISO/IEC 27018. Each instance runs in its own isolated network, with firewall rules and access controlled by you.

For most teams, renting an H200 is cheaper unless the GPUs will run close to 24/7 for several years. Buying means paying for more than the card: the server, power and cooling, rack space, the people to run it, and supply lead times.

On AceCloud, a 12-month term on a single H200 costs ₹21,65,376 in total (as of October 2026, excluding taxes). That includes the host CPU and RAM, and needs no upfront hardware spend. For a fuller look at H200 purchase prices in India against cloud rental, see our guide.

No. The H200 is the most capable GPU in NVIDIA’s Hopper generation, but the newer Blackwell GPUs, such as the B200, offer more memory and much higher compute. The H200 is still a strong choice for teams that need large memory today on a mature, well-tested software stack.

Use only if MIG is available on AceCloud instances: Yes. The H200 supports Multi-Instance GPU (MIG), which divides one H200 into as many as seven isolated instances, each with its own memory and compute. On AceCloud you can enable MIG on your dedicated H200 to run several smaller inference services side by side.

AceCloud’s H200 instances run the standard NVIDIA Hopper software stack: CUDA, cuDNN, PyTorch, TensorFlow, JAX, TensorRT and Triton Inference Server. Ready-to-use images are available, or you can bring your own containers.