Skip to main content
NVIDIA B200 · Blackwell

Rent NVIDIA B200 in India — ₹3,95,000/month ($4,499)

Blackwell, 192GB HBM3e — frontier-scale training and the largest open models, racked in Mumbai.

Per node, per month, Mumbai DC ₹3,95,000$4,499/mo 1-month minimum · ≈ ₹541$6.16/hr effective · managed ops +₹15,000$179/mo

B200 pricing

An NVIDIA B200 192GB node costs ₹3,95,000 ($4,499) per month in ZenoCloud’s Mumbai datacenter as of July 2026, about ₹541 per hour effective, on a 1-month minimum. Managed ops adds ₹15,000 ($179) per node per month. Multi-GPU NVLink clusters are quoted on scoping.

Config Total VRAM Per Node / Month ≈ Effective / hr
1× B200 192GB 192GB ₹3,95,000$4,499 ≈ ₹541/hr≈ $6.16/hr
2× B200 NVLink 384GB On request
4× B200 NVLink 768GB On request
8× B200 NVLink 1,536GB 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 B200?

Weight sizes at the stated precision; KV cache needs headroom on top. Unsure — the trial node settles it.

Model Params Precision Fits on 1× B200? Notes
Llama 3.1 70B 70B FP16 Yes ~140GB weights; 50GB+ KV headroom
Qwen2.5 72B 72B FP16 Yes ~144GB weights
Mixtral 8x22B 141B FP8 Yes ~141GB weights
Mixtral 8x22B 141B FP16 No ~282GB weights; 2× B200
Llama 3.1 405B 405B FP8 No ~405GB weights; 4× B200
DeepSeek V3 671B FP8 No ~671GB weights; 8× B200 cluster

B200 — or something else?

What is the B200?

The B200 is NVIDIA’s Blackwell datacenter GPU: 192GB of HBM3e, 8 TB/s of memory bandwidth, a dual-die design, and a second-generation Transformer Engine with FP4 support. One B200 holds Llama 3.1 70B at full FP16 precision with room left for long-context KV cache.

B200 vs H200

One generation apart: 192GB vs 141GB (36% more memory), 8 vs 4.8 TB/s (67% more bandwidth), plus FP4 through the second-generation Transformer Engine. Here the B200 costs ₹3,95,000 vs ₹2,50,000 per month. For training runs and FP4 inference the B200 earns the premium; for serving 70B–141B models the H200 remains the better rupee-per-token buy.

B200 vs H100

The B200 carries 2.4x the memory (192GB vs 80GB) and 2.4x the bandwidth (8 vs 3.35 TB/s) of the H100, at ₹3,95,000 vs ₹1,80,000 per month. One B200 replaces a 2× H100 NVLink pair for 70B FP16 serving with a simpler single-GPU deployment and no tensor-parallel overhead.

Buy vs rent at the frontier

Blackwell cards are scarce and priced well above the H100’s ₹25–30 lakh street price, before the server, power, and cooling around them. Renting at ₹3,95,000 per month with a 1-month minimum gets you on Blackwell now, with no capex, no import queue, and a trial node to validate the workload first.

NVIDIA B200 192GB — chip reference

Architecture Blackwell (dual-die)
Form factor SXM
VRAM 192GB HBM3e
Memory bandwidth 8 TB/s
Transistors 208 billion
FP4 Tensor 9 PFLOPS (dense)
FP8 Tensor 4.5 PFLOPS (dense)
FP16/BF16 Tensor 2.25 PFLOPS (dense)
Transformer Engine 2nd generation, FP4/FP8
Interconnect NVLink 5, 1.8 TB/s
TDP Up to 1,000W

B200 hosting questions

How much does a B200 cost per month in India? +

₹3,95,000 ($4,499) per node per month in our Mumbai datacenter, about ₹541/hr effective, as of July 2026. Monthly commitment, 1-month minimum; no hourly product. Managed ops adds ₹15,000 ($179) per node per month. 18% GST applies on INR invoices and is claimable as input tax credit.

B200 vs H200: which should I choose? +

Choose the B200 for training and FP4 inference: 192GB, 8 TB/s, and the second-generation Transformer Engine justify ₹3,95,000 per month when compute throughput decides. Choose the H200 at ₹2,50,000 when you are serving 70B–141B models and cost per token matters more than peak speed.

What models need a B200? +

Single-GPU: Llama 3.1 70B or Qwen2.5 72B at full FP16, and Mixtral 8x22B at FP8. Multi-GPU B200 clusters handle Llama 3.1 405B (4×) and DeepSeek V3 (8×). If your models are 70B quantized or smaller, an H100 or H200 costs less for the same result.

How long does B200 provisioning take? +

A single B200 node is ready in 5–7 business days, subject to Blackwell supply that we confirm before you commit. Multi-GPU NVLink clusters are scheduled during scoping. We share the exact lead time up front and keep you updated through provisioning.

Does B200 hosting satisfy DPDP data residency? +

Yes. The node runs in our Mumbai datacenter under Indian jurisdiction. Training data and 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.

Is there a B200 trial before the monthly commitment? +

Yes. We provision a B200 trial node so you can benchmark training throughput or tokens per second on the actual hardware before the 1-month commitment. Request one with your model, precision, and target batch size.

Benchmark your model on a B200 first.

We provision a trial node, you validate throughput on the exact hardware, then convert to a monthly node.