An NVIDIA H100 80GB costs ₹23 lakh (PCIe) to ₹35 lakh (SXM) to buy in India. Renting an H100 costs ₹255.55 to ₹395 an hour on demand, or ₹1,80,000 to ₹2,05,000 a month on a fixed plan, across six Indian GPU clouds checked on 12 October 2026. ZenoCloud's H100 is ₹1,80,000 a card a month, host included, in Mumbai: the lowest fixed monthly price in that set. For a few hours a month, AWS sells one H100 in Mumbai at $8.256 an hour.
Last verified: 12 October 2026. Figures marked "our maths" are calculated from the sources, not quoted by them. Monthly estimates use 730 hours.
How much does H100 rental cost in India?
ZenoCloud's H100 is ₹1,80,000 a card a month in Mumbai, host included, on one rupee invoice, with a one-month minimum. That is the lowest fixed monthly H100 price among the six Indian GPU clouds we checked; one other provider matches it and the other published fixed plans are higher.
We checked the public price pages and price guides of six Indian GPU clouds on 12 October 2026; the table shows the market range, not the names. Rates are per GPU. Tax is extra where the page says so.
| What you pay | Market low | Market high | ZenoCloud (Mumbai) |
|---|---|---|---|
| On demand, ₹ per GPU-hour (five of six publish one) | ₹255.55 | ₹395 | No hourly plan (₹1,80,000 ÷ 730 ≈ ₹247, our maths) |
| Fixed monthly plan, ₹ per GPU (three of six publish one; one does not state the term) | ₹1,80,000 | ₹2,05,000 | ₹1,80,000, plus taxes; one-month minimum |
| On demand for 730 hours, our maths | ₹1,86,552 | ₹2,88,350 | Not applicable |
| Spot, ₹ per GPU-hour (two of six publish one) | ₹70 | ₹145.80 | None |
| 12-month term, ₹ per GPU a month (two of six publish a rupee figure; a third offers 9% off) | ₹1,59,870 (₹219 an hour × 730, our maths) | ₹1,62,000 (₹19,44,000 ÷ 12, our maths) | Monthly plan only |
| Card and host | 80GB SXM is the usual listing; the ₹395 and ₹2,05,000 quotes are for the 94GB NVL; hosts run 16–26 vCPU and 200–250GB RAM | 80GB, PCIe or SXM stated in the quote; host included |
For a buyer who runs the card all month, the all-in monthly number is what counts: ₹1,80,000 for card and host, against ₹1,86,552 to ₹2,88,350 for the same 730 hours on demand (our maths) and ₹1,80,000 to ₹2,05,000 on the other fixed plans, where ours sits at the bottom.
Longer commitments can go lower: one cloud quotes as low as ₹155.90 an hour on three months or more, and one sells a 36-month plan at ₹1,60,054 a month. Those lock you in for the term; ours needs one month. One provider publishes 720 on-demand hours at ₹1,83,996, which is a sum, not a plan.
Prices on the page beat prices in guides: one provider's April 2026 price guide quoted ₹217.89 an hour, while its live rate in October is ₹330.48. Check the date on any H100 price you read, including this one.
Not in the table: a seventh provider's guide lists ₹249.40 an hour (checked 12 October 2026), and AWS, which sells the H100 by the hour in Mumbai in dollars. AWS gets its own section below, because it is the answer for occasional use.
How much does an H100 cost to buy?
The H100 is sold through NVIDIA's OEM and distributor channels, and quotes move with form factor, volume and the rupee. Confirm stock, configuration and delivery with the seller. These are the figures in published Indian GPU-cloud price guides and reseller listings, checked 12 October 2026.
| What | Price | Source |
|---|---|---|
| H100 80GB PCIe, one card | ₹23–30 lakh | Two Indian price guides: ₹23–30 lakh (undated) and ₹25–30 lakh (updated 19 Apr 2026) |
| H100 80GB SXM5, one card | ₹27–35 lakh | The same two guides: ₹27–34 lakh and ₹27–35 lakh |
| One card, variant not stated | ₹28–40 lakh | A third Indian price guide, updated 3 Sep 2026 |
| One card, live reseller listing | ₹32,56,800 | B2B marketplace listing, PCIe interface, memory not stated |
| HGX H100, 8× SXM system | ₹2.5–3.5 crore | Indian price guide; one Indian reseller lists an 8× system at about ₹3.5 crore (Google India result, 12 Oct 2026; the page blocked our fetch) |
| DGX H100, 8 GPUs | ₹3–4 crore | Indian price guide, updated 19 Apr 2026 |
| One card, global | $25,000–40,000 | IntuitionLabs, updated 14 Sep 2026; its text narrows to $27,000–40,000. NVIDIA publishes no list price for data-centre GPUs. |
On duty: three Indian price guides put import duties and taxes at 18–28%, 20–30% and 28–35%, and none cites a tariff line. The tariff entry for HSN 84718000, other units of automatic data processing machines, shows 0% basic customs duty and 18% IGST (Eximpe tariff finder, as at 13 May 2026).
Whether a given H100 shipment lands under that line is for your customs broker to confirm before you budget.
Need an H100 for a few hours? Use AWS on demand
Not every job needs a card for a month. A weekend fine-tune, a benchmark, a demo for an investor: for those, rent by the hour from AWS. We checked AWS's own pages on 12 October 2026.
AWS's P5 instances carry the H100. The small size, p5.4xlarge, has one H100 80GB, 16 vCPU, 256 GiB RAM and a 3.84TB NVMe drive. The large size, p5.48xlarge, has eight H100s on NVSwitch with 3,200Gbps EFA networking (AWS EC2 P5 instances; EC2 accelerated computing instance types). Both run in Asia Pacific (Mumbai), region ap-south-1 (EC2 instance types by Region).
| AWS instance | H100 cards | Mumbai on demand, Linux, $ per hour | Mumbai Capacity Block, effective $ per hour | Per GPU-hour, on demand |
|---|---|---|---|---|
| p5.4xlarge | 1 | $8.256 | $5.970 | $8.256 |
| p5.48xlarge | 8 | $66.048 | $47.757 | $8.256 (our maths) |
Sources: EC2 On-Demand pricing, Linux, Asia Pacific (Mumbai), and EC2 Capacity Blocks for ML pricing, both checked 12 October 2026. Prices are in dollars. US East (N. Virginia) is cheaper: p5.4xlarge is $6.88 an hour there. With Capacity Blocks for ML you reserve the accelerator capacity you need; AWS prices it as a reservation fee plus an operating-system fee, and the table shows its effective hourly rate.
The crossover, our maths: at ₹96.73 to the dollar (European Central Bank reference rate, 9 October 2026), $8.256 is about ₹799 an hour, so ₹1,80,000 buys about 225 on-demand hours in Mumbai. If your H100 would run under roughly 225 hours a month, AWS on demand costs less on compute alone than a monthly card, ours included. Above that, ZenoCloud's ₹1,80,000 card is the better value, and the gap widens with every hour: a full 730 hours on AWS on demand is about ₹5,83,000 (our maths).
We manage AWS. If you want the hourly H100 without running the account yourself, our Managed AWS desk watches it 24/7, patches what you approve and explains the bill. We can run that for you too.
Google Cloud's A3 High, A3 Mega and A3 Edge machine types also carry the H100; A3 Ultra is the H200 (Google Cloud GPU machine types, checked 12 October 2026). We did not verify a Mumbai price for them.
What moves the H100 price
SXM or PCIe. The SXM module has 80GB HBM3 at 3.35TB/s, up to 700W and 900GB/s NVLink; the NVL card has 94GB at 3.9TB/s (NVIDIA H100 specifications). The PCIe card runs at 2TB/s (our comparison page). SXM quotes run higher. Most rental listings we checked are SXM; ours states which one you get.
Card count and interconnect. Two cards do not pool memory. A model that does not fit in 80GB has to be split across cards by your framework, and multi-node training needs a fast fabric. One cloud's price list bills InfiniBand separately, and one provider's guide puts the networking for an owned multi-GPU setup at ₹50 lakh and up. Of AWS's two P5 sizes, only the 8-card p5.48xlarge has GPUDirect RDMA; the single-card size does not.
Commitment. On demand runs ₹255.55 to ₹395 an hour across the Indian clouds we checked, and about ₹799 on AWS in Mumbai (our maths). Twelve-month terms bring it to ₹219 an hour at the lowest. A fixed monthly card is the middle path: ₹1,80,000 is about ₹247 an hour if the card runs all month, and more per hour if it does not.
Availability. At ZenoCloud on 12 October 2026, H100 is sourced on request and we confirm availability before you commit; H200 141GB is ready now from ₹2,20,000. One Indian cloud lists H200 at ₹379.05 an hour against ₹255.55 for H100 (checked 12 October 2026). If your start date is fixed, ask us to price the H200 alongside the H100.
The host. A card without CPU, RAM, NVMe and network is not a server. Compare quotes on the whole box: the other ₹1,80,000 plan in our sample comes with 26 vCPU and 250GB RAM, the ₹356-an-hour VM with 24 vCPU and 240GB, AWS's p5.4xlarge with 16 vCPU and 256 GiB. Ours lists the host in the quote.
When is renting cheaper than buying?
Assumptions for this section: one H100 SXM at an assumed ₹30 lakh (inside the published range), rented at ₹1,80,000 a month, 730 hours a month. The crossovers are our maths; the cooling, networking and resale figures are quoted.
- The card price alone equals 16.7 months of rent (₹30,00,000 ÷ ₹1,80,000). That is a crossover on the card price, not a payback: the bare card still needs a host, power for up to 700W a card, cooling, rack space and someone to run it.
- For a multi-GPU setup, one Indian price guide puts cooling at ₹15 lakh to ₹1 crore and networking at ₹50 lakh and up.
- Eight cards at our per-card price would be ₹14,40,000 a month (8 × ₹1,80,000; multi-GPU HGX nodes are quoted separately). Against an 8× HGX system listed at ₹3.5 crore, that is about a 24-month crossover before power, networking and staff.
- Resale falls while you own it. A secondary-market price tracker (rates as of 12 October 2026) puts H100 SXM5 cards at $25,000–35,000, down from $40,000+ in early 2025, and forecasts a $30,000 card trading at $10,000–15,000 within 18 months.
Buying starts to make sense when the cards will run near full utilisation for their useful life and you already have a data-centre partner for power, cooling and hands. Otherwise rent.
Hourly or monthly is a utilisation question. At AWS's Mumbai on-demand rate, ₹1,80,000 buys about 225 H100 hours (our maths, at ₹96.73 to the dollar). Under that, pay by the hour on AWS. Over it, a monthly card costs less.
H100 vs H200 vs A100
| GPU | Memory | Bandwidth | FP8 | ZenoCloud, per card a month |
|---|---|---|---|---|
| A100 80GB | 80GB | about 2TB/s | No | from ₹99,000 |
| H100 80GB | 80GB HBM3 (SXM) | 3.35TB/s SXM, 2TB/s PCIe | Yes | ₹1,80,000 |
| H200 141GB | 141GB HBM3e | 4.8TB/s | Yes | from ₹2,20,000 |
Pick the A100 when your stack already runs on Ampere and 80GB is enough; at the starting prices it is ₹81,000 a card a month less (our maths). Pick the H100 for FP8 and the Transformer Engine on training and high-throughput serving.
Pick the H200 when 80GB is the limit: long context, large KV caches, workloads whose measured memory use exceeds one 80GB card. It costs ₹40,000 more than the H100 and is in stock. Specs and workload fits are on H100 vs A100 vs H200.
When ZenoCloud is the right choice
ZenoCloud is the right choice when you will run the card most of the month. You get the lowest fixed monthly H100 price we found, ₹1,80,000, with the host included, in Mumbai, on one rupee invoice, with a one-month minimum. Add GPU Infrastructure Care and we watch the host and runtime 24/7, P1 response under 15 minutes, from ₹25,000 a node a month.
We have run servers since 2009, more than 1,000 of them, on a 99.9% uptime SLA. The quote states PCIe or SXM, card count, CPU, RAM, storage and network, with the NVIDIA driver installed. You run your own software.
Not a fit: light or bursty use, a few hours or days a month, where AWS by the hour costs less (we manage AWS if you would rather not run the account); or an InfiniBand cluster on a fixed date, because H100 is on request.
Take a trial node before committing: a short paid trial on the card you plan to use, terms agreed before setup. If you keep it, the trial node becomes your production node. The whole range is on the GPU cloud page.
How to get an H100 quote
Send us five things and we reply within one business day:
- Model, framework and precision (FP8, BF16, 4-bit).
- Training, fine-tuning or inference, and the target: tokens a second, concurrent users or steps a day.
- Card count and whether you need PCIe or SXM.
- Storage size, dataset location and network needs.
- Start date and term.
Request an H100 quote. The quote states the card, the host, the price before tax and how long it holds.
Frequently Asked Questions
How much does an NVIDIA H100 cost in India in rupees?
₹23 lakh to ₹35 lakh for one 80GB card in published guides, PCIe cheaper than SXM, with some quotes up to ₹40 lakh. An 8-GPU HGX or DGX H100 system is ₹2.5 to 4 crore. Renting costs ₹255.55 to ₹395 an hour or ₹1,80,000 to ₹2,05,000 a month, before tax.
What is the H100 price per hour in India?
₹255.55 to ₹395 per GPU on demand across six Indian GPU clouds, checked 12 October 2026; spot from ₹70, 12-month terms from ₹219. AWS charges $8.256 an hour for one H100 in Mumbai. ZenoCloud sells by the month: ₹1,80,000, about ₹247 an hour, the lowest fixed monthly plan we found.
Can I buy an H100 GPU in India?
Yes, through NVIDIA's OEM and distributor partners and their Indian resellers, and through B2B marketplace listings, such as a ₹32.57 lakh single-card listing we found on 12 October 2026. Confirm stock, memory and form factor with the seller, and budget for IGST and a quote that moves with the rupee.
What does an 8-GPU DGX or HGX H100 cost in India?
₹2.5 to 3.5 crore for an HGX H100 8× SXM system and ₹3 to 4 crore for a DGX H100 in published Indian price guides; one Indian reseller lists an 8× HGX at ₹3.5 crore. Eight cards at ZenoCloud's per-card price would be ₹14,40,000 a month, our maths.
Why is the H100 so expensive?
The SXM card carries 80GB of HBM3 at 3.35TB/s, NVIDIA sets no public list price, and supply has been tight. In India, add freight, IGST, distributor margin and the exchange rate. Renting shares one card across many customers, which is why hourly rates run far below the purchase price.
Is the H100 better than the A100, and worth the price?
For FP8 training and high-throughput serving, yes: the H100 has FP8 Tensor Cores and up to 3.35TB/s of bandwidth against the A100's 2TB/s and no FP8. If your workload fits an A100 and runs at the speed you need, the A100 from ₹99,000 a month is the better buy.
Is the ZenoCloud H100 price per card or per server?
Per GPU card per month, with a baseline host included. The quote lists the card count, PCIe or SXM, CPU, RAM, storage and network, so you compare a complete machine, not a bare card. Two cards do not pool memory; your framework must split the work.
Is GST included in H100 rental prices in India?
ZenoCloud quotes before tax and adds GST on the invoice; at 18%, ₹1,80,000 a month bills at ₹2,12,400 (our maths). Other providers' pages do not all state tax treatment, so ask before you compare. Imported hardware under HSN 84718000 carries 0% basic customs duty and 18% IGST.
Sources
Checked 12 October 2026 unless dated.
- ZenoCloud H100 GPU cloud, GPU cloud pricing and Managed AWS
- Indian GPU-cloud prices: public price pages and price guides of six providers, plus a seventh provider's guide, two B2B reseller listings and one secondary-market price tracker, checked 12 October 2026 (list available on request)
- AWS, Amazon EC2 P5 instances
- AWS, EC2 accelerated computing instance types and EC2 instance types by Region
- AWS, EC2 On-Demand pricing, Linux, Asia Pacific (Mumbai) and US East (N. Virginia)
- AWS, EC2 Capacity Blocks for ML pricing
- Google Cloud, GPU machine types
- European Central Bank euro reference rates, 9 October 2026: USD/INR cross-rate ₹96.73 (via frankfurter.dev)
- IntuitionLabs, NVIDIA AI GPU prices, updated 14 September 2026
- NVIDIA H100 Tensor Core GPU specifications
- Eximpe tariff finder, HSN 84718000, rates as at 13 May 2026