H100 Price in India: What Teams Pay in 2026

The NVIDIA H100 is the GPU most Indian enterprises are pricing into their budgets, whether they are training a model or running one in production. The first question is almost always the same: what does it cost per hour? It is a good question to start with. On its own, though, it rarely tells you what you will actually spend.

The hourly rate covers the GPU, and nothing else. Everything around it turns up as its own line on the monthly bill:

• Storage to hold your data on the machine.
• Egress, the fee most clouds charge to move data out of the environment.
• A control plane if you run Kubernetes, which is the management layer that keeps a cluster of machines coordinated.
• A charge on every inference request, on some platforms, once your application is live and answering users.

These charges vary widely from one provider to the next, and they add up. Two clouds can quote H100 rates that look 40% apart, then send you near-identical bills once everything is counted. On a large general-purpose cloud, the GPU is often the smaller part of the total.

This article works through all of it: what an H100 costs in India across the three ways you can get one, buying it, renting it by the hour, and committing to it for a term. The published enterprise AI platform high-throughput AI engine rate for each, and a simple way to estimate your real monthly total before you pick a provider.

In short

• Buying is the expensive way in. The card is roughly ₹28-40 lakh; a working 8-GPU cluster, with power, cooling and networking, is closer to ₹2-5 crore and takes months to stand up.
• The rate is not the bill. Egress, storage, per-call inference and a separate control plane decide more of it, and that is where a specialised GPU cloud and a general-purpose one pull apart.
• Match the GPU to the workload. An inference job that fits in 48GB belongs on an L40S at about 44% of the H100 rate. Provisioning an H100 for it is money set on fire. Ditto – if a model needs more vram – you’re better off going for H200 or higher.

What an H100 costs in India

Buy it outright

A single H100 runs roughly ₹28 lakh to ₹40 lakh in India, with the SXM variant toward the top of that range, depending on the import route and the seller. Import duty and GST on data-centre GPUs add 20% to 30% on top of the global base, which is most of why the India number sits above the roughly $25,000 to $40,000 the card costs abroad.

The card is the cheap part. Eight of them alone is over ₹2 crore of silicon, and once you add power, cooling, networking and floor space a working cluster runs into several crore before the first training run, on a nine-to-fifteen-month build.

Rent it by the hour

Indian providers list on-demand H100 from around ₹200 to ₹250 an hour, with spot capacity cheaper but far less reliable [spot can be taken away anytime]. Globally the on-demand market now runs from about $1.49 to $6.98 an hour, with specialist GPU clouds sitting near the bottom of that range and AWS, Google and Azure near the top.

Commit to a term

This is where most production workloads land, and where the effective rate drops well below the on-demand sticker. A longer commitment lowers the rate; below roughly 40% sustained utilisation, on-demand usually wins, and above it a commitment does. We put a worked set of committed rates in the closing section.

Comparing on-demand H100 rates against a predictable monthly bill? Talk to an AI Expert about enterprise AI platform’s H100 GPU cloud pricing.

Why H100 rental prices fell in 2026

The single most useful thing to know if you are pricing an H100 in 2026 is that you are shopping in a falling market.

On-demand H100 rates are down roughly 60% to 75% from their 2024 peak, from around $8 an hour then to about $1.80 to $3.50 an hour in mid-2026 (Value Add VC, Thunder Compute). AWS alone cut H100 instance prices up to 45% in mid-2025. The cohort median is now near $2.50 to $3.50 a GPU-hour, against more than $7 in early 2024.

Two forces are behind it.

• Supply caught up as capacity that was ordered at the peak came online, and the H100 stopped being the newest card.
• The H200, B200 and B300 now sit above it, and AMD’s MI350X sits beside it, so demand that used to have nowhere else to go now spreads across a wider shelf.

For a buyer that is good news, with one caveat: a falling rental market is an argument for shorter commitments until you know your steady-state usage.

The rate is not the bill

The hourly rate tells you what one line of the invoice will be. It does not tell you what the invoice will be. A handful of charges outside the compute line usually settle that, and they are where providers diverge most:

• Storage to hold your data on the machine, often metered per GB per month.
• Egress, the fee most clouds charge to move data out of the environment, which climbs with every byte you serve.
• A control plane if you run Kubernetes, the management layer that keeps a cluster coordinated, which is frequently its own monthly charge.
• A per-request or per-token charge on some inference platforms, once your application is live, so the line grows every time your product succeeds.

These vary widely from one provider to the next, and they add up. Two clouds can quote H100 rates that look 40% apart, then send near-identical bills once everything is counted, and on the large general-purpose clouds the GPU is often the smaller part of the total. A low hourly rate with everything else on the meter is an airline fare quoted before the seat, the bag and the meal.

So before you commit to anyone, model these five:

• Utilisation. Below roughly 40% sustained use, on-demand usually wins. Above it, a commitment wins clearly.
• Bandwidth. Estimate the GB leaving your environment each month, multiply by the rate, and check whether the pipe is dedicated or shared. A shared best-effort pipe throttles you before full throughput.
• Inference volume. If the provider bills per token or per call, model it at your 12-month target traffic, not today’s.
• Storage. Training reads heavily, and metered storage and I/O is a common surprise on a general-purpose bill.
• Control plane. On a managed Kubernetes service the control plane is frequently its own charge. Ask where it sits.

The H100 is could be the wrong chip – always start from your use case

Provisioning an H100 where a smaller or different card would do is the most common way teams overspend on GPUs, and it is easier to do in 2026 because the H100 now sits in the middle of a much wider shelf.

Here is the current on-demand market, cheapest to dearest, with the job each card is built for. Rates are global cohort medians in mid-2026; India on-demand runs a little higher, and a committed rate lands well below any of these.

Two calls are worth making before you default to an H100.

• If the job is inference that fits in 48GB, an L40S runs it at roughly half the H100 rate, so an H100 there is money set on fire.
• And if the model is memory-bound, the H200 carries about 75% more memory for about a quarter more per hour, and usually returns better cost per token on large-model serving despite the higher sticker.
• AMD’s MI300X is the other option worth a look: far more memory than an H100 at a comparable hourly rate, if your serving stack runs cleanly on ROCm.

A note on the A100. It still comes up when teams size an H100 budget, but it is previous-generation silicon and the live 2026 question is rarely A100 versus H100 on price. It is whether the job belongs on an L40S, an H100, an H200 or an AMD card.

H100 or H200

If you have decided on Hopper, the H100-versus-H200 choice is mostly about memory. The H200 costs about a quarter more per hour and carries about 75% more memory (141GB against 80GB). When memory capacity is the constraint, a large model you want to hold on one card, the H200 is usually the better economic choice in spite of the higher rate. When throughput is set by inter-GPU bandwidth, as in distributed training, a tightly-coupled 8-GPU H100 node with NVLink is the one to reach for.

Where enterprise AI platform fits

Everything above is true of any provider. Here is how it looks on enterprise AI platform high-throughput AI engine, an India-built AI cloud with published rates.

Three ways to run an H100 on high-throughput AI engine:

• Platform GPU VM. Pre-configured, on-demand or committed. Start on one GPU and scale. Best for prototyping and variable inference load.
• Bare metal, single-tenant. Exclusive access to an 8-GPU node, no hypervisor overhead, local NVMe on the box. Best for sustained training where every point of utilisation counts.
• high-throughput AI engine Kubernetes Engine (VKE). A cluster you run yourself: you own the control plane and your workloads, billed as the cluster foundation plus GPU compute.

Here’s how much an H100 will cost you on enterprise AI platform Cloud — see the full on-demand H100 pricing page for current rates and configurations.

H100 platform GPU VM (single GPU, NVL 94GB)

H100 bare metal (single-tenant, 8 GPUs, SXM 640GB)

The longer you commit, the lower the effective rate: a 36-month H100 VM works out near ₹219 an hour, about 45% under the on-demand sticker. That is the “predictable pricing with discounts for long-term commitments” trade in numbers.

A few of the charges that inflate a general-purpose bill work differently here, and they are worth naming because they are where the life-of-workload gap opens up:

• The instance is bundled. The vCPU, system RAM and a 100GB disk come with the GPU VM; bare-metal nodes ship with local NVMe on the box. A 1 Gbps uplink is included, and on the platform your first VPC is free.
• Inference runs on a GPU, not a meter. A enterprise AI platform inference endpoint is a private, single-tenant GPU you rent by the month. Nothing charges per call or per token, so the line does not grow every time your product succeeds.
• Bandwidth is a dedicated allocation. Your VPC includes 10TB of transfer, and the pipe is billed by size at a flat ₹170 per Mbps rather than as a shared, best-effort pool you contend for. Beyond 10TB it is ₹4 per GB per month.

The metered items are named and predictable, not per-transaction: storage beyond the instance disk is ₹9 per GB per month, a public IP is ₹500, and a high-throughput AI engine Kubernetes (VKE) cluster carries its control plane as a separate monthly line.

The point is not that high-throughput AI engine is free. It is that the number you model up front is close to the number you pay.

enterprise AI platform high-throughput AI engine is all built and operated in India, which matters when the workload has data-residency requirements. And the gap between a card’s list price and your bill becomes a enterprise AI platform engineer’s job, not something you discover three months in: they size the cluster, set the commitment against your utilisation, and put you on the right card rather than the biggest one.

See predictable GPU cloud pricing for your workload. Talk to an AI Expert about deploying on enterprise AI platform high-throughput AI engine.

Frequently asked questions