H100 rental cost in 2026: what you actually pay

The H100 rental cost you see quoted can mean very different things: a PCIe card billed by the second, one GPU on a shared 8-GPU board, or a whole SXM module reserved for a year. If you are working out how much it costs to rent an H100 in 2026, the headline rate matters less than what it includes and how many hours you will really use. This guide breaks down what moves the price, shows our own numbers, and gives you a formula to compare any monthly price with any hourly quote.

H100 rental cost in 2026 at a glance

We rent the H100 in two configurations, both built on whole SXM GPUs assigned to one customer only. Prices are flat, billed monthly in advance, with no setup fee. The per-GPU-hour column divides the monthly price by the number of GPUs and by 730, the average number of hours in a month.

Two details are worth noticing. First, the 8-GPU server costs $620 a month less than eight single-GPU plans ($17,520), and its GPUs talk to each other over NVLink rather than over the network. If you need eight GPUs, the HGX server is the cheaper way to get them. Second, the per-hour figure is only an equivalent. You pay the monthly price whether the GPU runs 730 hours or 73, so whether it is a good deal depends on your utilization. The break-even section below shows how to check.

What drives the price of an H100 rental

Two quotes for "an H100" can differ widely and both be fair, because they describe different products. Six factors explain most of the gap.

SXM or PCIe

The H100 comes in two main form factors. The SXM5 module sits on an HGX baseboard: per NVIDIA's specifications, it has 80 GB of HBM3 at 3.35 TB/s, a 700 W power limit and NVLink 4 at 900 GB/s per GPU. The PCIe card keeps 80 GB but uses HBM2e at about 2 TB/s, runs within 350 W and has fewer streaming multiprocessors (114 versus 132). NVIDIA also sells the H100 NVL, a 94 GB PCIe variant. LLM token generation is usually limited by memory bandwidth, so a card with about 60% of the SXM bandwidth is a different product and should be priced like one. "H100" alone is not a specification: ask which variant a quote covers.

One GPU or a full HGX board

On an HGX board, all eight GPUs connect through NVSwitch, so tensor-parallel inference and multi-GPU training run over NVLink instead of PCIe. A full board also brings proportionally more CPU, RAM, NVMe and network. In our pricing, a GPU in the 8-GPU server works out to about $2.89 per hour, against $3.00 for the single-GPU plan, which is a dedicated VM with the full GPU passed through.

Billing model

The industry bills GPUs in four common ways: on demand by the second or hour, spot or interruptible capacity that can be reclaimed, reserved terms that commit you for months or years, and flat monthly dedicated servers. Interruptible and long-commitment rates are lower by design, because you carry the risk of losing the GPU mid-run or of paying for capacity you no longer need. Compare like with like: an interruptible GPU-hour and a guaranteed one are not the same purchase.

Term length

Longer commitments buy lower rates, but they lock in your budget. In 2026, GPU capacity is tight, many providers prefer one-year contracts, and on-demand stock is often sold out. Our minimum term is one month, and you can cancel any month with no fee. Once you have an active server, you can ask by ticket for volume and term pricing for 6 to 12 months, or for several servers (usually from three).

Egress and bandwidth

Hyperscalers typically bill outbound traffic per gigabyte, so serving a model, syncing checkpoints or exporting datasets adds a line that the GPU rate hides. Our plans include 20 TB of outbound transfer per month on single-GPU servers and 50 TB on 8-GPU servers, under fair use, with no egress fees within that allowance and inbound traffic not counted. Our help center explains how bandwidth and egress are counted.

Storage

A 70B model in BF16 is about 140 GB before you add datasets or checkpoints, so storage is never a rounding error. On many clouds, boot disks and volumes are billed per GB-month, and they keep billing while the instance is stopped. Our servers include local NVMe: 2 TB on the single H100 and 8× 3.84 TB on the HGX H100. We do not back up local disks unless agreed in writing, so plan an off-server copy of anything you cannot recreate. The local NVMe storage guide has the details.

How to compare a monthly price with an hourly quote

The comparison takes one division:

break-even hours     = monthly price ÷ hourly rate
required utilization = break-even hours ÷ 730

If you will run the GPU for more hours than the break-even in a month, the monthly plan is cheaper. If you will run it for fewer, the hourly plan is. The table applies the formula to a single H100 at $2,190 per month. The hourly rates are placeholders to show the math, not quotes from any provider: plug in your own.

Three adjustments make the comparison honest:

• Use the all-in hourly rate. Add storage, egress, public IP and support charges: all-in rate = GPU rate + (monthly extras ÷ hours used). With $300 of extras over 500 hours, a $3.50 rate becomes $4.10.
• Count the hours you will be billed, not the hours of compute. Setup, weight downloads, debugging and idle time between jobs all run on the meter, unless you tear the instance down every time and rebuild it later.
• Check the effective rate at your real usage. Divide the monthly price by the hours you use. An H100 at $2,190 used for 300 hours costs $7.30 per hour, which is why monthly plans suit steady workloads and hourly plans suit occasional ones.

For an 8-GPU server, compare against eight times the per-GPU rate. At a placeholder $3.50 per GPU-hour, eight GPUs cost $28 an hour, and the break-even against $16,900 per month is 604 hours, or 83% of the month. The GPU cost calculator runs these numbers for you, and our monthly vs hourly GPU pricing comparison goes deeper into the trade-offs.

Are H100 rental prices going down in 2026?

Not in the data we track. OverGPU's October 2026 survey of public price lists, covering 38 sources, found that over the past year the median on-demand price per GPU-hour rose by roughly 13% for the H100, 27% for the H200 and 20% for the B200.

The H100 is no longer NVIDIA's newest data-center GPU, so falling prices would be the intuitive expectation. The survey points the other way, which matches what buyers see in 2026: capacity is tight, many providers prefer one-year contracts, and on-demand stock is often sold out. A low hourly rate only helps if the GPU is available on the day you need it.

Two practical consequences follow. If your workload is steady, a fixed monthly price removes your exposure to on-demand swings. Ours is fixed for each paid month, and any change for future months comes with at least 30 days' notice and the option to cancel before it applies. If you want a price agreed for longer, ask by ticket for term pricing over 6 to 12 months once you have an active server, instead of betting on hourly rates staying where they are.

Hidden costs checklist

Before you sign, run every quote through these questions. Most surprises on a GPU bill come from what surrounds the GPU, not from the GPU itself.

• Egress. Is outbound traffic billed per GB, and how much is included?
• Storage. Are disks and volumes billed separately, and do they keep billing while the instance is stopped?
• Interruptions. Can the capacity be reclaimed? If so, what do checkpointing and restarts cost you in time?
• Idle time. Will the meter run during setup, data loading and the nights between experiments?
• Setup fees and minimums. Is there a setup fee, a minimum spend or a prepayment?
• Networking extras. Are public IPv4 addresses, load balancers or private links charged separately?
• Host resources. Do the CPU, RAM and local disk match the GPU, or will you need a larger instance to get them?
• Quotas and lead time. Do you need a quota approval before you can launch, and how long does it take?
• Support. Can you reach an engineer at any hour, or is that a paid tier?
• Taxes. Is the price before VAT or sales tax?

For reference, our monthly price includes the whole GPU, the listed CPU, RAM and NVMe, a public IPv4 address with an IPv6 /64, DDoS mitigation on public IPs, the monthly transfer allowance, Ubuntu images with NVIDIA drivers, CUDA and Docker preinstalled, and 24/7 support by engineers. It does not include backups of local storage, or taxes such as VAT, which are added where applicable.

H100 vs newer GPUs: cost per GB of memory

For LLM work, memory capacity often decides how many GPUs you need, so price per GB of GPU memory is a useful second lens. The table divides our monthly prices by total GPU memory.

Per GB, the single H100 is the most expensive option in this table. That does not make it a bad buy; it tells you where it fits. The H100 and the H200 have the same vendor-published compute (989 TFLOPS of dense BF16 and 1,979 TFLOPS of dense FP8), so for compute-bound work that fits in 80 GB, the H100 delivers the same math for 29% less. Once memory is the constraint, the H200 gives you 76% more capacity and 43% more bandwidth for 41% more money, while the B300 and the MI355X put far more memory behind each dollar. The MI355X runs on ROCm rather than CUDA, so check your software stack before price per GB makes the decision.

When an H100 is the right buy

The H100 is still the right choice for a lot of real work in 2026, in particular when:

• Your model and its KV cache fit in 80 GB. An 8B model in BF16 takes about 16 GB and a 32B model about 64 GB, leaving room for context. A 70B model fits in FP8 at about 70 GB, but only tightly, with little room for concurrent requests. Our guide to how much VRAM an LLM needs shows the full math.
• Your work is compute-bound. Fine-tuning and training smaller models, embedding jobs and batch processing care more about FLOPS than about capacity, which is where the H100's price per unit of compute holds up.
• You want a mature stack. Hopper is broadly supported by PyTorch, vLLM, TensorRT-LLM and the rest of the CUDA ecosystem, including FP8.
• You need eight GPUs on NVLink for LoRA on 70B models, full fine-tuning of smaller ones or tensor-parallel serving. At about $2.89 per GPU-hour, the HGX H100 is our lowest-priced 8-GPU server with HBM.

Move up when memory is the constraint: the H100 vs H200 comparison covers the most common upgrade, for 70B serving with real concurrency and long contexts. Move to Blackwell when you want a 70B model in BF16 on one GPU, or FP4. If the H100 fits your workload, the NVIDIA H100 page lists both configurations, and the server is delivered in under 10 minutes after payment.

Frequently asked questions

How much does it cost to rent an H100 per month?

We rent a single H100 SXM with 80 GB for $2,190 per month and an 8-GPU HGX H100 server for $16,900 per month, about $3.00 and $2.89 per GPU-hour over a 730-hour month. Prices are flat, billed monthly in advance, with no setup fee.

Is it cheaper to rent an H100 by the hour or by the month?

Divide the monthly price by the hourly rate to get the break-even hours. If you will use the GPU for more hours than that each month, monthly is cheaper; if you only need it for part of the month, hourly usually wins.

What is the difference between an H100 SXM and an H100 PCIe?

Per NVIDIA's specifications, the SXM module has 3.35 TB/s of HBM3 bandwidth, a 700 W power limit and NVLink across the HGX board, while the PCIe card has about 2 TB/s, a 350 W limit and fewer compute units. Ask which variant a quote covers before you compare prices.

Are there extra fees on top of the H100 price?

Our monthly price includes the CPU, RAM, local NVMe, a public IPv4 address with an IPv6 /64, DDoS mitigation and 20 TB of monthly transfer on single-GPU plans or 50 TB on 8-GPU servers, with no setup fee. Taxes such as VAT are added where applicable.

Can I cancel an H100 rental after one month?

Yes. The minimum term is one month and you can cancel any month with no fee; the server keeps running until the end of the paid month.