Renting an H100 by the Hour in Europe: What It Actually Costs

What an on-demand H100 hour costs in Europe today

An hourly H100 price is meaningful only with its GPU variant, region, billing rules and availability. The rate card below uses selected provider pages checked on 1 October 2026. It is not a European market median or a promise that a GPU is available to rent.

H100 listings do not all represent the same configuration. NVIDIA lists H100 SXM with 80GB memory and up to 900GB/s NVLink, while H100 NVL uses 94GB per GPU and a different interconnect specification. Ask for the exact form factor, usable memory and topology, especially for multi-GPU jobs.

Read the provider’s current rate, currency and region-specific offer before booking. Global pricing pages establish list prices, but do not by themselves establish European stock. The table keeps published USD prices without converting currencies.

• Unit: distinguish per-GPU and per-node prices
• Billing: confirm rounding, minimum charges and commitment length
• Extras: check storage, egress and public IP charges
• Availability: confirm the chosen European region and GPU count

When hourly beats a reservation

On-demand billing can suit an uncertain pilot or a short capacity burst. It generally follows allocated instance time, including idle time until termination. A reservation can reduce the hourly rate but may charge the entire committed period.

Consider an illustrative 30-day commitment, not a provider quote. At $2.50 per on-demand hour, 200 hours cost $500. A hypothetical $1.75 reserved rate for all 720 hours costs $1,260. The break-even is 504 on-demand hours, or 70% utilisation. Real reservations may require longer terms or different minimums.

The cost that kills hourly economics in practice is not the rate, it is the instance nobody tore down. Per-second billing and a teardown habit are what keep an on-demand bill tracking the job rather than the wall clock. Reservation economics invert that: idle hours are prepaid, so the utilization assumption has to hold before the discount is real.

One boundary, stated plainly: reservation terms and minimum commitments are covered separately in the piece on one-month-minimum H100 reservations. This section only decides whether the hour is the right unit to buy. For a time-boxed pilot or a burst window, it usually is. For a workload that runs hot every day for months, it usually is not.

• On-demand suits uncertain duration and short bursts when instances are terminated promptly
• A reservation needs sufficient utilisation across the full commitment
• Infrequent jobs may suit on-demand, interruptible capacity or a managed per-job service, depending on restart cost

Provider hourly rates, with the date they were read

Lyceum publishes this article and sells GPU capacity. Prices below were checked on 1 October 2026 in the provider’s displayed USD currency. Taxes and extra services depend on the offer. Confirm region and availability before ordering.

These selected rates are not a cheapest-provider ranking. Marketplace offers and other clouds may have lower prices, but a valid comparison also needs the same region, GPU variant, support terms and job requirements.

Compare the cost of completing your workload. A faster interconnect, larger memory allocation or more suitable CPU-to-GPU ratio can offset a higher hourly rate. Benchmark the offered configuration rather than assuming equal performance from the H100 name.

Lyceum lists its current H100 on-demand rate at lyceum.technology/pricing. When checked on 1 October 2026, its launch screen showed no available H100 capacity. A published rate is not a reservation or a promise of immediate access; obtain a capacity quote when timing matters.

OVHcloud’s USD rate card lists one-, 2- and 4-GPU H100 instances. Dividing each instance rate by its GPU count gives a comparison unit, but the resulting figure still includes the instance’s bundled resources.

The 4-GPU example is $11.97 divided by 4, or $2.9925 per GPU-hour before rounding. Keep the unrounded instance price in your budget. Dividing establishes a useful comparison unit, not a standalone price for a GPU stripped of CPU, RAM and storage.

• Divide a node price by its GPU count before comparing it with a per-GPU rate
• Keep the bundled CPU, RAM and storage visible in the comparison
• Distinguish on-demand and interruptible offers, and check termination conditions for both

Reading a quote: product mode, minimums and granularity

The headline rate is the least informative number on a quote. Four checks beside the number decide what the job actually bills.

1. Product mode: a VM, managed endpoint and training job are different offers
2. Billing: Lyceum documents per-second VM billing. Confirm minimums and rounding for every provider
3. Tax and extras: use the quote applicable to your billing country and add storage, network and IP charges
4. GPU partitioning: confirm whether the service actually offers a suitable partition; hardware support alone does not make it available

Billing granularity matters most for short jobs. With hypothetical rates, a 10-minute job costs about $0.48 at $2.89 per hour billed per second, but $2.50 with a one-hour minimum. Compare the actual rounding and minimum terms, not just the rate.

What the hourly rate leaves out

The hourly figure covers the listed instance bundle. Additional services and time spent idle can change the total.

• Network: Hyperstack publishes free egress, but check the terms and any other network charges
• Storage and IPs: add separately priced volumes, retained checkpoints and public IPs where applicable
• Idle allocation: one H100 at $2.50 per hour costs $420 over 168 hours, even without useful work
• Operations: include setup, monitoring and the engineering time needed to run the job

Run the numbers as total cost of compute, not hourly rate: compute plus egress plus storage plus idle time plus the engineering overhead of running on someone else's hardware. The full model, including the owned-hardware case and its break-even arithmetic, is in Total Cost of Ownership for a GPU Cluster in 2026.

Choosing between on-demand, reserved and cluster

The decision reduces to the shape of the workload, and each shape has a unit that prices it honestly.

• Short or uncertain work: compare on-demand offers and terminate promptly
• Sustained use: compare the full reservation bill with expected on-demand hours
• Distributed training: request a cluster quote with topology and capacity stated together
• Owned hardware: include financing, facilities, support and utilisation before comparing it with rental

When the workload outgrows 80 GB of VRAM, the same decision repeats one generation up, and the memory-bandwidth and precision trade-offs are laid out in NVIDIA B200 vs H200 GPU for Inference.

Lyceum’s on-demand GPU VMs provide raw GPU access over SSH, with supported GPU counts depending on the hardware profile and available capacity. Billing is per second with no minimum commitment. Confirm the placement and stock for your requested configuration, and terminate the VM when work ends.

• Use dated prices for exact configurations, then confirm European stock
• Model a reservation over its complete commitment period
• Include bundled resources, extra services and idle time in total cost

Check the current rate for the GPU and product mode you actually need.