GPU Prices in Europe: What an H100 Hour Costs

A single NVIDIA H100 costs between €2.80 ($3.20) and €7.79 ($8.89) per GPU hour in Europe, depending on which GPU cloud you rent it from. OVHcloud, Scaleway and Verda are all under €3. AWS and Azure charge more than double in their EU regions, and AWS does not list the H100 as an EC2 instance in Frankfurt.

Every price in this article comes straight from the providers' own price APIs and price lists. AWS and Nebius only publish US dollar prices, which we convert to euros at the ECB reference rate. We refresh the tables every month.

We compare the on-demand list price per GPU hour, excluding VAT: one hour with no contract. Reservations and commitments are structured differently by provider and term, so they are left out. Disclosure: Kontinent itself offers dedicated inference on GPUs in the EU.

Key takeaways

• H100 pricing from €2.80 ($3.20) per hour: OVHcloud, Scaleway and Verda all stay under €3.

• Hyperscalers charge more than double: Azure asks €7.79 ($8.89) for an H100 NVL in Frankfurt, AWS €6.45 ($7.36) in Stockholm.

• AWS lists no H100 in Frankfurt: eu-central-1 carries no P-series EC2 instance newer than the A100. The nearest AWS H100 in the EU is in Stockholm.

• Azure has no 8-GPU H100 nodes in Germany: in the EU, the ND H100 v5 series is only listed in the Netherlands, Poland, Sweden and Spain.

• Nebius raises HGX prices on October 1, 2026: the H100 goes from $3.85 to $4.50 per hour, up 17%.

• Managed costs extra: Scaleway's Managed Inference on the same H100 costs €3.40 ($3.88) instead of €2.87 ($3.27), about 19% more.

Table of contents

1. H100 pricing per GPU hour

2. H200, B200 and B300 pricing

3. A100, L40S and L4 for smaller models

4. Where the GPU runs: regions and headquarters

5. Nebius raises HGX prices on October 1, 2026

6. What a GPU server costs per month

7. Rent, reserve or run dedicated GPU servers

8. Methodology

9. Work out your GPU setup with us

10. Frequently asked questions

11. Sources

H100 pricing per GPU hour

Every provider in this table rents out the H100 as a cloud instance, meaning a GPU server billed by the hour. It comes as an 80 GB PCIe card, as an SXM module with NVLink across every GPU in a node, and as the H100 NVL with 94 GB, which is also a PCIe card. NVIDIA lists 3.35 TB/s of memory bandwidth at up to 700 W for the SXM H100, 3.9 TB/s at 350 to 400 W for the H100 NVL and 2 TB/s at 350 W for the 80 GB H100 PCIe. So the NVL has more memory bandwidth than SXM, but less FP8 compute (3,341 vs 3,958 TFLOPS with sparsity), and NVLink only bridges it in pairs. The cheapest SXM H100 in this comparison is Verda at €2.95 ($3.42).

If you want an SXM H100 at Azure, you have to take a node with eight GPUs and pay for all eight at €13.72 ($15.65) each. At Scaleway, the eight-GPU SXM node is the cheapest per GPU; smaller instances cost slightly more.

H200, B200 and B300 pricing

The H200 has 141 GB of HBM3e instead of 80 GB. That matters for large models and long contexts, where VRAM runs out before compute does.

H200

Azure charges €11.83 ($13.50) per GPU hour for the H200 in Frankfurt, Verda €3.89 ($4.50). That is a little over three times as much.

B200 and B300

B200 and B300 are the Blackwell generation; NVIDIA calls the B300 Blackwell Ultra. VRAM differs by offer: 180 GB for the B200, 268 to 288 GB for the B300 depending on the provider. According to its own region overview, Nebius offers neither in a public EU region: its B200 runs in Israel and the US, its B300 in the UK.

A100, L40S and L4 for smaller models

Not every project needs an H100. For models up to about 8B parameters and for LoRA fine-tuning, smaller cards are often enough.

The strongest P-series training GPU AWS lists in Frankfurt is the A100, as p4de with eight GPUs at €3.76 ($4.29) per GPU hour.

Where the GPU runs: regions and headquarters

If your data has to stay in Germany, or at least in the EU, the region matters as much as the price. Two findings from the price lists affect the architecture directly:

• AWS Frankfurt (eu-central-1): the price list carries no P-series EC2 instance newer than the A100 (p4d, p4de) there, so no H100, H200 or B200. Of the EU regions we query, AWS has the H100 only in Stockholm and the H200 in Stockholm and Spain. Outside the EU, London also has H100 instances.

• Azure Germany West Central: the H100 is only available there as the NC series with one or two H100 NVL. In the EU, the ND H100 v5 series with eight SXM GPUs for distributed training is only listed in the Netherlands, Poland, Sweden and Spain. Azure does offer ND nodes with eight H200s in Frankfurt.

The headquarters column shows where the parent company is based. AWS and Microsoft are US companies and fall directly under the US CLOUD Act (18 U.S.C. § 2713), even for data in EU data centers. For European providers with US subsidiaries or US regions, it depends on the corporate structure. More in Training AI models under GDPR and the AI Act.

Nebius raises HGX prices on October 1, 2026

Next to its current on-demand prices, Nebius's pricing page shows a second column effective October 1, 2026. For HGX instances from H100 to B300 the increase ranges from 17 to 21 percent; the L40S stays the same.

At today's exchange rate, Nebius's H100 goes from €3.37 to €3.94. It stays well below AWS and Azure, but the gap to OVHcloud, Scaleway and Verda grows to about one euro per hour.

What a GPU server costs per month

A GPU running around the clock adds up to roughly 730 hours a month. At list price, a single H100 costs:

For distributed training, what matters are nodes with eight SXM GPUs connected by NVLink:

For a single GPU, the difference between the cheapest EU provider and Azure in Frankfurt is a little over €3,600 ($4,100) a month, and a little over €3,200 ($3,600) against Azure in Ireland. For an eight-GPU SXM node, it is about €63,000 ($71,000) between Verda and Microsoft Azure.

Rent, reserve or run dedicated GPU servers

List prices carry no discount at all. AWS and Azure discount through Savings Plans and reservations, and Nebius states up to 35% off for multi-month reservations of large clusters on its pricing page. Spot instances at Verda cost half the on-demand price but can be reclaimed at any time. That suits batch jobs, less so production inference.

Managed offerings fold operations into the hourly price. At Scaleway, the same H100 costs €3.40 ($3.88) per hour as Managed Inference and €2.87 ($3.27) as a bare instance.

A model that serves requests all day needs reserved GPUs configured for the workload, and someone to run them. That is what we offer as dedicated inference. If you want to train first, LLM fine-tuning in Europe lays out the four routes, and post-training on EU infrastructure shows how we run it for you.

Methodology

• As of: prices retrieved on September 24, 2026.

• Sources: OVHcloud order catalog API, Scaleway's GPU pricing page data (exact per-minute price times 60), Nebius pricing page, region overview and VM preset list, Verda instance-types API (api.datacrunch.io), AWS EC2 Price List, Azure Retail Prices API.

• Regions queried: AWS eu-central-1, eu-north-1, eu-south-1, eu-south-2, eu-west-1 and eu-west-3 in the EU, plus eu-west-2 (London) outside the EU; Azure every EU region where the Retail Prices API lists the VM.

• Conversion: AWS and Nebius only publish US dollar list prices; we show those as published and convert them to euros at the ECB reference rate of September 23, 2026, €1 = $1.1411. Verda publishes its own euro and US dollar prices, and both are shown as published. Microsoft Azure, OVHcloud and Scaleway publish euro prices; their US dollar figures in brackets are converted at the same ECB rate.

• Calculation: the instance's on-demand list price divided by its number of GPUs. Per provider and GPU variant we show the cheapest EU instance, plus Frankfurt for Azure.

• Not included: VAT, storage, data transfer, reservation and volume discounts, spot prices.

• Not compared: Google Cloud, because Google only publishes machine-readable prices through the Cloud Billing Catalog API with an API key. Dedicated GPU servers rented by the month are also out of scope; this comparison covers hourly cloud instances.

• Listed does not mean available: an instance in a price list is not automatically bookable right away.

• ¹ OVHcloud lists GPU instances in Gravelines, Limburg, London and Warsaw on its availability page, but does not say which model runs where.

• ² Verda (formerly DataCrunch, based in Helsinki) names EU locations but not a location per GPU.

Work out your GPU setup with us

Which GPU, region and pricing model fit depends on model size, utilization and where your data is allowed to live. We work it out with you in concrete numbers. Book a demo

Frequently asked questions

How much does an H100 cost per hour in Europe?

A single H100 costs between €2.80 ($3.20) at OVHcloud and €7.79 ($8.89) at Azure in Frankfurt, as the on-demand list price per GPU hour excluding VAT, as of September 24, 2026. In an Azure node with eight SXM GPUs it is €13.72 ($15.65) per GPU.

Where can I rent an H100 in Germany?

Among the providers compared here, only Azure lists an H100 located in Frankfurt, as the H100 NVL at €7.79 ($8.89) per hour. OVHcloud names Limburg as a GPU region but does not say which models run there. AWS lists no H100 as an EC2 instance in Frankfurt.

Why is the H100 so much more expensive at AWS and Azure?

Hyperscaler instances include more, such as large amounts of RAM, local NVMe and fast networking between nodes. Large customers also pay less than list price through Savings Plans and enterprise agreements.

Is the H100 NVL the same as an H100 SXM?

No. The H100 NVL is a PCIe card with 94 GB and, according to NVIDIA, 3.9 TB/s of memory bandwidth, more than the SXM variant with 80 GB and 3.35 TB/s. The SXM H100 has more FP8 compute and NVLink across every GPU in the node, while NVLink only bridges the NVL in pairs. For distributed training, SXM is the better choice.

Do the prices include VAT?

No. All prices are net, excluding VAT, storage and data transfer.

Is buying a GPU cheaper than renting?

It depends on utilization. At 730 hours a month, a rented H100 at €2.80 ($3.20) costs a little over €24,500 ($27,900) a year. The full calculation with purchase, power and operations is in LLM fine-tuning in Europe.

How often are the prices updated?

Monthly. The tables are generated by a script straight from the providers' price sources; the retrieval date and exchange rate are listed under Methodology.

Why is Google Cloud missing?

Google only publishes machine-readable prices through the Cloud Billing Catalog API with an API key. We will add Google Cloud once we check its prices automatically.

Sources

• OVHcloud Public Cloud Order Catalog

• OVHcloud: Regions availability

• Scaleway: GPU Instances pricing

• Nebius: AI Cloud pricing

• Nebius: Regions

• Verda: Pricing

• AWS: EC2 On-Demand Pricing

• AWS: EC2 instance types by Region

• Microsoft: Azure Retail Prices API

• NVIDIA: H100 Tensor Core GPU

• NVIDIA: H100 PCIe GPU Product Brief

• Nebius: VM types and presets

• Verda: Instance types API (EUR)

• Cornell LII: 18 U.S.C. § 2713 (CLOUD Act)

• ECB: Euro foreign exchange reference rates

As of September 24, 2026 · prices from the providers' own price sources, ECB rate of September 23, 2026 · kontinent.ai · Not legal advice.