European GPU Clouds: 8 Alternatives to AWS and Azure for AI Workloads

Key points:

• The eight European GPU cloud alternatives to shortlist against AWS and Azure: Scaleway and OVHcloud (EU-owned full-service clouds with managed Kubernetes, H100 and L40S), Nscale and Verda (AI-first clouds with H100/H200/GB200-class fleets), Hetzner and IONOS (cheaper EU-owned capacity with a narrower GPU catalogue), Exoscale (compliance-focused, EU-owned), and the Vast.ai/RunPod marketplace tier for burst and experimentation.
• On published on-demand list prices in EU regions, European GPU specialists undercut AWS p5 and Azure ND H100 v5 per GPU-hour, and marketplace hosts go lower still. Hyperscaler list prices are pre-discount, so compare against your actual Savings Plan or committed-use rate, not the sticker.
• An EU region on AWS or Azure delivers data residency, not corporate or jurisdictional independence: the operating entity's ultimate parent stays US-headquartered, which is the gap AWS European Sovereign Cloud and Microsoft's EU Data Boundary were built to close.
• Capacity, not price, is usually the binding constraint. Most European H100/H200 fleets are allocated by quota request, hourly reservation, or multi-month committed cluster, so confirm region-level availability in writing before you benchmark cost.
• Egress and data gravity, not compute, decide whether a migration pays back. Price per-GB transfer out of your current provider against a multi-terabyte training set before you compare GPU-hours, and check what the EU Data Act's switching rules change for your exit.
• Qovery does not sell GPUs. Qovery is the developer platform layer running inside your own cloud account (BYOC) on AWS, GCP, Azure, Scaleway, or any existing Kubernetes cluster, so you can move GPU workloads to whichever EU provider has capacity at the right price while keeping one deployment workflow.

---

What are the best European GPU cloud alternatives to AWS and Azure?

The eight European GPU clouds worth shortlisting against AWS and Azure are Scaleway and OVHcloud (EU-owned full-service clouds with managed Kubernetes, H100 and L40S), Nscale and Verda (AI-first clouds with H100, H200 and GB200/GB300-class fleets), Hetzner and IONOS (cheaper EU-owned capacity with a narrower GPU catalogue), Exoscale (compliance-focused and EU-owned), and the Vast.ai/RunPod marketplace tier for burst and experimentation. CoreWeave shows up later as a price benchmark, not a ninth pick, because it is US-headquartered with European data centres rather than an EU-owned provider. Here is the one-line verdict for each, written so you can lift any single line on its own.

• Scaleway (France, Iliad group): H100 PCIe, H100 SXM, L40S and L4 with managed Kubernetes (Kapsule) and public per-hour pricing. Best for EU teams that want a full cloud and GPUs behind one console.
• OVHcloud (France): H100, A100, L40S single cards plus 8xH200 nodes, managed Kubernetes (MKS), free egress and public pricing. Best for the widest published EU GPU catalogue.
• Nscale: an AI-first cloud running hydro-powered data centres in Glomfjord, Norway (EEA), with H100, H200 and GB200/GB300-class systems allocated by quote. Best for large committed training clusters.
• Verda (formerly DataCrunch, Finland): public on-demand and spot pricing for H100, H200, B200 and GB300. Best for price-sensitive training and fine-tuning with transparent rates.
• Hetzner (Germany): monthly dedicated GPU servers (NVIDIA RTX 4000 SFF Ada, RTX PRO 6000 Blackwell), not hourly cloud. Best for cheap, steady inference where you can commit monthly.
• IONOS (Germany, United Internet): NVIDIA H200 cloud VMs at a flat per-GPU-hour rate, EU hosting and ISO 27001. Best for regulated EU workloads that want a simple H200 rate.
• Exoscale (Austria, A1 Group): GPU families around NVIDIA A40, A30, V100 and RTX PRO 6000 Blackwell across Frankfurt, Vienna, Geneva and Zagreb. Best for compliance-heavy EU workloads that do not need H100 or H200.
• Vast.ai and RunPod (GPU marketplaces): H100 SXM and A100 by the hour at the cheapest headline rates, but European hosting, DPA coverage and uptime depend on the individual host you rent from. Best for burst and experimentation, not regulated production.

"European" here means two different things, and the distinction matters more than any price. Scaleway (Iliad), OVHcloud, Hetzner, IONOS (United Internet) and Exoscale (A1 Group) are EU-headquartered and EU-owned; Verda is Finland-based and Nscale runs data centres inside the EEA. AWS, Azure and CoreWeave operate EU regions or EU data centres while their ultimate parent stays US-headquartered, which is a data-residency story, not an ownership one.

On procurement speed, Scaleway, OVHcloud, Verda, IONOS, Exoscale, RunPod and Vast.ai publish list prices you can read today, while Nscale and CoreWeave's committed tiers are quote-only. That one fact often decides how fast you can start: a public price page means a card and a launch, a quote means a sales cycle. For a provider-by-provider look at managed Kubernetes support in Europe, control-plane pricing and SLAs, see our companion piece Which EU Cloud Providers Support Kubernetes?. This article stays on GPU availability, per-GPU-hour pricing, egress and sovereignty.

How does GPU pricing on European clouds compare to AWS and Azure per GPU-hour?

On published EU on-demand list prices checked in September 2026, AWS charges about $7.36 per H100 GPU-hour (p5.48xlarge at $58.89/hour for 8 GPUs in eu-north-1, Stockholm) and Azure charges about $15.98 per H100 GPU-hour (ND96isr H100 v5 at $127.82/hour for 8 GPUs in West Europe). European specialists list H100 well below both, with Scaleway at about EUR 2.87, OVHcloud at $2.99 and Verda at $3.25 per GPU-hour, and the Vast.ai/RunPod marketplace tier goes lower still with the least predictable quality. Treat marketplace rates as a floor, not a guarantee.

List prices checked September 2026. Every figure below comes from the provider's live public price page, linked inline. The 8-GPU hyperscaler and CoreWeave nodes are normalised to per GPU-hour by dividing the node price by 8; European single-card instances are already per GPU. Prices are shown in the currency each provider publishes (USD or EUR), and I have not converted between them because exchange rates move.

Here is the same workload priced three ways, on-demand, for one 8x H100 node over 30 days (720 hours):

• AWS p5.48xlarge (eu-north-1): $58.89/hour x 720 = about $42,401 for the month.
• OVHcloud (8x h100-380, Gravelines): $2.99 x 8 = $23.92/hour x 720 = about $17,222.
• Scaleway (8x H100 PCIe, Paris): EUR 2.87 x 8 = EUR 22.96/hour x 720 = about EUR 16,531.

Same eight H100s, same 30 days, and the AWS list bill runs roughly 2.5x the Scaleway or OVHcloud bill, before you apply a single discount and before AWS egress and storage are added on top.

That last clause is where honesty cuts the other way: those AWS and Azure numbers are list prices, and few teams at scale pay list. AWS Compute Savings Plans advertise up to 66% off on-demand and EC2 Instance Savings Plans up to 72% on a one- or three-year commitment (AWS Savings Plans). Azure reserved VM instances advertise up to 72% versus pay-as-you-go, and the Azure savings plan for compute up to 65% (Azure reservations). Compare a European provider's rate against your actual committed rate, not the sticker: a three-year AWS commitment closes a lot of the gap, though rarely all of it.

A cheap European per-GPU-hour can also hide what it excludes. Check whether the price includes fast block storage IOPS, the interconnect you need (PCIe H100 and SXM H100 are not the same for multi-GPU work, and InfiniBand or RDMA fabric is often a separate line), your support tier, and the idle time between jobs. And spot or interruptible capacity, like Verda's H100 spot at $1.63 versus $3.25 on-demand, is a real discount only if you checkpoint. A preempted multi-day training run without checkpoints can lose days of progress, which is far more expensive than the hours you saved.

Does an EU region on AWS or Azure count as data sovereignty?

An AWS or Azure EU region gives you data residency and real operational controls, but not jurisdictional independence, because the operating entity's ultimate parent is US-headquartered and remains in scope of US law such as the CLOUD Act, which compels US-based providers to disclose data in their control "regardless of whether such communication, record, or other information is located within or outside of the United States" (US Congressional Research Service). Closing that gap is the whole point of AWS European Sovereign Cloud and Microsoft's EU Data Boundary. Teams conflate three separate things, and it helps to keep them apart.

Both hyperscaler programs deserve real credit. AWS European Sovereign Cloud is a EUR 7.8 billion investment through 2040, with its first region in Brandenburg, Germany, generally available since 15 January 2026, operated by a new EU parent company and three German subsidiaries with EU-citizen leadership and, in AWS's words, "zero operational control outside of EU borders" (AWS). Microsoft's EU Data Boundary now keeps customer data, pseudonymized personal data and support data for Microsoft 365, Dynamics 365, Power Platform and most Azure services inside the EU and EFTA, with the final phase completed in February 2025 (Microsoft). In April 2025 Microsoft also pledged to contest in court any government order to suspend its European cloud operations (Microsoft European Digital Commitments).

The residual point is narrow, and worth stating precisely. The Court of Justice's Schrems II ruling (Case C-311/18, 16 July 2020) struck down the EU-US Privacy Shield and left transfers to US-controlled providers dependent on standard contractual clauses plus a case-by-case assessment of US surveillance law (CJEU). Residency programs reduce that exposure; they do not, on their own, move the parent company outside US jurisdiction. This is not legal advice, and if you carry that risk you should take the specifics to counsel.

The drivers behind all of this are concrete regulations, not vibes: GDPR, Schrems II, and for financial services DORA (Regulation (EU) 2022/2554), which has applied since 17 January 2025 and requires documented ICT third-party risk management and exit strategies (EUR-Lex), alongside national health and public-sector rules and France's SecNumCloud. The market backdrop is lopsided: US hyperscalers hold roughly 70% of the European cloud market while European providers sit around 15%, down from about 29% in 2017, per Synergy Research Group (Synergy Research Group).

On ownership, the EU-owned tier is easy to name: Scaleway (Iliad), OVHcloud (listed in France), Hetzner (Hetzner Online, Germany), IONOS (United Internet, Germany) and Exoscale (A1 Group, Austria). On certifications, SecNumCloud (ANSSI, France) qualifies a specific offering at a specific scope and includes protections against non-EU extraterritorial law, with OVHcloud, Outscale and Scaleway among qualified providers; BSI C5 (Germany) is an audited criteria catalogue (121 criteria in the C5:2020 version); ISO 27001 and SOC 2 are baseline and provider-agnostic. When a vendor waves a certification at you, ask what offering and scope it actually covers.

A copyable three-question test cuts through most sovereignty pitches: (1) Who owns the operating entity that runs the region? (2) Who can touch the control plane, and are they all in the EU? (3) Which law governs the contract, and what happens under a foreign disclosure order? If the honest answer to any of these lands outside the EU, you have residency, not independence.

Can you actually get H100 and H200 capacity in Europe right now?

High-end European GPU capacity is mostly allocated by quota request, hourly or monthly reservation, or a multi-month committed cluster rather than instant on-demand, so getting H100 or H200 capacity is a procurement question you settle before you benchmark per-GPU-hour cost. Instant on-demand is realistic for L4, L40S and marketplace GPUs; H100 and H200 at node scale and up usually mean a request form or a sales conversation. In my experience talking to teams, the sticker price is rarely what blocks a launch. Availability in the specific region you need is.

There are four access patterns, and each provider leans on one or two:

• Instant on-demand: L4, L40S and single H100/L40S cards on Scaleway, OVHcloud, IONOS and Exoscale, plus marketplace GPUs on Vast.ai and RunPod. Click and run.
• Quota request: hyperscaler H100/H200 (AWS p5, Azure ND v5) frequently gated behind a service-quota increase per region.
• Hourly or monthly reservation: Verda spot and reserved, Hetzner monthly dedicated servers.
• Multi-month committed cluster: Nscale and CoreWeave's larger fleets, sold as reserved capacity with a term.

Before you benchmark anything, put these in writing with the provider: region, node count, exact GPU model, interconnect (InfiniBand versus Ethernet), minimum commitment, notice period, exit terms, and the remedy if they cannot deliver on the promised date. A verbal "we have H100s" is not capacity; a signed availability date in a named region is.

The build-out is real and worth tracking. Nscale runs a hydro-powered site in Glomfjord, Norway, described as 30MW expandable to 60MW on 100% renewable power (Nscale). CoreWeave committed an incremental $2.2 billion to continental Europe in June 2024, adding data centres in Norway, Sweden and Spain on top of its UK operations (CoreWeave). At the policy level, the EU has selected 19 EuroHPC AI Factories and, through InvestAI, is mobilising around EUR 200 billion for AI with roughly EUR 20 billion earmarked for AI gigafactories (European Commission). Reported H100 lead times eased through 2024 and 2025 from several months toward roughly 8 to 12 weeks (Tom's Hardware), while H200 demand has repeatedly run ahead of supply. Specific cluster GPU counts move fast enough that I would confirm them against a provider's own newsroom before quoting a number.

The practical strategy most teams land on is multi-provider: committed capacity for the training runs you can schedule, and elastic or marketplace capacity for inference and experiments you cannot. That only works if your workloads are portable, which is the whole argument for section six. A pipeline welded to one scheduler, one storage class or one managed endpoint cannot chase capacity to whichever region opens up next.

How do you choose between European GPU clouds for training, fine-tuning, and inference?

Choose by workload shape, not by brand. Multi-node training needs InfiniBand-class interconnect plus committed capacity; fine-tuning fits a reserved single 8x H100 or 8x H200 node; batch and real-time inference favour the cheapest per-GPU-hour with autoscaling and low cold start on L40S, L4 or RTX-class cards. Get the shape right and the provider shortlist mostly picks itself.

For multi-node training, interconnect bandwidth and parallel storage throughput decide throughput more than the raw GPU rate, and you will run either Slurm or Kubernetes with the NVIDIA GPU Operator. For fine-tuning, the card's memory is the constraint: an NVIDIA H100 SXM carries 80GB of HBM3 at about 3.35 TB/s, while an H200 carries 141GB of HBM3e at about 4.8 TB/s (NVIDIA H100, NVIDIA H200), so the H200's roughly 1.76x memory and 1.43x bandwidth let you fit larger models or longer context without sharding. For inference, an L40S or L4 with scale-to-zero and request-level autoscaling usually beats an H100 on cost per request, and remember egress is charged on response traffic. For compliance-heavy workloads, start from an EU-owned provider with clear DPA terms, a published subprocessor list, and the certifications your auditor wants (ISO 27001, SOC 2, SecNumCloud or C5).

If you are an EU startup shipping an inference API, choose Scaleway or OVHcloud L40S with autoscaling. If you are a scale-up fine-tuning open models, reserve a single 8x H100 or H200 node on Verda or Scaleway. If you are a regulated enterprise with residency obligations, choose an EU-owned provider (or a sovereign-cloud tier) and lead with certifications. If you are a research team needing a burst cluster, use a marketplace or a short reservation, and checkpoint aggressively.

Scaleway versus OVHcloud is the head-to-head most EU teams actually run, and they are close. Both are French, EU-owned, and ship managed Kubernetes with GPU node pools. Scaleway lists H100 PCIe at about EUR 2.87 per GPU-hour, offers H100 SXM with MIG and fast scratch storage, and wires GPUs into Kapsule through the NVIDIA GPU Operator. OVHcloud lists H100 at $2.99 per GPU-hour, carries a wider catalogue (A100, L40S and 8x H200 nodes), and gives you free egress out of Gravelines. If you want the broadest published catalogue and free egress, OVHcloud; if you want H100 SXM with MIG and scratch storage tightly integrated with managed Kubernetes, Scaleway. The prices are close enough that catalogue, region and egress usually decide it.

How do you run the same application stack across multiple GPU clouds without rebuilding your pipeline?

Standardise on Kubernetes and keep the deployment workflow in a layer no GPU provider owns. That is how teams move workloads between Scaleway, OVHcloud, an AI-first GPU cloud or their own cluster without rewriting CI/CD every time pricing or capacity shifts. The GPU vendors already made the runtime portable for you: the same NVIDIA GPU Operator that runs on Scaleway Kapsule and OVHcloud MKS also runs on Amazon EKS, Google GKE, Azure AKS and OpenShift.

Kubernetes is the portability contract. The NVIDIA GPU Operator installs and manages the drivers, the container toolkit, the Kubernetes device plugin, GPU feature discovery and DCGM monitoring, so a GPU pod looks the same everywhere (NVIDIA GPU Operator docs). Workloads request GPUs through the standard nvidia.com/gpu resource limit, with node selectors, taints and tolerations steering pods onto GPU node pools (NVIDIA k8s-device-plugin). Both Scaleway Kapsule (Scaleway docs) and OVHcloud MKS (OVHcloud docs) support GPU node pools this way.

What is portable: your container images, Kubernetes manifests, Helm charts and CI pipelines. What is not: proprietary schedulers, provider-specific storage classes, and managed model endpoints. Keep the first list large and the second list small and you can move.

Qovery does not sell GPUs, and it does not replace a GPU provider. Qovery is the developer platform layer that deploys and operates your apps inside your own cloud account (BYOC) on AWS, GCP, Azure, Scaleway, or any existing Kubernetes cluster including self-managed and on-prem. Because it runs in your account, the GPU bill and any committed-use discounts stay in your name, and you keep one deployment workflow while you pick whichever EU provider has capacity at the price you want.

The capabilities that matter for GPU work are the mundane ones: git-push deployments, a preview environment per pull request, environment auto-stop for non-production (a direct cut to the idle GPU spend that quietly dominates dev and staging bills), managed cluster upgrades, per-environment RBAC, and databases backed by managed cloud services. None of this frames anything as AWS-only: the same workflow runs on GCP, Azure, Scaleway or your own Kubernetes cluster. For the provider-by-provider Kubernetes detail this article does not repeat, see Which EU Cloud Providers Support Kubernetes?.

What does a realistic migration off AWS or Azure GPU instances cost?

Moving GPU workloads is a data-gravity and pipeline problem more than a compute problem, so price egress and dataset replication before the GPUs. For a multi-terabyte training set, transfer-out charges alone can wipe out your first month of GPU savings. The step order that works: inventory your GPU jobs, benchmark on the target hardware, price egress and transfer time, move inference first and training last, and keep a documented fallback the whole way.

Egress is the bill people forget. AWS gives 100GB of internet egress free per month, then charges about $0.09/GB for the first 10TB from EU regions (AWS EC2 on-demand); Azure gives the first 100GB free, then about $0.087/GB in Zone 1, which includes Europe (Azure bandwidth). Move a 50TB training set out of AWS and, after the free tier, roughly 50,000GB at $0.09/GB is about $4,500 for a single copy. By contrast, OVHcloud public cloud charges nothing for egress (OVHcloud prices), and Hetzner includes 20TB per month on a 10 Gbit uplink then bills about EUR 1 per additional TB (Hetzner traffic). That asymmetry is often the whole business case.

The rules are shifting in your favour. The EU Data Act (Regulation (EU) 2023/2854), applicable since 12 September 2025, forces cloud providers to make switching easier and phases out switching and data-transfer charges, with all switching charges due to be prohibited from 12 January 2027 (European Commission). If you are planning a multi-year GPU strategy, that timeline changes what your exit will cost.

Transfer time matters as much as transfer cost. A 50TB set over a sustained 10 Gbps link is about 11 hours at line rate, and real throughput is lower, so plan for days; past a few dozen terabytes, physical transfer appliances start to beat the wire. On dual-running, keep the overlap deliberately short and move inference before training so you are never paying for two committed GPU fleets at once. And be honest about the team cost: someone has to patch and upgrade the new cluster, manage RBAC and watch it, which is exactly the work you can hand to a platform layer instead of hiring for.

A copyable pre-migration checklist:

1. Inventory every GPU job, its GPU model, and its data dependencies.
2. Benchmark the top three jobs on the target hardware before committing.
3. Price egress for a full dataset copy out of your current provider.
4. Confirm H100/H200 availability in writing for your target region.
5. Check the interconnect (InfiniBand versus Ethernet) on the new cluster.
6. Move inference first; keep training on the old cluster until it is proven.
7. Set a short, dated dual-running window and stick to it.
8. Verify DPA, subprocessor list and certifications for compliance workloads.
9. Standardise the deploy workflow on Kubernetes so CI/CD does not get rewritten.
10. Document a fallback path and a commitment-break plan before you cut over.