European alternatives to AWS and Azure in 2026: 24 providers checked

Most comparisons of European clouds stop at the price of a virtual machine. This one prices the same two workloads on 24 cloud providers headquartered in Europe and on AWS and Azure in Frankfurt, then looks at what the price leaves out: what a zone physically is, where the database standby sits, who controls the company, what the SLA actually pays and how long a new account waits before it can build anything.

It is written for developers, tech leads, founders and product managers deciding where to build, or whether to move off AWS or Azure. Everything was checked on 15 and 16 October 2026, and each provider has its own article in this series with the detail behind the summary here.

The method, in one sentence: every fact comes from the provider's own documentation, legal terms, price data and status page, from public filings and registers, or from user reports, and each one was checked at its source. What we did not do: we did not open accounts or run workloads. Nothing here measures performance, and where a provider's documents and its behaviour may differ, we can only report the documents and what its users have written.

The short answer

You want managed services close to what AWS gives you. Start with Scaleway: managed Kubernetes with a free control plane, PostgreSQL, serverless functions and containers, queues and KMS, in a Paris region with three zones in separate buildings. IONOS and UpCloud are the cheapest providers that sell every part of our production workload. IONOS costs EUR 436.79 a month with 24/7 phone support included, or EUR 524.39 if a database core counts as one thread rather than two; its price page once said two and no longer does. UpCloud costs EUR 520.33. STACKIT promises at least three physically separate zones per region and sets service levels in a legal certificate per service, but only companies in Germany, Austria and Switzerland can sign up on their own. T Cloud Public (Open Telekom Cloud until January 2026) has a hyperscaler-shaped catalogue and a synchronous database standby you can place in another zone, at a price above AWS.

You run your own stack and want cheap compute. Hetzner includes at least 20 TB of egress per EU cloud server and sells dedicated servers in the same parks, but it has no managed database, managed Kubernetes or zones, it raised prices steeply in 2026, and since 26 June it has restricted new cloud servers for new customers and for some existing ones, chosen at random. Cyso Cloud builds our production workload, minus the managed database it offers only as a closed beta, for EUR 254.40 in a three-zone Frankfurt region. Infomaniak charges nothing for instance traffic and keeps data in Switzerland. OVHcloud includes egress in all its European regions. Leaseweb sells bare metal in volume next to its cloud.

You need GPUs. Nebius lets a new account run up to 32 H100s in Finland after a USD 25 card charge, but it raised GPU prices twice in 2026. UpCloud lists the cheapest self-serve H100 we found, at EUR 1.79 an hour in Helsinki, with new prices announced from 20 October 2026; how many GPUs an account may run depends on its history with UpCloud. Gcore (EUR 2.73) and OVHcloud (EUR 2.80, though a paid quota tier may be needed first) also sell H100s without a sales call. For Blackwell, Evroc charges EUR 8.00 per B200 GPU-hour in Stockholm, about half of AWS's Stockholm price, but every account starts with a GPU quota of zero.

You sell to the public sector or a regulated industry. OVHcloud holds three SecNumCloud qualifications on the French cybersecurity agency's list. STACKIT, Scaleway and a partnership of Post Telecom, OVHcloud and Clever Cloud were rated SEAL-3 in the European Commission's sovereign cloud tender. For BSI C5, Hetzner has a Type 2 attestation with a clean auditor's opinion, and STACKIT, Exoscale and T Cloud Public provide theirs on request or under NDA. The full Type 2 report that plusserver publishes carries a qualified opinion: two C5 controls, and a third reported in the same document, were not effective in 2025. In Sweden, Elastx is 53% owned by the Swedish investor Sobro, keeps data in Sweden, says it is covered by Sweden's NIS2 law and offers security protection agreements.

You want a platform, not servers. Clever Cloud runs Heroku-style deployments on its own servers in Paris under a French contract, but has no self-serve database failover. Northflank gives you services, databases, preview environments and GPUs on one per-second bill, but its managed cloud is Google Cloud and three venture firms (Bain Capital Ventures, Vertex Ventures US and Kindred) hold a majority of its shares. Upsun runs six of its seven European regions on AWS, Google Cloud or Azure. Koyeb, now part of Mistral AI, deploys containers quickly but sells its H100-class GPUs only in North America.

What it costs, like for like

Our production workload (workload B) is three VMs, block storage, a load balancer, a highly available PostgreSQL, object storage, a Kubernetes control plane and 2 TB of egress. Of the 26 clouds in this comparison, only nine sell every part of it, AWS and Azure included. These are their totals, as list prices in EUR per month, VAT excluded, checked 15 and 16 October 2026.

The cheapest totals and the strongest zones do not line up. IONOS and UpCloud, the two cheapest, have no zones documented as separate buildings in Frankfurt. Scaleway is the cheapest with three zones in separate buildings, but its database standby stays in the same data centre. Of the European providers in this table, only two document a standby in another zone, T Cloud Public (if you choose it) and Elastx (when possible), and both cost more than AWS. On dedicated rather than shared vCPUs, IONOS rises to EUR 603.23 (on the two-thread database reading) and Scaleway to EUR 676.78. The workloads and the rules we priced them by are described below.

The lower numbers you will see elsewhere usually leave out the database or the Kubernetes control plane. Hetzner's EUR 241.11 for workload B has neither, because Hetzner sells neither: your team runs both. Infomaniak's EUR 132.10 has no PostgreSQL: its managed PostgreSQL is marked "Available soon", and an April 2025 post already said more managed databases were "coming soon". Cyso Cloud's PostgreSQL is a closed beta with no public price. Aruba Cloud's EUR 267.91 prices MySQL in place of PostgreSQL and has no load balancer for VMs on its new platform. Gcore's EUR 384.60 includes a PostgreSQL that is a free beta. The database is not a small line: at Exoscale the highly available PostgreSQL is EUR 384.20 of the EUR 908.88.

How we priced

Workload A, a starter setup: one VM with 2 vCPU and 4 GB of RAM, at least 40 GB of disk, one public IPv4 address, 1 TB of internet egress a month and 100 GB of object storage.

Workload B, a small production system: three VMs of 4 vCPU and 16 GB, each with a public IPv4 address; 300 GB of SSD block storage; one managed load balancer; managed PostgreSQL with 2 vCPU, 8 GB and 100 GB of storage, primary plus standby; 1 TB of object storage plus 1 million PUT and 10 million GET requests; one managed Kubernetes control plane (the highly available one where offered, control plane only); 2 TB of internet egress a month.

List prices on demand, with no commitment or discount, 730 hours a month, VAT excluded, in the provider's region nearest to Frankfurt (named in the tables). Where no exact size exists, the next larger one. Where a part is not sold, it is left out of the total and named, rather than replaced by a guess; where a provider sells a close substitute (MySQL for PostgreSQL), we priced it and said so. Developer platforms are priced on their nearest equivalent. AWS is priced on m7i.xlarge and Azure on D4s_v5; on burstable instances AWS's workload B falls to EUR 985.89 and Azure's to EUR 1,017.63. Prices in USD, CHF and SEK are converted at the ECB reference rates of 15 October 2026 (1 EUR = 1.1208 USD = 0.9305 CHF = 11.1765 SEK), except where a provider has its own EUR price, which we used.

Everyone else

The remaining providers, with what their total leaves out. Read the last column before the price.

For comparison, workload A costs EUR 112.37 on AWS and EUR 116.15 on Azure, and on AWS the 1 TB of egress alone is USD 81.00 of the USD 125.95. Nscale publishes no price for VMs or storage, so it is not in either table. CloudSigma's totals are pay-as-you-go: on one-month subscriptions in Geneva, workload A is EUR 34.19.

Zones and high availability, honestly

On AWS and Azure a zone is one or more data centres with their own power, cooling and network, kilometres from the other zones and within about 100 km of them. A design spread over three such zones survives the loss of a building. Many European providers use the word for something smaller, or have no zones at all, and that changes what "highly available" means.

Zones in separate buildings, in the region we priced: AWS and Azure (three in Frankfurt), Scaleway (three in Paris), Elastx (three data centres in Stockholm up to 20 km apart), plusserver (two, in Cologne and Düsseldorf) and Clever Cloud, whose count depends on the page: five Paris data centres on its infrastructure page, two availability zones in its status entries and three in its Kubernetes docs.

Three zones on two sites: T Cloud Public's three zones sit on two sites 18 km apart, so two of them share a site.

Separate data centres on one campus: Aruba Cloud's three Bergamo zones "do not share physical infrastructure, such as Data Centers" by Aruba's definition, but all three sit on its Bergamo campus, and Aruba does not say how far apart they are.

Separate by the provider's own definition, sites not named: Evroc describes its three Stockholm zones as physically separate data centres but does not name the facilities.

Physically separate, but possibly in one building: STACKIT promises "at least three physically separate AZs" per region, but its Compute Engine certificate says "Several Availability Zones can be located in the same building", and it does not say which form eu01 takes.

Zones in some regions only, or not physically described: OVHcloud has three zones about 30 km apart in Paris, and in Milan, but its Frankfurt region is one zone. Leaseweb has two zones in Amsterdam, defined in its contract as separate power, cooling and network, and none in Frankfurt or London. Cyso Cloud has three zones in Frankfurt and in Amsterdam without saying what a zone physically is.

Zones that are smaller than a building: IONOS describes its zones as different hosts or storage systems, and no IONOS page says they are separate buildings or power. To get a second building you deploy in a second Frankfurt location (IONOS has three), and traffic between IONOS data centres in the same country is free. Infomaniak's zones have "different network connectivity and power inputs", for compute only, and their names suggest they sit in one data centre.

No zones within a region: Hetzner (spread placement keeps your servers on different hosts), Exoscale (a zone is one facility in one city), UpCloud (one data centre per city, except Helsinki), Gcore, Cleura's self-service Public Cloud, CloudSigma, Nebius (none documented) and Upsun (none defined). On these, protection against losing a building means a second city, and at Upsun that is a manual rebuild in a second region with no automatic failover. Koyeb's Frankfurt region spans three data centres and reschedules instances within the region when one fails, but Koyeb does not document whether they are separate buildings.

Where the database standby sits. "Highly available" on a managed database tells you little until you know this.

AWS and Azure state a failover time of 60 to 120 seconds, T Cloud Public "within 5 minutes", and OVHcloud fails over after a 60-second timeout. Scaleway offers multi-zone placement only for read replicas. At Clever Cloud a replica is set up by sales or support and failover is manual, which by Clever Cloud's own page takes "from a few dozen minutes to a few hours".

Zones do not stop shared layers from failing. OVHcloud's Object Storage returned errors in at least nine regions on 16 July 2025, all three Paris zones among them, for 2 hours 10 minutes, and on 23 September 2026 all three Paris zones had network latency and packet loss for 3 hours 19 minutes. On 10 March 2025 a router configuration error disrupted every product in Scaleway's Paris region: the network for five minutes, Instances until 17:10 UTC, and Object Storage and Kubernetes until 18:00 UTC. A power loss in fr-par-2 on 22 May 2026 degraded the other Paris zones. On 17 June 2026 a STACKIT Kubernetes incident cut every eu01 cluster for 2 hours 37 minutes. Elastx lost routing across all three zones for 8 minutes on 26 September 2025. Cyso Cloud's Amsterdam region lost external connectivity region-wide three times in 2025. Control planes are often shared too: Exoscale's API and portal were down in all eight zones for about seven hours on 1 September 2026 (running servers were unaffected), and UpCloud could not deploy new servers in any of its 15 zones for about two hours on 11 August 2026.

The test that went the other way: a fire at the NorthC data centre in Almere took T Cloud Public's eu-nl zone AZ2 offline from 7 May to 3 June 2026. By the provider's account AZ1 and AZ3, at another site in Aalsmeer, kept running, and customers were told to fail over to them.

Who controls each provider, and whose hardware it runs on

Every provider in this series is headquartered in Europe; that was the only filter. Control was reported, not used to exclude anyone. These are facts from filings, registers and the providers' own documents, not legal conclusions about which laws can reach them.

Three points the table compresses. Exoscale's ultimate control is in Mexico. Northflank, Upsun and Koyeb are European companies whose platforms run, wholly or in part, on US clouds: Northflank's regions API maps every region to a Google Cloud region, Upsun's region list puts six of seven European regions on AWS, Google or Azure, and Koyeb's DPA places its control plane on Google Cloud in Belgium. And Koyeb's contracts still name Koyeb SAS, which was struck off the French register on 19 May 2026 after its assets passed to its sole shareholder.

The sovereign versions of AWS, Microsoft and Google

Many teams first ask whether they can simply buy a sovereign edition of the cloud they already use. On 16 October 2026 the answer differs by vendor.

AWS European Sovereign Cloud is a separate AWS partition with one region, in Brandenburg, generally available since 14 January 2026 and open to anyone through a separate, self-serve account. Customers in the EU contract with Amazon Web Services EMEA SARL in Luxembourg; German companies set up for the purpose run it, with staff who are EU residents or EU citizens located in the EU (AWS says it is moving to EU citizens only, with no end date). The catalogue is smaller: 109 products on its price page, 311 of Frankfurt's 1,163 EC2 instance types, and one GPU family (L4). EC2 is priced in EUR, about 10 to 11% above Frankfurt's USD price at the ECB rate of 15 October. Its contract commits AWS to redirect a government request to the customer, notify the customer where allowed and challenge requests that conflict with EU law; if AWS then "remains compelled to disclose Your Content", it "will disclose to the Requesting Party only the minimum amount of Your Content necessary to satisfy the Request."

Microsoft's Sovereign Public Cloud is not a separate cloud. It is a set of controls (Data Guardian, external key management, tamper-evident access logs) on the existing European Azure regions, plus the EU Data Boundary. Microsoft wrote in February 2026 that residency, encryption and regulated-environment controls are available in all European regions; we found no dated general availability for Data Guardian, which lets authorised staff residing in Europe approve and monitor remote access by engineers elsewhere. Microsoft's National Partner Clouds are owned by others: Bleu (100% Orange and Capgemini) runs Azure and Microsoft 365 in two French regions for named customers and is still in SecNumCloud qualification, and Delos Cloud (an SAP subsidiary, operated by security-cleared German staff of Arvato Systems) sells only to the German public sector; on 1 October 2026 its homepage still described it in the future tense, with a pilot customer programme.

Google offers controls on its public regions (Data Boundary) and, through partners, separate clouds on Google Cloud Dedicated. The live one is S3NS PREMI3NS, run by Thales Cloud Sécurisé in France with no connection to Google's network: generally available since 16 October 2025, SecNumCloud 3.2 qualified on 17 December 2025, about 40 products in one region, sold through sales without published prices. A German version run by a new Thales entity is in preview.

What each said under oath. Two French parliamentary inquiries asked Microsoft, Google and AWS, under oath, whether they would hand over French public or private data on a lawful US order. Microsoft France's director of public and legal affairs, Anton Carniaux, before the Senate on 10 June 2025: "Non, je ne peux pas le garantir, mais, encore une fois, cela ne s'est encore jamais produit." ("No, I cannot guarantee it, but, once again, it has never happened yet.") And: "Lorsque nous sommes obligés de les donner, nous les donnons." ("When we are obliged to hand them over, we hand them over.") Before the Assemblée nationale on 13 May 2026, Google Cloud's Frédéric Geraud de Lescazes: "si un juge allemand, canadien, chilien, américain ou français nous demande de coopérer, nous le faisons." ("If a German, Canadian, Chilean, American or French judge asks us to cooperate, we do.") AWS's Arnaud David said the question "ne s'est pas posée à nous" ("has not arisen for us") and that "selon la loi, on ne peut fournir des données que dans la mesure où on y a accès sur le plan technique" ("under the law, data can only be provided to the extent that one has technical access to it"). Microsoft France's president, Corine de Bilbao, the same day: "nous devons nous conformer aux obligations des différents États où nous opérons" ("we must comply with the obligations of the various states in which we operate"). The rapporteur, Cyrielle Chatelain, noted that only one company had said it would cooperate with the courts, and read the others' failure to answer yes or no as a yes; that is her reading, not the companies' statement. The translations are ours.

The one EU-level rating so far. On 17 April 2026 the European Commission awarded four contracts in a sovereign cloud tender worth up to EUR 180 million over six years, scored with its Cloud Sovereignty Framework. Bidders needed SEAL-2 to be eligible. Post Telecom with Clever Cloud and OVHcloud, STACKIT and Scaleway reached SEAL-3 ("EU actors exercising meaningful but not full influence"); Proximus, with S3NS on Google technology, reached SEAL-2. The Commission's conclusion: "even non-European technologies, when operated within a strict and appropriate framework, can meet the minimum level of sovereignty required." Neither AWS nor Microsoft is among the four, and the Commission does not say whether they bid. We found no act adopting the EU's cloud certification scheme, EUCS, and the Cloud and AI Development Act, which would add four Union assurance levels, is a proposal: our article on the Cloud and AI Development Act covers what each level would require.

GPUs

On-demand prices per GPU-hour in EUR, in whichever region sells the card. USD prices are converted at the ECB rate of 15 October 2026.

Three things the table does not show. AWS does not sell H100 in Frankfurt at all, and its GPU quotas start at zero. Scaleway's contract lets it terminate an on-demand H100 PCIe instance with 10 minutes' notice. Nebius raised GPU prices on 1 June and again on 1 October 2026, taking the H100 from USD 2.95 to USD 4.50 an hour, and its contract allows three days' notice of a price change. Below the top end, Hetzner sells RTX PRO 6000 Blackwell cards as dedicated servers (GEX63: EUR 997.30 a month cap plus EUR 999.00 setup), Exoscale sells the same card at EUR 2.15 an hour after manual screening, and an L40S costs EUR 0.63 an hour at Infomaniak (on request) and EUR 1.11 at UpCloud, against EUR 2.08 on AWS.

Two European GPU companies sell almost entirely to very large customers. Nscale (London) publishes no GPU price and sells GPU instances only through reserved contracts; self-serve covers its model API. Its S-1 filing of 18 September 2026 lists Microsoft statements of work worth up to about USD 43.8 billion through December 2033 and Anthropic agreements worth up to about USD 44.6 billion, says its largest customer was 73% of 2025 revenue, and describes on-demand clusters, offered "to support smaller customers", as "not currently a material component". Fluidstack is described at the end of the provider list.

Guarantees, support and getting an account

The patterns below matter more than any single SLA percentage.

The credit is the only remedy, and it is capped. OVHcloud's SLA credits are its sole remedy, capped at 50% of an instance's monthly cost and 30% for databases and load balancers. After the 2021 fire in Strasbourg, the Douai court of appeal upheld OVHcloud's liability cap on 24 April 2025 and awarded a VPS customer EUR 1,800.48 against a proven loss of EUR 47,900. Gcore caps credits at 20% (25% for Object Storage) and issues none under EUR 25; Nebius caps them at 30%. Hetzner pays only in cloud credits, Scaleway only in vouchers. Clever Cloud pays its SLA penalties in euros rather than vouchers. T Cloud Public's standard terms pay no credits at all. Credits come only with paid add-ons: from EUR 3,500 a month under the Enterprise Support Agreement 2.0, which is its EUR 2,500 base fee plus the EUR 1,000 service credits option, or from the Gold level at EUR 5,500 under the older Enterprise Agreement 1.0.

Downtime may start when you report it. At OVHcloud, for instances and storage, downtime is "calculated by OVHcloud from the moment the incident ticket is opened". At Gcore, for most services it counts only from the customer's emailed defect report (for Object Storage, from the incident ticket Gcore opens).

Claim windows are short. Leaseweb gives five days, Aruba Cloud 10, Hetzner 14, IONOS two weeks with request logs, UpCloud 15, and OVHcloud, Scaleway, Gcore and Cyso Cloud 30. Azure allows 60 days; AWS until the end of the second billing cycle. Keep your own monitoring logs, because several providers ask for them.

Some platforms have no SLA unless you pay or negotiate for one. Northflank has none below Enterprise, and caps its liability at the lesser of fees paid or USD 100. Koyeb has none on its Pro plan; 99.9% comes with Scale at USD 299 a month. Upsun's default is "commercially reasonable efforts" for 99.5%, with no credits; 99.9% and 99.99% are list-priced add-ons. Cleura publishes no SLA document. plusserver sends its SLA with the order confirmation, so you cannot read it before you order.

Included support ranges from nothing to 24/7. AWS's free Basic tier opens no technical cases, and its Developer, Business and Enterprise On-Ramp plans end on 1 January 2027. Developer and Business give way to Business Support+ from USD 29 a month or 9% of charges, whichever is greater; Enterprise On-Ramp customers move to Enterprise Support. IONOS includes 24/7 phone and email in German and English with a contractual first response under one hour for malfunctions. Elastx includes 24x7 support with 15 minutes for urgent cases, and Nebius gives every account a 30-minute first-response target for critical tickets. OVHcloud's included support targets 8 business hours for a critical incident; Hetzner and Aruba Cloud publish no response targets.

Not everyone may sign up. IONOS, UpCloud, Koyeb and T Cloud Public accept businesses only, and Leaseweb refuses individuals. STACKIT and plusserver offer self-service only to businesses in Germany, Austria and Switzerland; everyone else goes through sales. Elastx is sales-led: "Sign up" means requesting a test account.

Identity checks can stop you. Hetzner runs a risk check on every account; flagged accounts must show ID (to staff or through an iDenfy selfie) or prepay, and when a verification fails Hetzner says it does "not provide specific details" about why. Scaleway asks for photo ID and a live face check before you can use GPUs, outbound SMTP, more than 500 GB of block storage or the larger instance types in its quota table (POP2-4C-16G, PRO2-XS and up); that table did not yet list the newer families we priced. Leaseweb's terms allow up to five business days of KYC before an order is confirmed, and it quotes 1 to 3 business days before new Kubernetes customers reach the portal. OVHcloud may hold an order for manual verification for 24 to 48 business hours.

Default quotas can block a production build on day one. AWS's documented default for standard on-demand instances is 5 vCPUs; workload B needs 12, so a new account must ask first. On Hetzner neither version of workload B fits a new account's default vCPU quota, and raises come after a month and a paid invoice. An IONOS trial is capped at 8 cores and 20 GB of RAM. On Leaseweb's older Frankfurt platform a new customer may run two instances for two months. On Azure, new subscriptions can no longer deploy the B-series v1 and several other older VM families since the July 2026 capacity restrictions, which catches anyone following an old tutorial.

Capacity can run out. Since 26 June 2026 Hetzner has restricted the creation of cloud servers for new customers and for existing customers selected at random, and the restriction was still open on 16 October. IONOS had a GPU capacity shortage with provisioning errors until 21 September 2026. Nebius's contract lets it refuse on-demand service based on available capacity, and after its August 2026 outage in us-central1 reserved capacity came back first and on-demand days later.

Running out of credit can cost you data. UpCloud is prepaid: when the balance runs out all servers are shut down, and its docs say storage is deleted after two months, though its terms allow termination and deletion after "a minimum fourteen (14) days". Aruba Cloud deletes cloud resources 17 days after credit runs out. At Gcore, one failed payment suspends every product and deletes VMs after 48 hours.

The providers, one by one

Each name links to the provider's own article in this series.

Infrastructure clouds

Hetzner (Germany). Bandwidth is close to free (at least 20 TB per EU cloud server, then EUR 1.00 per TB), dedicated and GPU servers sit next to the cloud, and its C5 Type 2 attestation carries a clean auditor's opinion. But the CCX23 went from EUR 23.99 to EUR 85.99 a month in 2026, new cloud servers have been restricted for new and some existing customers since 26 June, and there is no managed database, managed Kubernetes or zones.

OVHcloud (France). Egress included in every European region, a broad catalogue (PostgreSQL, Kafka, OpenSearch, KMS, H100 and H200 GPUs) and three SecNumCloud qualifications. Outside Paris and Milan every region is one zone, Frankfurt has no SLA-backed Kubernetes plan, prices rose twice in 2026, and SLA credits are the sole remedy, with downtime on instances and storage counted from your ticket.

Scaleway (France). The widest AWS-style managed catalogue of the European-controlled providers here, egress included on instances and load balancers, a public price API with 5,655 SKUs, and detailed postmortems. The database standby sits in the same data centre, larger instances need an identity check, and in June 2026 load balancers rose 44% to 46% and IPv4 25%.

IONOS Cloud (Germany). With UpCloud, the cheapest like-for-like production workload here, with the first 2 TB of egress free and 24/7 support included. Compute, storage, IPv4, egress and load balancer prices did not change from October 2024 to October 2026, though Kubernetes node prices changed in April 2025 and database backups have been charged since September 2026. No IONOS page describes its zones as separate buildings, only businesses can sign up, GPU VMs have no SLA, and in May 2026 provisioning was slow or unavailable in every location for about 36 hours.

STACKIT (Germany). At least three physically separate zones per region (its Compute Engine certificate says several may share a building), a legal certificate per service with credits up to 100%, no egress charge and a family-owned German parent. Self-service is limited to companies in Germany, Austria and Switzerland, the Kubernetes control plane costs EUR 72.70 a month, and incidents in 2025 and 2026 crossed its zones.

UpCloud (Finland). A managed cloud (PostgreSQL, Kubernetes, object storage) with 15 zones and one price list for nearly everything (Helsinki charges more for a few storage items), 24/7 engineers on every account, and workload B at 48% of AWS. Each city is one data centre (Helsinki has two), the SLA pays only account credit with claims within 15 days, and its posted ISO 27001 certificate expires on 20 October 2026 with no renewal visible on 16 October.

Exoscale (Switzerland). Managed Kubernetes and eight database engines, IPv4 included, no list price rise in 24 months, and data kept in the chosen country by contract under Swiss law. Each zone is one facility in one city, the control plane is shared (the API and portal failed in all eight zones on 1 September 2026), and the ultimate parent is in Mexico.

T Cloud Public (Germany; Open Telekom Cloud until January 2026). A hyperscaler-shaped catalogue, a synchronous database standby you can place in another zone, and, by the provider's account, zones that held through a four-week fire outage. It costs more than AWS and Azure at list price, its 99.95% is a "target" with no credits in the standard terms, and only businesses in Europe can sign up.

Infomaniak (Switzerland). Free instance traffic, the lowest workload A total in this series (EUR 13.12), Swiss data under a public-benefit foundation, and an ISO 27001 certificate seen. Managed PostgreSQL is still marked "Available soon", data stays in Switzerland rather than the EU, and its zones, by their names, appear to share one data centre.

Gcore (Luxembourg). Free, unlimited egress from VMs, a free highly available Kubernetes control plane, CDN and DDoS protection on the same bill, and public GPU prices. Its owners are not published. Northern Data took two board seats and an option on a majority in April 2025, but after amended arrangements in July 2026 its accounts say it "no longer has significant influence" over Gcore. Managed PostgreSQL is a beta, and for most services SLA credits stop at 20% and downtime counts from your report.

Leaseweb (Netherlands). Bare metal in volume (an 8-core, 64 GB server listed at EUR 118.37 a month on the check date), egress at EUR 2.20 per TB after a pooled 1 TB, and two zones in Amsterdam. Frankfurt and London run an older platform without zones, signup is for businesses only, its terms allow up to five business days of KYC before an order is confirmed, and prices are indexed yearly by up to 5% or CPI if higher.

Aruba Cloud (Italy). Owned Italian data centres with Rated-4 certificates, free VM egress, three zones in Bergamo and CNCF-conformant Kubernetes. No managed PostgreSQL, no VM load balancer on the new platform, every managed service except Object Storage in Bergamo only, and no public status page.

Elastx (Sweden). A Swedish majority owner (Sobro, 53%), data kept in Sweden, three data centres in Stockholm with a multi-zone SLA, and 24x7 support in every price. The most expensive like-for-like option here (171% of AWS), signup is a test-account request, and it documents OpenStack Yoga, a release now unmaintained upstream.

Cleura (Sweden). An EU-only data path written into the contract, the same prices in all three regions including Frankfurt, and Public Cloud egress at a flat EUR 0.02875 per GB (free on its Compliant Cloud). No published SLA document, no zones on the self-service Public Cloud, and no self-service managed database.

Cyso Cloud (Netherlands; Fuga Cloud until 1 January 2025). Cheap compute, instance egress described as free on its platform page (its billing FAQ calls bandwidth a variable cost), two regions with three zones each, and ISO 27001 and NEN 7510 certificates seen. No generally available managed database or GPU, and Amsterdam outages in 2025 crossed its zones. A March 2026 Cyso post on "moving away from OpenStack", mostly about replacing VMware, names CloudStack as the direction, with no timeline and nothing yet for existing public cloud customers.

CloudSigma (Switzerland). Servers sized freely by CPU, RAM and disk, cheap on one-month subscriptions in Geneva under a Swiss contract. Frankfurt and Düsseldorf no longer accept signups, pay-as-you-go costs 3 to 5.5 times the subscription, and ticketing has been broken since 28 September 2026. There are no zones and almost no managed services: object storage in Düsseldorf only, and a Kubernetes service that the web app calls managed but nothing documents.

plusserver (Germany). No egress charge, zones in separate buildings, an HA Kubernetes control plane at EUR 65.70 a month, and a C5 Type 2 report you can download, though with a qualified opinion. Its owners are unnamed former lenders, the SLA and DPA arrive only with the order, and each zone is a separate OpenStack environment, so cross-zone HA is your job.

Evroc (Sweden). No sub-processor for the infrastructure services, B200 at EUR 8.00 per GPU-hour, and three zones in Stockholm in unnamed facilities. No managed database or Kubernetes, egress at EUR 0.10 per GB with no allowance found, and no certificate of any kind seen.

Developer platforms

Clever Cloud (France). Git push deploys in 18 runtimes with managed PostgreSQL, Redis and S3-style storage, its own servers in Paris, six detailed postmortems and SLA penalties paid in euros. No self-serve database failover, failures in one part of the Paris region have repeatedly spread beyond it, and workload B costs more than on AWS.

Upsun (France; formerly Platform.sh). A French contract, a full copy of the stack for every Git branch, and egress at EUR 30 per TB. Six of its seven European regions run on AWS, Google Cloud or Azure, the self-serve product has no database failover, and workload B costs EUR 1,911.83.

Koyeb (France). A fast path from a container to an autoscaling HTTPS service with scale to zero, and per-second GPUs from USD 2.50 for an H100. Part of Mistral AI since spring 2026, its H100-class GPUs are only in North America, and there is no SLA below the USD 299 a month Scale plan.

Northflank (United Kingdom). Services, cron jobs, databases, preview environments and GPUs in one platform, which can also deploy into your own AWS, Google or Azure account. Its managed cloud is Google Cloud, three venture firms hold 52.69% of its shares, and there is no SLA below Enterprise.

GPU clouds

Nebius (Netherlands). H100, H200 and B300 on demand without a sales call, managed Slurm and Kubernetes, free internet egress from VMs (Object Storage egress is billed) and free public IPs, and seven detailed postmortems. GPU prices rose twice in 2026, it documents no zones in its regions, and its contract lets it refuse on-demand service when capacity runs short.

Nscale (United Kingdom). Active GPU sites in Norway, Iceland and Portugal (much of its contracted capacity is in the US, including a leased site in Ward County, Texas), and an OpenAI-compatible model API that anyone with a Google account and a card can top up from USD 5. GPUs need a reserved contract with no public price, and by its own S-1 it is still "in the process of obtaining" ISO 27001 and SOC 2.

Not in the series: Fluidstack

Fluidstack builds large GPU clusters for AI companies under negotiated contracts. Its group parent is still Fluidstack Ltd in London, but since 4 December 2025 it has called New York its headquarters ("The new Midtown Manhattan office will serve as the company's global headquarters"), and the series includes only providers headquartered in Europe. Its website links to no product, price list or signup, and its terms say orders are placed "via a mutually agreed statement of work or purchase order".

Before you move: what to check

1. Run an export test. Copy a disk image, a database dump and a bucket out of the provider and time it. Hetzner's API has no endpoint to download an image, so disks have to be copied out by hand. Ask whether the contract has an EU Data Act switching clause: STACKIT, UpCloud, Exoscale, Gcore, Leaseweb and Aruba Cloud have one, and Exoscale and Clever Cloud can waive egress fees when you leave, though Exoscale's terms also provide for a data recovery fee of up to EUR 0.12 per GB.
2. Price egress and IPv4 on their own. Egress beyond the free allowance costs EUR 80.30 per TB on AWS and EUR 100 on Evroc, and nothing at STACKIT. IPv4 is moving: OVHcloud started billing it separately on Gen3 instances on 1 October 2026, Scaleway raised it 25% in June, and Evroc charges EUR 7.30 a month per address.
3. Read the SLA's claim rules, not its percentage. Who starts the clock, how many days you have, what evidence you must send, the cap, and whether you get money or credit.
4. Check default quotas and verification before you set a migration date. Open the account weeks early, request the quota for your target size, and get it in writing.
5. Find out where the database standby sits. Ask for the location, synchronous or not, and the failover time. IONOS, STACKIT and Nebius do not document all three.
6. Read the provider's own status history. Note how far back it goes: about a week on Hetzner's front page, two weeks at Evroc, 60 days at Infomaniak, no history at all in Leaseweb's public API and no public status page at Aruba Cloud, against UpCloud's back to 2015. Check whether incidents in one zone spread to others.
7. Re-check every price on the day you decide. Prices changed often in 2026. Hetzner's cloud servers rose 30% to 37% on 1 April and, for new orders and rescales, 38% to 192% on 15 June. OVHcloud changed prices in April and October, Scaleway in June, Nebius in June and October and Elastx by 10% in July, and UpCloud announced new H100 and B200 prices from 20 October. Notice periods range from the three days Nebius's contract allows to the three months Elastx gave in practice, though Elastx's terms state none.
8. Check who you actually contract with. The entity, its country and its sub-processors. Koyeb's documents still name a struck-off company. Upsun lists US sub-processors in every region. We found no DPA or sub-processor list for Northflank, whose trust centre's sub-processor page did not render. CloudSigma's Frankfurt terms name a German subsidiary that was merged into enthus in October 2025.

Sources

Checked on 15 and 16 October 2026.

Hetzner

• Price adjustment tables and price API
• Cloud API specification
• Cloud SLA
• Status page and cloud capacity incident of 26 June 2026
• Legal notice and BSI C5 Type 2 attestation

OVHcloud

• Public Cloud prices and October 2026 pricing details
• Specific Conditions for Public Cloud Services, 26 August 2026
• Status hub
• Universal Registration Document 2025
• Cour d'appel de Douai, 24 April 2025

Scaleway

• Instances pricing and price update of 27 April 2026
• SLAs
• Status page and postmortem of 10 March 2025
• iliad Group Annual Financial Report 2025

IONOS Cloud

• EUR price list
• SLA and GTC
• Status page
• IONOS Group Annual Report 2025

STACKIT

• Price API and price list
• Terms of Use v1.3.2
• Status page
• Schwarz Group results, 11 June 2026

UpCloud

• Pricing (archived 7 October 2026) and calculator
• Terms of Service with SLA (archived 2 October 2026)
• Status page
• Tesi portfolio entry

Exoscale

• Pricing and Terms and Conditions
• Instances SLA
• Incident 247488, 1 September 2026
• A1 Group Annual Financial Report 2025

T Cloud Public (formerly Open Telekom Cloud)

• Price API
• Service description, 9 October 2026
• Status events
• Deutsche Telekom Annual Report 2025, shareholders' equity

Infomaniak

• Public Cloud prices
• Legal documents, including the Public Cloud terms and SLA
• Public Cloud status
• Investors page

Gcore

• Cloud pricing and AI pricing
• Legal documents: MSA, SLA, DPA
• Status page
• Northern Data and Gcore partnership, 2 April 2025 and Rumble closes acquisition of Northern Data, and Northern Data interim statements, Q1 2026, filed by Rumble on 28 August 2026

Leaseweb

• Public cloud pricing, Amsterdam and older platform
• Sales Terms and Conditions, 1 August 2026
• Service Level Agreement, 1 September 2026
• GLEIF ultimate parent record

Aruba Cloud

• Pricing
• Terms and Conditions v1.4
• Service Level Agreement v1.3
• Aruba statement to Consob, 16 November 2017

Elastx

• Pricing
• Availability SLA and Support SLA
• Status page
• Sobro portfolio entry

Cleura

• Pricing and price JSON
• Terms of Service for Public Cloud
• Status page
• EQT sale of Iver, 2 July 2021

Cyso Cloud

• Pricing
• Service Level Agreement
• Status page
• From OpenStack to CloudStack, 18 March 2026

CloudSigma

• Geneva price list and Geneva cloud status
• Geneva SLA
• Status page
• Commercial register, CLOUDSIGMA Holding AG

plusserver

• pluscloud open
• General Terms and Conditions v3.1.4
• BSI C5 Type 2 report 2025
• Status page
• PlusHolding GmbH, Re [2024] EWHC 828 (Ch)

Evroc

• Virtual machines and prices
• General Terms and Conditions and Service Terms
• Status page

Clever Cloud

• Pricing and price API
• General Terms and Conditions of Use
• Status page
• Company register entry

Upsun

• Pricing
• Service Specific Terms, 17 September 2026
• Regions
• Status page

Koyeb

• Pricing
• Service Level Agreement and Data Processing Agreement
• Koyeb is joining Mistral AI, 17 February 2026
• BODACC notice B202600692856, 10 April 2026 and BODACC open data

Northflank

• Pricing
• Terms of Service
• Regions API
• Companies House, Northflank Ltd

Nebius

• Prices and compute pricing
• Services Agreement and SLA
• Status page
• Annual report on Form 20-F for 2025

Nscale

• Form S-1, 18 September 2026
• Companies House, Nscale plc
• Service Level Objectives

Fluidstack

• Fluidstack relocates its global headquarters to New York City, 4 December 2025
• Terms and Conditions
• Companies House, Fluidstack Ltd

AWS and Azure

• AWS Price List API and EC2 On-Demand pricing
• AWS Regions and Availability Zones and RDS Multi-AZ deployments
• Amazon Compute SLA, EC2 instance quotas and end of the Developer, Business and Enterprise On-Ramp plans
• Azure Retail Prices API and Azure availability zones
• Azure Database for PostgreSQL high availability and VM retirements and capacity restrictions
• SLA for Microsoft Online Services and Azure support plans
• ECB euro reference rates, 15 October 2026

Sovereign offerings and EU frameworks

• AWS, Opening the AWS European Sovereign Cloud, ESCA, FAQ and pricing
• Microsoft, What is Microsoft Sovereign Cloud?, Data Guardian and How Microsoft is addressing digital sovereignty in Switzerland
• Google Cloud, Delivering a secure, open, and sovereign digital world, 7 February 2026
• S3NS, PREMI3NS SecNumCloud qualification, general availability and products; Thales, sovereign cloud in Germany, 20 May 2026
• Bleu, Qui sommes-nous and J1 of the SecNumCloud qualification; Delos Cloud, homepage, archived 1 October 2026
• ANSSI, catalogue of qualified services, updated 16 October 2026 and SecNumCloud providers in qualification
• French Senate, commission of inquiry on public procurement, hearing of Microsoft France, 10 June 2025
• Assemblée nationale, commission of inquiry on digital dependencies, round table with Microsoft, Google and AWS, 13 May 2026
• European Commission, Commission advances cloud sovereignty through strategic procurement, 17 April 2026, and Cloud Sovereignty Framework v1.2.1
• European Parliamentary Research Service, Cybersecurity Act review: what to expect, 5 January 2026