SCALEWAY · TECH
Scaleway: French hoster from Paris, ISO 27001, strong API, NVIDIA H100 AI cluster
Scaleway operates data centres in Paris, Amsterdam and Warsaw. ISO 27001, S3-compatible object storage, H100 cluster for AI training with transparent hourly pricing.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What is Scaleway?
Scaleway is a French cloud platform headquartered in Paris, originally founded in 1999 as a hosting provider under the name Online.net (Iliad group). In 2015 Scaleway was positioned as a standalone cloud brand; since the 2020 rebranding it is the cloud division of the Iliad group (Free, Free Mobile, Free Pro). As of May 2026 Scaleway operates seven data centres across three EU regions: Paris (fr-par-1, fr-par-2, fr-par-3), Amsterdam (nl-ams-1, nl-ams-2) and Warsaw (pl-waw-1, pl-waw-2). Own data centre DC5 in Saint-Ouen-l-Aumone has been running with adiabatic cooling since 2023 – one of the most energy-efficient facilities in Europe.
The product portfolio targets modern cloud architectures. Instances (virtual cloud servers) from EUR 0.0086/hour for Stardust (1 vCPU, 1 GB RAM, local 10 GB SSD), up to high-RAM and CPU configurations. Apple Silicon (M2 Mac mini) as bare-metal server from EUR 96/month – interesting for iOS builds and macOS CI/CD. Elastic Metal as dedicated servers with hourly billing logic. Object storage (S3-API compatible) from EUR 0.0144/GB/month for multi-AZ standard. Managed Kubernetes (Kapsule) from EUR 8/month for the control-plane fee. Managed databases (PostgreSQL, MySQL, Redis). Serverless Containers, Serverless Functions, Serverless SQL Databases – for event-driven workloads.
As of May 2026 the AI portfolio of Scaleway is particularly actively built out. NVIDIA H100 PCIe cluster with InfiniBand networking in Paris, from about EUR 2.73/hour per H100 in the on-demand variant. NVIDIA L4 from EUR 0.65/hour. NVIDIA L40S for inference from EUR 1.20/hour. Generative APIs as managed inference service for open-weight models (Llama 3.3, Mistral, Qwen) with pay-per-token. Embeddings API with French-optimised models. Vector database (managed Qdrant). This bundle positions Scaleway as an EU alternative to Together AI or Replicate, with clear EU data residency.
The platform is rigorously API-first. An official CLI, a Terraform provider, a Python SDK, a Go SDK and a REST API with OpenAPI spec. New instance provisioning in May 2026 is among the fastest in the EU – typically under 30 seconds.
Why it matters
Three points make Scaleway interesting for the Swiss and EU market: extremely fast provisioning, transparent GPU hourly pricing, and a modern managed AI offering in the EU.
Provisioning speed: as of May 2026 Scaleway is the fastest EU cloud at spinning up new instances. A standard VM is up in 20-30 seconds, a Kubernetes cluster in 3-5 minutes, a managed Postgres instance in 5-10 minutes. For test-driven development and CI/CD pipelines that is a clear productivity advantage. AWS and Azure typically need 2-3x longer for comparable operations.
GPU hourly pricing: H100 from EUR 2.73/hour is in May 2026 one of the lowest hourly prices for an H100 in the EU with an ISO 27001 hoster. For a 24-hour finetune on an H100 you pay around EUR 65 – comparable to RunPod Secure Cloud, but with guaranteed EU residency and transparent ISO certification.
Generative APIs: Scaleway hosts open-weight models (Llama 3.3 70B, Llama 3.3 405B, Qwen 2.5 72B, Mistral Large) on its own hardware in Paris. The inference API is OpenAI schema compatible and can be deployed as a drop-in replacement for OpenAI calls without code changes – with pay-per-token billing instead of pay-per-hour. For workloads without constant load that is a marked cost advantage.
Regulatory position: ISO 27001 for all three regions, DPA per GDPR Art. 28 available as standard. No US parent – the Iliad group is a French public company, no US group, no CLOUD Act risk. For EU mandates the compliance argument is unambiguous. For mandates with strict Swiss residency Scaleway does not fit, but for EU standard with Swiss clientele the package is attractive.
Price-wise Scaleway sits between Hetzner and OVHcloud. A comparable VM is about 1.5x more expensive than Hetzner but easier to automate with a modern API. Compared to OVHcloud, Scaleway is at the same price level, with an edge on API comfort and provisioning speed.
How it works
Ordering: through the portal console.scaleway.com. Account creation with European VAT logic, payment by credit card or SEPA direct debit. Instance provisioning typically in 20-30 seconds, Kubernetes cluster (Kapsule) in 3-5 minutes, managed PostgreSQL in 5-10 minutes.
Sample prices May 2026: Stardust1-S (1 vCPU, 1 GB RAM, 10 GB SSD): EUR 0.0086/hour, equivalent to EUR 6.20/month at constant runtime. PRO2-S (8 vCPU, 32 GB RAM, 80 GB SSD): EUR 0.117/hour, EUR 84/month. GP1-XS Elastic Metal bare-metal (Xeon Silver, 96 GB RAM): EUR 254/month. NVIDIA L4 24GB GPU: EUR 0.65/hour, around EUR 470/month. NVIDIA H100 PCIe: EUR 2.73/hour, around EUR 1,970/month. Multi-AZ object storage: EUR 0.0144/GB/month. Kapsule control plane: EUR 8/month plus node cost. Managed PostgreSQL Hobbyist: EUR 30/month.
Generative APIs May 2026: Llama 3.3 70B inference from EUR 0.84/million input tokens, EUR 0.84/million output tokens. Qwen 2.5 72B similar. Mistral Large 2 from EUR 2/million input. Embeddings (bge-multilingual-gemma2) from EUR 0.05/million tokens. Managed vector DB (Qdrant) from EUR 49/month.
Network: every instance gets a public IPv4 and IPv6. Private networks (VPC) as a standard feature, free between instances in the same region. Load balancer from EUR 12/month. Anti-DDoS protection at the infrastructure level included.
Storage: local SSD on every instance (NVMe on most tariffs), plus separate block storage as persistent volume mount from EUR 0.10/GB/month. Object storage One-Zone (cheaper variant) from EUR 0.006/GB/month. Cold storage from EUR 0.002/GB/month. Backup-as-a-Service with automatic snapshot management.
Contract details: all contracts billed monthly with month-end cancellation, hourly billing as default. DPA per GDPR Art. 28 generated in the console. No US parent, clear EU-law position. Iliad group as French SA.
Migration: a typical migration path from AWS or Azure to Scaleway takes 2-5 days. Terraform scripts can be re-mapped directly, image import supported. Generative API migration from OpenAI is a one-line endpoint swap in the SDK in May 2026.
Scaleway setup for an EU AI stack in 5 steps
- 01Create an account at console.scaleway.com, register company details, generate the DPA per GDPR Art. 28, set up SEPA or credit card payment.
- 02Choose region: fr-par-1 for Paris (H100 GPU availability), nl-ams-1 for Amsterdam (alternative EU region), pl-waw-1 for Warsaw (multi-region DR).
- 03Set up a Terraform project, define instances, Kapsule cluster and managed Postgres as code; provisioning takes under 30 seconds per instance.
- 04Configure the generative API as drop-in for OpenAI: api.scaleway.com/llm-inference/v1/chat/completions as endpoint, OpenAI schema compatible.
- 05Roll out the AI stack: Llama 3.3 70B via Generative APIs, plus managed Qdrant vector DB, plus object storage for RAG sources, plus Kapsule cluster for n8n workflows.
When to use Scaleway
Scaleway is the right choice when (a) modern API-first architecture with fast provisioning is important, (b) generative AI workloads should run via an EU API, or (c) H100 GPUs in the EU at transparent hourly prices are coming up. Concrete cases: Swiss startup with RAG pipeline and CI/CD drift tests. EU mid-market planning to switch from OpenAI to an EU inference API. AI team needing temporary H100 hours for finetuning.
For serverless architectures Scaleway is especially interesting. Serverless Containers can be used like Cloud Run or Fargate, with pay-per-compute-second. Serverless Functions as FaaS for event-driven pipelines (n8n webhooks, Stripe events, etc.). Serverless SQL Databases scale automatically with load.
For Kubernetes workloads Kapsule is a cost-efficient choice. Control plane from EUR 8/month, worker nodes as standard instances. Cilium as CNI, automatic upgrades, integrated monitoring. For GitOps with ArgoCD or Flux, Scaleway is well positioned.
For EU multi-region setups (Paris-Amsterdam-Warsaw) the geographic distribution fits disaster recovery or regional latency optimisation. Egress between Scaleway regions is EUR 0.01/GB – predictable, no egress shock.
When not to use
Anyone needing strict Swiss residency does not get there with Scaleway – no region in Switzerland, only Paris, Amsterdam, Warsaw. For Swiss mandates Infomaniak, Exoscale or Safe Swiss Cloud is cleaner.
Anyone seeking strictest budget discipline is better served by Hetzner. Scaleway is about factor 1.5x more expensive than Hetzner at comparable configuration. The API depth and AI portfolio justify the premium only for mandates that actually need this functionality.
For maximum multi-GPU clusters (64+ H100 with full InfiniBand topology) CoreWeave or Lambda Labs is more established. Scaleway has H100 clusters with InfiniBand but not in the league of hyperscaler training.
General caveat: Scaleway is a growing provider with occasional service constraints – as of May 2026 H100 slots are occasionally on a waitlist, especially in peak periods. Anyone needing a guaranteed H100 24/7 should consider a reserved contract or keep OVHcloud in parallel as a backup option.
Mail hosting and classic web hosting are not the Scaleway profile. Anyone looking for a mailbox plus static web hosting is better served by Infomaniak or OVH. Scaleway is focused on compute, storage, Kubernetes and AI.
Trade-offs
STRENGTHS
- Fastest EU cloud for provisioning – VM in 20-30 seconds, cluster in 3-5 minutes
- Transparent GPU hourly prices: H100 from EUR 2.73/h, L40S from EUR 1.20/h, L4 from EUR 0.65/h
- Generative APIs as OpenAI drop-in with EU residency (Llama, Qwen, Mistral in Paris)
- No US parent – Iliad group is a French public company
WEAKNESSES
- No CH region – for strict Swiss residency, Infomaniak, Exoscale or Safe Swiss Cloud
- Factor 1.5x more expensive than Hetzner at comparable configuration
- H100 availability occasionally on waitlist, reserved contract recommended
- Mail hosting and classic web hosting not in the Scaleway profile
FAQ
Is Scaleway nFADP-compliant for Swiss mandates?
Yes, for EU regions without third-country transfer issues. France, Netherlands and Poland are recognised in the nFADP annex as countries with adequate data protection. The Scaleway DPA per GDPR Art. 28 is solid under nFADP standard. For mandates with strict Swiss residency (professional secrecy, cantonal rule) Scaleway does not fit – then Infomaniak or Exoscale. For EU-standard mandates Scaleway is a mature choice.
What does a 24-hour H100 finetune cost?
An H100 PCIe instance at Scaleway costs EUR 2.73/hour in May 2026. A 24-hour job therefore lies at EUR 65.50. Comparison: at RunPod Secure Cloud EU-Sweden USD 2.59/hour, around EUR 56 for 24 hours – slightly cheaper but with Akamai Sweden location. At Lambda Labs USD 2.49 for 24h, around EUR 53 – cheaper but US-hosted without EU residency. At OVHcloud H100 from EUR 4.50, around EUR 108 for 24h – more expensive but SecNumCloud-certifiable. Scaleway is in May 2026 the cheapest H100 with clear EU residency and ISO 27001.
How does the Scaleway Generative API work?
Scaleway hosts open-weight models (Llama 3.3 70B, Qwen 2.5 72B, Mistral Large 2) on its own H100 hardware in Paris. The API is OpenAI chat completions schema compatible – an existing OpenAI integration can be redirected to Scaleway with a single endpoint swap. Billing per million tokens (input and output separately). Pro: EU residency, no third-country transfer discussion, pay-per-token instead of pay-per-hour. Con: the model is not GPT-4o level on every task – for coding and mathematics, proprietary models stay better.
What happens with H100 waitlists?
In May 2026 H100 slots in Paris (fr-par-1) are occasionally on a waitlist during peak periods – typically 1-3 days. Workarounds: (a) reserved contract for guaranteed availability, (b) switch to L40S for inference workloads, (c) multi-GPU job on 4x L4 for smaller models. Generative APIs are usually immediately available because Scaleway maintains dedicated inference hardware for them. For 7/24 production workloads, keep OVHcloud H100 in parallel as a fallback.