Scaleway GPU & Bare Metal

Who owns it

Scaleway SAS is French, based in Paris, and is the cloud arm of the Iliad group controlled by Xavier Niel. Iliad was taken off Euronext Paris in 2021, so the parent is privately held. The brand absorbed Online.net and Dedibox, and Dedibox is still sold as its own dedicated server line. What matters for GPU buyers is that Scaleway owns the machines. The H100 clusters sit in its own DC5 site in the PAR2 region, the Dedibox estate runs out of DC2, DC3, DC5 and AMS1, and the company states plainly that its infrastructure is "100% European" and "free from the US CLOUD Act". There is no resale of another cloud and no third-party host marketplace. The contracting entity for a UK buyer is not stated on the pages read, so assume Scaleway SAS in France until the contract says otherwise.

## What it costs

Prices are in euros, excluding VAT, read on 22 September 2026. There is no sterling price list, so a UK buyer carries the currency movement. One H100 PCIe 80 GB (H100-1-80G, 24 vCPU, 240 GB RAM, 3 TB of scratch NVMe, 10 Gbps) costs EUR 2.8665 an hour, which is EUR 2,092.55 over 730 hours. SXM starts at two GPUs: H100-SXM-2-80G is EUR 6.6198 an hour, or EUR 3.3099 per GPU-hour, and the four-GPU node is EUR 12.771. The eight-GPU node is listed at EUR 25.3308 an hour, EUR 18,491.48 a month, but it does not appear in the API that lists orderable machines, so treat it as priced rather than available. A Blackwell B300 eight-GPU node is EUR 60 an hour. The cheapest GPU is an L4 24 GB at EUR 0.7875 an hour, EUR 574.88 a month. The older RENDER-S with a P100 costs EUR 1.221 an hour, more than an L4, which is worth knowing if you are pricing a render farm.

There is no committed rate, no reserved rate and no spot tier anywhere in the GPU range, so the hourly price is the price. Billing is per minute. Block storage is EUR 0.0949 per GB a month at 5,000 IOPS and EUR 0.1292 at 15,000. Object storage is EUR 0.01606 per GB a month multi-AZ, with the first 75 GB of egress each month free and EUR 0.01 a GB after that. Bare metal starts at EUR 27.99 a month for an x86 Elastic Metal server, or EUR 0.077 an hour with no setup fee and no commitment, and a RISC-V box comes in at EUR 15.99. Monthly Elastic Metal billing adds a commitment fee equal to one month of rental. Dedibox starts at EUR 4.99 a month plus a EUR 4.99 setup fee, and taking 12, 24 or 36 months waives the setup fee and cuts the rate by 5%, 10% and 15%. The Dedibox GPU server with two H100s is EUR 3,499.99 a month, or EUR 2,974.99 on 36 months. A small team running one L4 continuously with 100 GB of block storage pays 730 times EUR 0.7875, which is EUR 574.88, plus EUR 9.49 of storage, so about EUR 584 a month. Scaleway does not publish whether a stopped instance is still billed, or an uptime figure with service credits.

## What you get and what to watch

Signing up is self-serve with a card, but a new account does not get a GPU by default. Quotas start low, verifying your payment method and identity unlocks more, and Scaleway documentation tells customers to "contact our Support team if you want to increase your quotas to more than 1". No turnaround time is published. Stock is the second obstacle and the more serious one: on 22 September 2026 Scaleway's own availability API marked every H100 type "scarce" or "shortage", all four L40S sizes "shortage", and only the L4 and the RENDER-S instance as freely available. A team that needs an H100 this week should check that API before planning around one.

There is no UK region. GPUs run in three zones only: Paris fr-par-1 has L4s, Paris fr-par-2 has the H100 PCIe and SXM machines, the B300 node and RENDER-S, and Warsaw pl-waw-2 has H100 PCIe, the whole L40S line and L4s. Amsterdam has no GPUs at all. Nodes go up to eight GPUs, and above that the AI Supercomputer line (Nabuchodonosor, 1,016 H100 SXM5 on InfiniBand NDR at 400 Gb/s with DDN storage) is sold by sales contact with no published price, minimum term or minimum size. Kubernetes Kapsule gives a free control plane, autoscaling and multi-AZ clusters with a Terraform path. The public API is the standout: it returns specifications, prices and live stock without a login, which is how every number above was checked. There is an official CLI and Terraform provider. Scaleway publishes no MCP server of its own, only community projects, and Slurm is not documented.

Scaleway holds ISO 27001:2022 and the French HDS health-data certification. SecNumCloud is described on its own security page as under assessment rather than held, and SOC 2 is not mentioned. Support is free by ticket 24 hours a day with an eight-hour target on a P1, then EUR 50 a month or 5% of net spend for a two-hour target and a weekday hotline, EUR 250 or 10% for 30 minutes and a 24/7 hotline, and EUR 990 or 20% for 15 minutes with a technical account manager. The percentage matters: a customer spending EUR 5,000 a month on GPUs pays EUR 500 for the middle tier, not EUR 250. No uptime percentage or credit schedule was found on the pages read. Trustpilot is the hardest reading here, at 1.5 out of 5 from 351 reviews, 79% of them one star against 13% five star, with no invited label. The dated complaints run through 2026: support replying with AI-generated tips and outdated documentation (14 September), accounts restricted without explanation and charged early (3 to 19 August), production servers locked over a billing problem while still being billed (12 to 14 September), undisclosed charges for IP addresses and storage on top of the instance price (4 August), and hardware not matching the advertised specification with refunds refused (12 to 29 July).

Not yet tested by Saaskly. Assessed on published pricing, vendor documentation, company filings, the vendor's own status and trust pages, and third-party reviews.