Private AI/vs Scaleway
Comparison

Private LLM hosting vs Scaleway

By Samuel Seidel · Published September 9, 2026 · 8 min read

I run GPUwerk, which rents DGX Sparks, so weigh everything below with that in mind. Scaleway is a French cloud provider, so like GPUwerk it operates from the EU. Its GPU Instances run in Paris and Warsaw, from an NVIDIA L4 at €0.79 an hour up to H100 and B300 machines, inside a full cloud catalogue with object storage, managed Kubernetes and databases. GPUwerk sells one thing: a dedicated DGX Spark with 128 GB of unified memory, billed by the minute. Pick Scaleway for platform breadth and GPU choice, a Spark for one large-memory machine with nothing else to configure.

Side by side

Scaleway GPU InstancesDedicated DGX Spark (GPUwerk)
What it is A GPU compute product inside Scaleway's broader cloud platform, alongside storage, managed databases, Kubernetes, and more. A single-purpose GPU host: one dedicated DGX Spark, rented by the hour, nothing else attached. (advantage)
Region Paris: L4 instances in PAR-1 and PAR-2; L40S, H100 and B300 instances in PAR-2. Warsaw (WAW-2): L4, L40S and H100 PCIe, often in short supply. Per Scaleway's GPU pricing page and its public instance API. Own Sparks in Hradec Králové, Czechia; partner Sparks in other EEA countries, the US and Hong Kong. The console offers one Europe (EEA) location.
Hardware choice NVIDIA L4 (24 GB), L40S, H100 PCIe and SXM, and B300 SXM, in 1 to 8 GPU configurations. One machine: the DGX Spark, 128GB unified memory, with an optional two-node cluster for larger workloads.
Who can see your data Governed by Scaleway's own terms and data processing addendum; check their legal pages for the current text. No other customer, and GPUwerk only under the DPA. GPUwerk's DPA states GPUwerk "hosts the machine but does not access, read, copy, index, or analyse the content of the controller's instance." (advantage)
Model choice Whatever you deploy, unrestricted by the provider. Larger GPU instance types can run bigger models or serve higher concurrency than a single Spark. Whatever you deploy too, but bounded by 128GB unified memory: gpt-oss-120b, Llama 3.3 70B, Qwen3-Coder 30B, DeepSeek, and similar fit; models needing substantially more memory don't, without the two-node cluster.
Speed and concurrency Depends on the GPU type and count rented; Scaleway publishes instance specs but tok/s for a given model is yours to benchmark. A single Spark, single-stream: gpt-oss-120b decodes at 33.5 tok/s and reaches 862.8 tok/s aggregate at 256 concurrent requests (Dendro Logic's concurrency benchmark, run on their own Spark). Raw logs and methodology on GPUwerk's benchmarks page. (advantage)
Pricing model Per hour, by instance type. On Scaleway's pricing page (checked 9 November 2026), L4-1-24G is €0.79/hour, about €575 a month; larger types cost more. Per hour, billed per minute: $0.79/hour on-demand, $0.59/hour while stopped and held for you. One rate regardless of workload, per pricing. (advantage)
Worked cost example See the full arithmetic below the table.
Contracts and DPA Governed by Scaleway's own customer agreement and data processing terms; both companies are EU entities, so check the specifics against your own requirements. A standard GDPR Article 28 DPA published free at /legal/dpa, no negotiation required. Sub-processor list at /legal/sub-processors states none are engaged for instance workloads. (advantage)
What you operate yourself Everything on the instance: OS, model server, monitoring, backups. Scaleway manages the physical hardware and network underneath, and offers additional managed products if you want to build on top of them. The same: root access, from the browser terminal or over SSH with your own key, to your own container, the model server, any RAG or agent layer, monitoring, and backups. GPUwerk keeps no copy of /workspace, not even for a failed drive, so that responsibility is entirely yours.

The worked cost example

Take a workload that generates 500 million output tokens a month, roughly what a busy internal support bot or a document-processing pipeline running continuously might produce. Two ways to serve it:

Scaleway. GPU Instance pricing is per hour and varies by card. A single L4 (24 GB) costs €0.79/hour but cannot hold gpt-oss-120b; for that model you need an H100-class instance in PAR-2. Take its hourly rate from Scaleway's pricing page, measure your model's tokens per second on it, and work out the hours for 500 million tokens.

Dedicated Spark. gpt-oss-120b reaches 862.8 tok/s aggregate at 256 concurrent requests, per Dendro Logic's concurrency benchmark on a single Spark (cited on GPUwerk's benchmarks page). That is 862.8 × 3,600 = 3,106,080 tokens per hour if the node stays saturated at that concurrency the whole time. 500,000,000 ÷ 3,106,080 ≈ 161 hours. At $0.79/hour, 161 × $0.79 = $127.19, before tax.

That Spark figure holds only if the node is busy at 256 concurrent requests for all 161 hours. A larger Scaleway instance can clear the same 500 million tokens in fewer hours at a higher hourly rate, so compare cost per token, not the sticker price. Run both with your own model and traffic pattern.

Migration path

Both are GPU machines you deploy your own stack on. Run vLLM on either and you get the same OpenAI-compatible API, because the endpoint comes from vLLM, not from the hosting provider. Put LiteLLM in front of either for request logging and key management. What changes is the hardware underneath: x86 and discrete GPUs on Scaleway, ARM64 and 128 GB of unified memory on a Spark. A model tuned for a large Scaleway instance may need re-quantizing to fit a single Spark, and your container images need an arm64 build.

When Scaleway is the right choice

When a dedicated Spark is the right choice

FAQ

Is Scaleway a good alternative to GPUwerk?

It depends what you're optimizing for. Both are EU-headquartered infrastructure providers (Scaleway is French; GPUwerk is Czech), so data residency concerns are similar in kind. Scaleway is a full general-purpose cloud, and its GPU Instances in Paris and Warsaw are one product line among object storage, Kubernetes and managed databases. GPUwerk sells one thing: a dedicated DGX Spark by the hour. If you need only GPU compute for LLM inference and want to avoid picking an instance type, GPUwerk is simpler. If you need the GPU alongside other Scaleway infrastructure you already run, staying on one platform may be worth more than the simplicity.

Does Scaleway offer dedicated hardware or is it shared/virtualized?

Scaleway sells GPU Instances in 1 to 8 GPU configurations; check its GPU Instances documentation for what a given type shares with other tenants. A GPUwerk DGX Spark is a whole physical machine for the duration of your rental: no other tenant runs on it while it's yours.

Is a DGX Spark cheaper than Scaleway GPU Instances?

It depends on the GPU type you'd compare it against. In the worked example on this page, 500 million generated tokens a month costs about $127 in Spark compute at $0.79/hour, assuming the node sustains 256 concurrent requests the entire time. On Scaleway's pricing page, checked 9 November 2026, the smallest GPU Instance (one L4, 24 GB) is €0.79/hour; instances big enough for gpt-oss-120b cost more. Pull the current rate for the instance you would actually run before comparing.

Why pick a single dedicated Spark over a cloud GPU instance from a general-purpose provider?

Mainly simplicity. Scaleway's catalogue spans many products and GPU instance types, which is useful if you need that range, but it also means comparing specs and prices before you can start. A Spark is one machine at one rate, configured by GPUwerk, so there is less to set up before you are running a model.

Scaleway claims checked on 9 November 2026 against scaleway.com/en/pricing/gpu (instance types, GPUs, Paris zones, L4 price). Prices change; check the instance you would run. Spark throughput figures come from Dendro Logic's concurrency benchmark, detailed on GPUwerk's benchmarks page.

Related pages

First top-up: pay $20, get $30 in credit

Run your own numbers before you commit either way.

Deploy a dedicated DGX Spark and test your actual prompts against an open model before comparing quotes.

Deploy now Read the benchmarks