[ Dedicated GPUs ]
Dedicated GPUs
Run an open model or your own model on GPUs reserved for you in the EU, behind an OpenAI-compatible endpoint.
With dedicated GPUs, a model runs on hardware that is reserved for you in an EU data center. No other customer's traffic shares it. You pay for the capacity you reserve, not per token.
Dedicated GPUs are set up with you as a project. There is no self-service yet: the Inference page in the console is a preview. To start, book a call.
When it pays off
| Gateway | Dedicated GPUs | |
|---|---|---|
| Billing | Per token | Per reserved capacity |
| Models | Every model in the catalog | Any open model, or one trained for you |
| Traffic | Shared with other customers of the provider | Only yours |
| Setup | None, a key is enough | A project with Kontinent |
| Good for | Getting started, variable or low volume, many models | High steady volume, stable latency, a model that is not in the catalog |
Most teams start on the gateway. When volume on one model grows to the point where reserved GPUs cost less than tokens, or when latency has to stay stable under load, dedicated GPUs are the next step.
What you get
- An OpenAI-compatible endpoint. The same request and response shape as the gateway, so your code does not change beyond the base URL, key and model id you receive at handover.
- A region you choose. The deployment runs in the EU region agreed in the offer and stays there.
- A configuration for your workload. Context length, batch size and parallelism are set for your target latency and throughput. Interactive traffic and batch jobs can get separate deployments of the same model.
- No data retention by default. Like the gateway, Kontinent does not store prompts or completions. Anything else, such as collecting traces for later training, happens only if you ask for it in writing.