Hosting a fine-tuned model as a production API in Europe
Hosting a fine-tuned model as a production API in Europe: dedicated GPU rates at eight providers and the break-even against per-token pricing.
In this article7

Why a fine-tuned model cannot ride a shared endpoint
A team that has fine-tuned its own model and needs it behind a production API can serve it from a Lyceum on-demand GPU VM with full root access, A100 80GB at $1.59/hr or H100 at $2.79/hr, billed per second (read 2026-09-22),[1] and keep the weights on a machine it controls. The same question at seven European providers is answered below from their own published pages, read on 22 and 30 September 2026.
A shared per-token endpoint is a managed API that fronts a fixed catalogue of models. The provider operates the GPU fleet, keeps each catalogue model loaded, and bills per million tokens. A fine-tuned model is absent from every such catalogue: the artefact belongs to the customer, so the provider has no catalogue row to bill against and no mechanism to load the customer's weights onto shared capacity.
- Billing flips from per million tokens to per GPU-hour, so cost becomes a function of reserved time, not traffic.
- The team owns the serving stack: inference engine, scaling, monitoring and updates are its responsibility, not the provider's.
- Capacity is billed whether or not tokens flow, which is the price of exclusivity on the hardware.
Some providers offer a managed dedicated endpoint that removes the serving work while keeping custom weights. Lyceum's Dedicated Inference product is labelled "Coming soon" on its own site as of 14 September 2026,[2] so the self-managed VM is the route available today. The cost comparison that follows therefore prices dedicated GPU capacity, the input a team actually controls.
What dedicated capacity costs
All rates are on-demand unless the page states otherwise.
| Provider | Listed GPU hourly rate (GPU named) | Custom artefact or container documented | Page read |
|---|---|---|---|
| Lyceum | H100 $2.79/hr, on-demand VM | Yes, full root access | 2026-09-22 |
| Exoscale | A30 €1.23/hr, RTX PRO 6000 €2.15/hr (smallest instance sizes) | Yes, Dedicated Inference serves any Hugging Face model on dedicated GPUs | 2026-09-30 |
| Hetzner | – (GEX45 and GEX131 listed; monthly and hourly prices are rendered by script and were not captured) | Yes, full root access | 2026-09-30 |
| IONOS | H200 PCIe €3/hr pay-as-you-go (GPU VM H200 – S) | – | 2026-09-30 |
| Nebius | H100 $3.85/hr, on-demand | Yes, Serverless AI endpoints run a supplied container image | 2026-09-22 (rate), 2026-09-30 (artefact) |
| OVHcloud | H100 $2.99/hr excl. VAT (h100-380 instance) | Yes, AI Deploy runs a Docker image on GPU nodes, billed per minute | 2026-09-30 |
| Scaleway | L4 €0.79/hr (L4-1-24G) | Yes, Managed Inference deploys an own model (e.g. Hugging Face) on dedicated GPUs | 2026-09-22 (rate), 2026-09-30 (artefact) |
| STACKIT | GPU Server-n2.14d.g1-EU01 at €1.74/hr, GPU type not named on the page | – | 2026-09-22 |
A dash means the provider's page lists no value for that cell. IONOS publishes its GPU VM rate on its German price list, where the H200 – S instance costs €3 per hour pay-as-you-go (read 30 September 2026).[6]
The break-even against per-token pricing
The break-even follows from published rates on both sides. A dedicated H100 VM at $2.79/hr billed for a full month breaks even against output tokens at $2.70 per 1M for Qwen3.8-27B on the IONOS AI Model Hub[7] at roughly 754 million output tokens a month: 730 hours × $2.79 is about $2,037, and $2,037 ÷ $2.70 per 1M is about 754 million tokens. Below that volume, per-token billing on a comparable base model costs less; above it, the dedicated hour does. For context on how widely per-token rates vary, see the compiled LLM API pricing statistics.
The same-provider variant avoids cross-provider comparison: a Lyceum H100 VM at $2.79/hr billed for a full month breaks even against its own Qwen3.8-27B output rate of $2.40 per 1M (read 30 September 2026)[15] once monthly output volume is high enough that the per-token bill would exceed the month's GPU cost. One product mode applies per figure: the hourly rates are VM compute and the token rates are per-token inference on a catalogue model, and the two are never combined in one cell. The floor of the per-token market is tracked separately in the token price comparison.
Which European providers accept a custom artefact
Acceptance of a custom artefact is documented unevenly across these providers. Lyceum's VMs give full root access, so any weights and any serving stack run on them, at the H100 VM rate of $2.79/hr.[1] Separately, its serverless GPU execution accepts a Python or Docker artefact billed per second, with the H100 at $3.59/hr on that serverless rate card, a different product mode from the VM (read 2026-09-22)[2]. Exoscale's Dedicated Inference serves any model on Hugging Face, public, gated or private, as an API on dedicated GPUs (read 30 September 2026).[4] Hetzner's dedicated GPU servers give root access with remote management via IPMI, and its page states the servers run in its data centres in Germany and Finland, at Nuremberg, Falkenstein and Helsinki.[16]
OVHcloud's AI Deploy runs a customer's Docker image on GPU nodes billed per minute[11], Nebius's Serverless AI runs a supplied container image as an endpoint[9], and Scaleway's Managed Inference deploys an own model such as a Hugging Face checkpoint on dedicated GPUs (all read 30 September 2026).[13] The pages of STACKIT and IONOS list GPU instances but state no custom-artefact or custom-model upload feature; teams evaluating those two need to confirm the mechanism directly. How ten European providers disclose catalogue, region and pricing terms on shared endpoints is compared in the European serverless inference providers guide, and open-weight availability across ten providers in the open-weight provider comparison.
Choosing a home for your own weights
- The hourly rate at the GPU class the fine-tuned model actually needs, read from the provider's own price page on a stated date.
- Whether that page documents bringing an own artefact or container, since a GPU instance without a documented path to custom weights serves only catalogue models.
- Jurisdiction: where the compute sits legally, which determines which data-protection rules apply to the weights and to the prompts processed through them.
For a team holding its own weights, the decision reduces to those three checks. The hourly rate sets the fixed cost, the artefact policy determines whether the deployment is possible at all without a workaround, and jurisdiction decides what the endpoint can be used for under European data-protection law. A provider that publishes all three verifiably shortens the audit a buyer has to run; one that publishes none of them leaves the verification to the buyer.
Frequently asked questions
Can I serve my own fine-tuned model?
Yes, but not on a shared per-token catalogue endpoint. Those endpoints serve a fixed model list, and custom weights are not in it. The route is dedicated capacity you control: a GPU virtual machine or server with root access, where you load your own weights and run a serving stack of your choice.
What does a dedicated endpoint cost?
Hourly rates read from provider pages on 22 and 30 September 2026: published H100 VM rates ran from $2.79/hr at Lyceum to $3.85/hr, with OVHcloud at $2.99/hr; in euros, a Scaleway L4 instance cost €0.79/hr, an Exoscale A30 €1.23/hr and an IONOS H200 €3/hr. Hetzner's GPU server prices are rendered by script on its page and were not captured, so no figure is stated for Hetzner.
When is dedicated cheaper than per token?
Dedicated becomes cheaper once monthly output volume is high enough that the per-token bill on a comparable base model would exceed a month of dedicated GPU hours. Worked from the published rates above, an H100 VM at $2.79/hr against output tokens at $2.70 per 1M (IONOS AI Model Hub, Qwen3.8-27B) crosses that line at roughly 754 million output tokens a month: 730 hours × $2.79 is about $2,037, and $2,037 ÷ $2.70 per 1M is about 754 million tokens; below it, per-token pricing on a comparable base model is cheaper.
Which EU providers accept a custom model?
Only where the provider's own page documents it. Lyceum's on-demand GPU VMs give full root access, so any weights and serving stack run on them. Exoscale's Dedicated Inference serves any Hugging Face model on dedicated GPUs, OVHcloud's AI Deploy runs a customer's Docker image, Nebius's Serverless AI runs a supplied container image, Scaleway's Managed Inference deploys an own model, and Hetzner dedicated GPU servers give root access and remote management via IPMI. STACKIT and IONOS list GPU instances but their pages state no custom-artefact or custom-model upload feature. Any GPU VM with full root access accepts custom weights in practice.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Training →Where self-hosting an open model beats the API
When self-hosting an open-weight LLM beats a managed API: the break-even formula, a worked H100 example and European per-token and GPU-hour prices.

Where to run Kimi K3 in Europe, and what it costs
Compare published Kimi K3 token prices across European inference providers, including cached rates, and learn what rate cards do not cover.

Serverless vs dedicated inference cost: where the curves cross
Serverless vs dedicated inference cost as a model with named variables, and which European providers sell each mode. Read 26 October 2026.
