What self-hosting AI costs a company in 2026
Control has a price, but it is often lower than buyers expect, and the most expensive option is rarely the most sovereign. Worked examples from published European prices.
Contents
Self-hosting is the most visible form of AI sovereignty, and the most expensive one to run. Published European prices show where the money goes, and why owning the system often matters more than owning the GPU.
The question behind the search
A company that searches for "self-hosted AI" usually wants one of three things: to keep data away from other parties, to stop depending on one vendor, or to avoid rising per-seat fees. Each can be met at very different costs. The figures below come from providers' own price pages, read on the date shown; they exclude VAT and change often.
Four levels of control, four price models
| Level | What the company controls | Price model | Published example |
|---|---|---|---|
| EU model API | Which model; data stays with an EU provider | Per token | STACKIT gpt-oss 20b: €0.15 per million input, €0.25 per million output tokens1 |
| Rented GPU instance | The model and its runtime, on a provider's cloud | Per hour | OVHcloud H100 from $2.99 an hour, L40S from $1.802 |
| Dedicated GPU server | The whole server, on a provider's premises | Per month plus setup | Hetzner GEX63 (96 GB GPU) €999 a month plus €999 setup3 |
| Company-owned system | The system, data and choice of model and host | Hosting plus usage | OpenLazarus managed hosting from $150 a month, movable to your own cloud4 |
Worked examples
Take a moderate business workload of 50 million input tokens and 10 million output tokens a month. The volume is an illustration, not a benchmark; measure your own before deciding.
- STACKIT, gpt-oss 20b: 50 × €0.15 + 10 × €0.25 = €10 a month1.
- STACKIT, Qwen3 VL 235B: 50 × €1.50 + 10 × €1.75 = €92.50 a month1.
- Mistral Large via Mistral's API: 50 × $0.5 + 10 × $1.5 = $40 a month, with batch processing at half price5.
- OVHcloud H100 instance, running continuously: $2.99 × 730 hours = about $2,183 a month2.
- Hetzner GEX63, dedicated: €999 a month plus €999 setup, or €12,987 in the first year3.
- Hetzner GEX45, dedicated (24 GB GPU): €214 a month plus €209 setup, or €2,777 in the first year3.
At the blended price of the larger STACKIT model in this example (about €1.54 per million tokens), the €999 monthly cost of the GEX63 would buy roughly 648 million tokens of API use. Dedicated hardware pays off on cost only at sustained, high volume. It is usually chosen for control, not for savings.
What the price pages leave out
- People. Someone must install the model runtime, patch the operating system, monitor capacity and update models. None of the hardware prices above include that time.
- Idle capacity. A dedicated GPU costs the same at three in the morning as at noon. APIs bill only for use.
- Model updates. Open-weight models improve quickly. A self-hosted model needs to be replaced and re-tested; an API or a router can switch with a configuration change.
- The system around the model. A model alone does no business work. The software that connects it to the company's records, permissions and processes is a separate cost, whether built in-house or bought.
Where ownership is cheaper than hardware
The most durable form of control is often owning the system rather than the GPU. A company that owns its system and data can call an EU model API today, move to a dedicated server later and switch models as prices fall. OpenLazarus is one example: the code and data belong to the company, European companies can run on servers in the EU, and any model can be used through its AI router6. Its published prices are plans at $100 or $300 a month including the same amount in AI credits, managed hosting from $150 a month and optional engineering support from $4,500 a month6.
A path to the right level of spend
- Measure token volume for one real workload for a month before buying hardware.
- Start on an EU API with a written no-training commitment, such as those published by STACKIT and OVHcloud7.
- Own the system layer so that the model and host can change without rebuilding.
- Move to dedicated hardware only for data that may not leave infrastructure you control, or when steady volume makes it cheaper.
- Recalculate every six months, because token and GPU prices both fall.
The result is a cost that follows use, a data path the company can explain, and the freedom to move. That combination, more than any single server, is what most companies mean when they ask for sovereign AI. See our main guide for how the providers compare.
Questions buyers ask
How much does it cost to self-host an AI model?
On published European prices, a dedicated server with a 24 GB GPU starts at €214 a month at Hetzner and one with a 96 GB GPU at €999 a month, plus a one-time setup fee. Renting an H100 GPU instance at OVHcloud starts at $2.99 an hour, about $2,183 for a month of continuous use.
Is self-hosting AI cheaper than using an API?
Usually not, at the volumes most companies have. At STACKIT's published price for its largest listed chat model, the €999 monthly cost of a 96 GB GPU server buys roughly 648 million tokens of API use. Self-hosting pays off on cost only at very high, steady volume; it is chosen more often for control.
What hidden costs does self-hosting AI have?
Staff time for setup, security patching, model updates, monitoring and capacity planning, plus the cost of idle hardware. None of these appear on the providers' price pages.
Sources
- STACKIT AI Model Serving. STACKIT (Schwarz Digits). Accessed 8 Nov 2026.
- Cloud GPU instances. OVHcloud. Accessed 8 Nov 2026.
- Dedicated GPU servers. Hetzner. Accessed 8 Nov 2026.
- Pricing. OpenLazarus. Accessed 8 Nov 2026.
- Pricing. Mistral AI. Accessed 8 Nov 2026.
- Frequently asked questions. OpenLazarus. Accessed 8 Nov 2026.
- AI Endpoints. OVHcloud. Accessed 8 Nov 2026.
Updated 8 Nov 2026: First published with prices read on 8 November 2026.