Translate on a rented GPU

Explanation cards comparing an API provider with an open model run on a rented GPU in a region you choose (Apertus as the example): the trade-offs of starting, paying by the hour and waiting, where it is set up (Integrations), and who sees the text.

1. What an API provider gives you

   With an API provider you send your text to its models, run where it chooses. That is the simplest route, and often the best one.

   

2. What a rented GPU adds: an open model in a region you choose

   On a rented GPU you run an open model yourself, in a region you choose. The example here is Apertus, an open model from Switzerland; it is an example, not a recommendation.

   

3. What it costs you

   The trade-offs: you start and stop the server, you pay by the hour while it runs, and it takes minutes to start.

   

4. Open Integrations

   Rented GPUs are set up under Integrations: the servers running, their configurations, and the region they may run in.

   

5. Who sees the text

   Your text goes to the GPU provider’s machine, which runs the AI model. AV masks it as the project’s confidentiality settings say, exactly as for any other provider.

   

A RunPod key and the EU/EEA GPU Region are set, then a configuration for Apertus 8B is created and started from the launch console. The server starts and loads, is used from a project as vLLM Custom (self-hosted) to translate a segment, and is shut down. Exoscale's Zurich zone is described for Switzerland, and a RunPod launch under Switzerland only is refused.

1. Open Integrations, enter your RunPod API Key and Save

   Enter your RunPod API Key and Save. Apertus needs no Hugging Face token: that is only for gated models.

   

2. Set the GPU Region (EU/EEA only)

   The GPU Region decides where AV may start a server. EU/EEA only is recommended if your data must stay in the EU/EEA. RunPod has no datacenter in Switzerland.

   

3. Click Create New Configuration, AV starts it, template Apertus 8B

   Create New Configuration, AV starts it, then the template Apertus 8B.

   

4. Choose Remote / Cloud and the provider RunPod

   Choose Remote / Cloud and the provider: RunPod.

   

5. Choose the hardware tier and Save

   AV offers the GPUs the AI model fits on, here the Ampere tier, with RunPod’s price per GPU-hour. Save.

   

6. Click the rocket to open the launch console

   The rocket opens the launch console: the datacenters your region allows, the GPUs in stock there, and the price per hour.

   

7. Click Start (billing starts)

   Start: from here it is billed by the hour. RunPod places it in one of the datacenters, here in Sweden, and the console closes itself.

   

8. Watch Running services during the cold start

   The machine starts first. Running services follows it; it usually takes a few minutes.

   

9. Wait while the AI model loads (Starting)

   Then the AI model loads. The row stays Starting until the AI model answers.

   

10. Running: datacenter, run time and cost so far

    Running: the row gives the datacenter, the time it has run and the cost so far, and what to choose in a project: vLLM → Custom (self-hosted).

    

11. In the project's LLM Configuration, choose vLLM, Custom (self-hosted)

    In the project’s LLM Configuration, set the role to vLLM and Custom (self-hosted). AV fills in the server’s address, key and model when it sends.

    

12. Translate a segment (Ctrl+Shift+M)

    Translate a segment as usual (Ctrl+Shift+M): it goes to the rented server. The translation was recorded on 21 October 2026.

    

13. Click Shutdown in Running services

    When you are done, Shutdown in Running services. AV asks, then deletes the server on your account.

    

14. Stopped: billing stops, the configuration stays

    The server is deleted and billing stops. The configuration stays, ready for next time.

    

15. For Switzerland: Exoscale's Zurich zone

    For a GPU in Switzerland itself, AV also offers Exoscale, whose Dedicated Inference has a zone in Zurich. Exoscale’s Zurich zone may need approval from Exoscale, and capacity can be fully booked.

    

16. Under Switzerland only, a RunPod launch is refused

    Under Switzerland only, a RunPod launch is refused before anything is created: RunPod has no datacenter there.

    

17. In short

    Video recorded on 5 November 2026 with AV 0.2.11; the launch screens, stock and prices on 24 October 2026, and they change daily. For Switzerland itself, Exoscale’s Zurich zone is the route.