# New RTX 5090 and RTX 5080 servers in Switzerland
18 November 2026 · Fleet · SwissGPU
In October we said more GPU capacity was coming. The new machines have arrived: additional NVIDIA RTX 5090 and RTX 5080 servers have joined the SwissGPU fleet. They are hosted in Switzerland and rented as dedicated bare-metal machines.

## At a glance

- New capacity: NVIDIA RTX 5090 (32 GB VRAM) and RTX 5080 (16 GB VRAM), both on the Blackwell architecture.

- Dedicated bare metal: the GPU, CPU, memory and NVMe storage are yours alone for the whole rental.

- Hosted in Switzerland, on hardware that SwissGPU owns and operates.

- On-demand prices on 18 November 2026: RTX 5090 at CHF 0.50 per hour and RTX 5080 at CHF 0.38 per hour, metered by the second.

- Ubuntu with SSH access. You install the drivers, frameworks and models you need.

## What the new servers offer
 |  | GPU | GPU memory | CPU | RAM | Storage | On-demand price
 | [NVIDIA RTX 5090](https://www.swissgpu.ch/gpus/rtx-5090) | 32 GB | 16 cores | 32–64 GB | 1–2 TB NVMe | CHF 0.50/hour
 | [NVIDIA RTX 5080](https://www.swissgpu.ch/gpus/rtx-5080) | 16 GB | 16 cores | 32 GB | 1 TB NVMe | CHF 0.38/hour

## Why bare metal
Most GPU clouds rent you a slice of a shared host. A SwissGPU rental is the whole machine. No other tenant runs on your server, and there is no virtualization layer between your workload and the hardware. You have root access and choose the driver, CUDA version and frameworks yourself.
The servers, and the team that runs them, are in Switzerland. If you need to know where your workload runs, the answer is short.

## What to run on them
GPU memory usually decides which model fits. With 32 GB, the RTX 5090 suits larger image-generation workflows and single-GPU model inference. The RTX 5080 covers development, smaller image-generation workflows and inference that fits in 16 GB. Check that your framework build supports Blackwell GPUs before you start.

- [ComfyUI](https://www.swissgpu.ch/guides/comfyui) for image generation with your own checkpoints and custom nodes.

- [Ollama](https://www.swissgpu.ch/guides/ollama) for LLM inference with an API you reach through SSH.

- [PyTorch](https://www.swissgpu.ch/guides/pytorch) for training runs and experiments on a full Linux machine.

- [Blender](https://www.swissgpu.ch/guides/blender) for GPU rendering with CUDA or OptiX.
Need more memory? The [GPU catalogue](https://www.swissgpu.ch/gpus) also lists the RTX 6000 Ada with 48 GB and the [RTX PRO 6000 Blackwell Workstation Edition](https://www.swissgpu.ch/gpus/rtx-pro-6000-blackwell) with 96 GB.

## How to rent one

- Choose a model in the [GPU catalogue](https://www.swissgpu.ch/gpus) and sign in.

- Add account credit, select the Ubuntu image and set a server password.

- Allow around 20 to 30 minutes for installation. Usage billing starts when the server is ready, not before.

- Connect over SSH, run your workload, back up your results and release the server when you finish.
Prefer automation? You can create the same rental through the [customer API](https://www.swissgpu.ch/news/gpu-rental-api), and read current prices as JSON from the [public pricing API](https://www.swissgpu.ch/news/gpu-pricing-api).
Capacity is limited, and a rental is confirmed when your request is accepted. Planning a longer project? Fixed-term prepaid plans lower the compute rate where offered; see the [pricing page](https://www.swissgpu.ch/pricing) or write to [info@swissgpu.ch](mailto:info@swissgpu.ch).

## What comes next
Virtual machines and reusable templates are on our [roadmap](https://www.swissgpu.ch/roadmap). We will post fleet and product updates here as they ship.