Guide · 6 min read

Pick a GPU for a model you run yourself

Choose between the RTX 4090, RTX 5090 and H100 plans by VRAM and job type, without invented benchmark scores.

Start from VRAM, not from a brand adjective

The catalog has three dedicated cards: an RTX 4090 with 24 GB, an RTX 5090 with 32 GB, and an H100 80 GB PCIe. The card is yours for the month. It is not a shared slice and not an hourly spot instance.

We don't publish benchmark leaderboards. Fit the model to the memory, then read the framework's own notes for the weights you downloaded.

24 GB — RTX 4090

This is the plan for quantised 7B to 70B models when you are serving, not training a large model from scratch. The machine around it is 16 vCPU, 64 GB RAM and 1 TB NVMe.

List price starts at $420 per month before the term discount. Switzerland is about 12% higher.

32 GB — RTX 5090

The extra memory is room for a larger context, a less aggressive quantisation, or a LoRA next to the weights. The machine is 24 vCPU, 96 GB RAM and 2 TB NVMe. List price starts at $690 per month.

80 GB — H100

Use this when the weights don't fit in 32 GB, or when the job is a heavier training run. The machine is 32 vCPU, 192 GB RAM and 4 TB NVMe. List price starts at $2,700 per month.

It is a poor fit for a two-hour experiment. There is no hourly SKU. Prepay a longer term only if the card will stay busy.

Where it can run

GPU plans are offered in Iceland, Romania, Switzerland and the Netherlands. The order form shows the locations that have a plan. Panama, Moldova and Russia don't.

Put a small agent VPS in the same country if something else calls the model. The uncensored AI page explains what we mean by that phrase: no third-party filter, and the AUP still applies.

Put it into practice

A VPS S plan is enough for most guides.

See VPS plans