H200 141 GB
From$4.39per hour, excluding tax
Since you opened this page, an idle H200 has cost
At the rental price of an H200: $4.39 per hour, excluding tax. The counter runs for as long as the page stays open.
Your GPUs are expensive. How many hours do they sit idle?
We rent out GPUs, train your open-source AI models and tune your cards so that they work. Your data stays in Europe.
Book a GPU assessmentFor teams that run AI.
- AI startups
- Dedicated cards to train and serve your models.
- Software vendors
- AI inside your product, served through a standard API.
- IT services firms
- Capacity and a GPU team for your clients’ projects.
- Research labs
- Compute by the hour or by the month for your experiments.
- Studios
- Cards for your renders and image generation.
- Small and mid-sized businesses
- Your own model, without your data leaving Europe.
Rent GPUs.
Dedicated bare-metal cards, by the hour or by the month.
By the hour or by the month
A 3, 6 or 12-month commitment gets a better price.
Cards for you alone
No virtualisation: the whole card works for you.
Once your use is approved
We make the cards available once your use has been approved.
Run by Maxime
Our GPU lead handles capacity sourcing, commissioning and follow-up.
Starting price, firm price on quotation. Prices in US dollars.
Train your open-source AI models.
From scoping to the delivered model, measured on your own cases.
Scoping
Which tasks, which data, and whether training is really useful. Often, connecting the model to your documents is enough and costs less: we say so.
Data preparation
Collection, cleaning, anonymisation.
Evaluation set
Built on your real business cases, measured before and after.
Fine-tuning
LoRA or QLoRA on Mistral, Qwen, Llama or Gemma, licences checked.
Illustration: the loss goes down during training, and the evaluation on your cases is better after than before.
What you receive
Your model (the weights belong to you), its measured evaluation and its documentation.
Where the training runs
For your projects, on H100 cards we rent or on your own infrastructure. For our research and benchmarks, on the supercomputers of the BSC AI Factory, which selected Propul AI.
Around your cards.
Optimise your GPUs
Measurement of real utilisation, tuning of model serving (vLLM, SGLang), quantisation, batching, dashboards.
You receive a before-and-after report.
Deploy privately
On your servers, with a European host or on a Kap Box. A standard API: your existing tools plug in without being rewritten.
Documented GDPR and AI Act compliance.
Size and buy your cards
Sizing of your needs, comparison of buying, renting and cloud over 3 years.
You receive the supplier quotations.
Five steps. The first one is free.
GPU assessment Free
30 minutes to look at your cards or your project.
Audit or scoping
We measure your cards, or we scope your project.
Costed plan
What needs doing, and what it costs.
Delivery
We carry out the plan with your team.
Monthly follow-up
Optional.
Frequently asked questions
Does my data leave?
No. It stays on your infrastructure, or with the European host you choose.
Do I have to buy GPUs?
Not necessarily. We compare buying, renting and cloud.
Train a model, or connect it to my documents?
We tell you what is enough. Often, there is no need to train.
Who owns the model?
You do, weights included.
How is this different from Kap?
Kap is a ready-to-use assistant. GPU services are made to measure.
Your cards can work harder.
Thirty minutes to look at your GPUs together. It is free.