AI on your own data,
on our own GPUs.

For a lot of work you could, and your data would sit in the EU. What an EU region does not change is who the provider answers to. The US CLOUD Act lets American authorities compel any provider under US jurisdiction to hand over data in its possession, custody or control, wherever it is stored, and that reaches a European subsidiary through its US parent. DIVERSITY has no US parent: no US company owns or controls us, so there is no entity above us for that order to be served on.

It depends on the task, and it is the first thing the discovery call settles. If a closed model is the better tool for what you want to do, we will say so.

Open-weight models, usually from the Qwen, GLM and gpt-oss families. We do not publish a fixed list, because the catalogue changes. Tell us the task and we will tell you which fit. On a dedicated endpoint you can also bring your own model: we agree it with you, by its licence, its size and the GPUs it needs.

Model as a service is shared: models we host that several customers call, billed per use. A dedicated endpoint is yours, serving your own model or an open-weight one we host, agreed with you by the model's licence, its size and the GPUs it needs.

The AI services on this page run in Amsterdam, including model as a service, dedicated endpoints, GPUs by the hour and training clusters. The fleet there is currently Nvidia H200, H100, L40S, L4, A10, V100S and RTX PRO 6000 Blackwell Server Edition; tell us the workload and we will say which fits. The RTX PRO 6000 Blackwell is also the GPU tier of our Bare Metal, usually two to a server, and those servers are available in all four of our datacenters (Madrid, Munich, Helsinki and Amsterdam), on Bare Metal's six-month minimum term. If you need something the fleet does not have, we build to order and will tell you where it goes and how soon.

No to both. We do not offer a private coding assistant, and we do not run models on hardware at your site.

Yes, if that is your decision. We do not recommend it: we host open-weight models on our own GPUs, and an outside service puts your data back where you came to us to avoid. We will tell you on the call what you would gain and what you would give up.

We run the model endpoints, the shared models and the vector database. Once you are a customer, you reach us through the customer console or at support@ondiversity.eu. Support is free on every product we sell, with a 4-hour target response time on support requests, and the paid advanced plan adds shorter targets plus weekend and out-of-hours cover.

99.9% uptime, by contract, with automatic credits if we miss it, on every product we tie to an SLA. It measures whether the service is up. A hardware fault on a GPU server you rent through Bare Metal is different: that is a repair rather than an SLA event, and we aim to have it back in service within four working hours.

That is MLOps, and it is part of what our engineers do: deploying the model, monitoring it once it runs, and keeping a registry of what was trained and what is deployed.

Model as a service is billed per use. We do not publish prices for any of it; ask us.

The discovery call, and it is free: 60 minutes, an engineer rather than a salesperson, and a written diagnosis inside a week. Already a customer? Same call, same form: tell us what you run with us and it starts from there.

No. Your data never trains anything except your own models. That includes the shared models: what you send them is not used to train them.

No provider can promise that, and we do not. The Act puts most of its obligations on whoever provides and deploys the AI system, which is usually you. What we offer is a readiness assessment: how your use case is classified for risk, the documentation it needs, and the governance around it.

DIVERSITY is a European company with no parent company, US or otherwise, so there is no entity above us for an order from outside the EU to be served on. That is a structural answer rather than a policy one.