Where should your AI run? Four setups and how to choose

Every AI feature sends data to a model somewhere. Which model, run by whom and where, is a choice, and there are four realistic ways to make it, from "change one setting" to "run it on hardware you own". For most companies the answer is a mix: the strongest model for everyday work, and a tighter setup for the data that needs it.

Four ways to run a model

1. A major provider's model, in a region you choose. The large AI providers offer their models with data processing in several regions, including the UK and Europe. This is the least work and gives you the most capable models. It answers where the data is processed. It does not change whose law the provider answers to, which we covered in Where your data lives.

2. A European model provider. Mistral AI, based in Paris, is the best-known example. You call an API much as you would with any other provider, and the company you contract with is European. Check where the API processes data and what the contract says, as you would with anyone.

3. Open-weight models served by a cloud provider. Clouds now run open-weight models behind an API and charge per use. Scaleway's Generative APIs, for example, serve models including OpenAI's open-weight gpt-oss and Mistral's models from a data centre in Paris. OVHcloud, another French provider, offers a similar service called AI Endpoints. You get the convenience of an API, from a provider you pick, with models whose weights are public.

4. Open-weight models on infrastructure you control. Rent GPU servers from a host of your choice, or run the model on your own hardware. This gives you the most control: nothing leaves your environment, and you choose exactly which model version runs. It is also the most work, and you pay for the hardware whether it is busy or not.

	Where data is processed	Provider's home law	Effort to set up	Model choice
Major provider, chosen region	The region you pick	Usually US	Low	Widest, including the strongest models
European model provider	Check the contract	EU	Low	That provider's models
Open-weight model, cloud provider	That provider's data centres	Where the provider is based	Low	The open-weight models offered
Open-weight model, your infrastructure	Where you put it	Yours	High	Any open-weight model you can run

What "open-weight" means, and why the licence matters

An open-weight model is one whose trained weights are published, so anyone can download and run it. That is what makes options 3 and 4 possible. It does not automatically mean you may use it for anything.

Licences vary. Mistral publishes several of its current models, including Mistral Small 4 and Mistral Large 3, under the Apache 2.0 licence, which allows commercial use. OpenAI released its gpt-oss models in August 2025 under Apache 2.0 as well. Meta's Llama models come with Meta's own community licence, and the licence for Llama 3.2 states that the rights for its multimodal models are not granted to companies with their principal place of business in the EU.

That last example is the reason to read the licence before you build on a model, not after.

A quick way to decide

1. Does the workload touch personal data, client documents or trade secrets?

   No →

   If no

   Use the model that does the job best, with a data processing agreement in place.

   Yes

2. Does it need a provider outside US jurisdiction, because of your customers, contracts or the data itself?

   No →

   If no

   A major provider's model, in a region that suits you and with the data processing terms checked, is usually enough.

   Yes

3. Is an API you do not operate acceptable?

   No →

   If no

   Run an open-weight model on infrastructure you control, in your cloud account or on your own hardware.

4. If yes to all three

   Use a European model provider, or a cloud with no US parent that serves open-weight models.

What it costs you

The shape of the cost differs more than the price.

• Per-use APIs (options 1 to 3) cost nothing when idle and scale with use. The risk is a bill that grows quietly with traffic, so it needs a budget and monitoring.
• Your own GPUs (option 4) cost the same every month whether you use them or not. That suits steady, heavy workloads and sensitive data. It does not suit an experiment.

Either way, work out the running cost before you commit, at the volume you actually expect. We do that as part of every AI build plan, because a model that is cheap in a pilot can be expensive in production.

A setup that works for most companies

1. A single gateway that every AI feature goes through. It decides which model handles which request, keeps the logs, and holds the API keys in one place.
2. A routing rule by data type. Sensitive data goes to the setup you chose for it: a European provider, a provider in a particular region, or a model on your own infrastructure. Everything else can use whichever model does the job best.
3. Keys and accounts in your name. The gateway, the model accounts and the logs belong to you, so changing a provider is a configuration change, not a migration.

That gives you the strongest models where they are fine to use, and a tighter path where it matters, without running a GPU cluster for tasks that do not need one.

Sources

1. Models overview Mistral AI documentation
2. Introducing gpt-oss OpenAI, 5 August 2025
3. Qwen3-235B and OpenAI's gpt-oss models now available on Scaleway's Generative APIs and Managed Inference Scaleway
4. OVHcloud launches AI Endpoints OVHcloud
5. Llama FAQs Meta
6. Llama 3.2 Community License Agreement Hugging Face