Instance model and model hosting: two questions to keep apart

What is the instance model?

The instance model answers the question: How is my platform separated from that of other customers? This covers sign-in, permissions, apps, documents and logs. Three levels are common:

• Shared environment: All customers work in one installation and are separated logically (tenants).
• Separate environment: Each customer has its own installation with its own account management, on servers that several customers share.
• Hardware for one customer alone: at the provider or on the customer’s premises.

What is model hosting?

Model hosting answers the question: Where does the AI model compute, and who operates it? Typical answers are:

• The platform provider operates the model itself, in its data centre or on the server at the customer’s site.
• A model provider computes, for example OpenAI, Anthropic, Google or Mistral AI, often via a cloud in the EU. The platform sends every question there. More on this in the guide Wrapper or standalone platform?
• The customer provides the model, for example on its own computing cluster.

On top of this comes the question of whether several customers share one model. That is the case with almost all cloud offerings and is not a problem in itself, as long as each request is processed separately and not used for training.

Shared model or own instance?

An AI model needs powerful GPUs. Because they are expensive, most providers let one model compute for many customers at the same time. This is called a shared interface (“shared API”) or multi-tenancy (“multi-tenant”). Each request is processed on its own; customers are separated logically, not by their own hardware. An own instance (dedicated, “single-tenant”) is a model that computes for one organisation only, on hardware available to it alone, at the provider or on the customer’s premises.

Together with the instance model, this gives four levels of separation:

Level	Platform and data	AI model	Typical for
1. Everything shared	One environment for all customers, separated logically	Shared	Entry-level offerings, test environments
2. Separate environment	Separate environment and account management per customer, on shared servers	Shared	Cloud offerings for organisations
3. Own instance at the provider	Separate environment	For this customer only, on the provider’s hardware	Large organisations, special requirements
4. Hardware on your own premises	On the customer’s server	On the customer’s server	Organisations with their own data centre, operation without an internet connection

Providers use different terms. Langdock, for example, names a shared cloud as standard, an environment for one customer only from 2,000 seats and operation in the cloud or in the customer’s data centre from 5,000 seats. Mistral AI also offers its enterprise edition in a private cloud or on the customer’s premises. As of 11 October 2026, according to the providers.

Is a shared model a data protection problem?

Not in itself. The GDPR does not prescribe a particular design; it requires technical and organisational measures appropriate to the risk (Art. 32 GDPR). For health data in the cloud, § 393 SGB V likewise requires proof under BSI C5 or equivalent proof, not dedicated hardware. More important than the level are the questions further down: who operates the model, whether anything is stored, logged or used for training, and how customers are separated.

When is an own instance worthwhile?

• Fixed computing power: Answers do not slow down when other customers ask a lot.
• Your own choice of model: for example a specialised model that only your organisation uses.
• Your organisation’s requirements: Some security concepts or tenders require hardware for one organisation only.
• Operation without an internet connection: only with hardware on your own premises.

Against this stands the price: GPUs for one organisation alone cost money even when they sit idle at night.

How are the two connected?

Not necessarily at all. This results in four common combinations:

Platform runs …	Model computes …	Example
at the provider	at the same provider	Cloud platform that operates its own models
at the provider	at a model provider	Interface in Germany, questions go to a model provider
on your premises	on your premises	Server with platform and model, without an internet connection if you wish
on your premises	at another location	Hybrid operation (when it makes sense)

For data protection, what counts is the place where data is stored and the place it is sent to for each answer. In the second row, the platform is in Germany, but every question, together with the passages from your documents, goes to the model provider. That provider is then a sub-processor, and its law also applies to this data.

Which questions should you ask every provider?

1. How is our platform separated from other customers: shared environment, separate environment or dedicated hardware?
2. Do we have our own account management?
3. Where does the AI model compute, and who operates it?
4. Do we share the model with other customers? If so: is anything stored or logged, and for how long?
5. Is our data used for training?
6. Does data go to a different model under certain circumstances, for example under load or for certain features?
7. Where is all this stated in the contract, the TOMs and the list of sub-processors?

How does it work at basebox?

Option	Instance model	Model hosting
Cloud	separate virtual Kubernetes environment and separate account management (Keycloak) per organisation, on a shared server in the Munich-East data centre	AI model in basebox on our servers, shared; each request without permanent storage and logically separated. A model for your organisation alone is available on request; you will find the price in your personal quote.
Server	a server belongs to exactly one customer, at your site or in the data centre	on the server; the GPUs are shared only by your own services
Hybrid	platform on your hardware in your network	at another location, for example in your data centre or at a provider you connect

An external model is only available as a separate app, and only if you expressly have it set up; it runs on Google Vertex in the EU (app with an external model). Your data does not train any model.

What does each option cost at basebox?

• Cloud: €490 to €178 per person per year, the more people, the lower the price; €240 for 200 people, €199 for 1,000 people; 20 million tokens per person per year for the whole organisation.
• Server: buy once, one licence per server, plus the service contract; spread over five years, around €267 per person per year for 200 people and around €73 for 2,000 people. No quota, anyone may sign in.
• Own model in the cloud: Calculated in your quote. You will find the price in your personal quote.

List prices, excluding VAT. From around 240 people, the server is cheaper than the cloud over five years, including electricity; more in the guide Cloud or server?. The calculator works out every figure.