# Control is a deployment decision with an operating cost.

> Decide where the AI service should run by looking at your information, access rules, budget and ability to support it. Owning a model is one option, not the starting assumption.

Canonical URL: https://marmation.com/expertise/cloud-and-private-inference/

## The decision

Which data controls do you need, and who will keep them working?

Follow prompts, logs, saved responses, search indexes and backups as well as the model request. Compare hosting options against those restrictions, response-time needs and the people required to operate them.

## The data restriction must survive real service operation

An IT team requires data to remain in a specified region and access to be restricted through its network and identity controls. The team also needs reliable latency and a clear path for model updates.

## Hosting location does not control every copy of the data

Hosting does not guarantee governance. Model weights, logs, prompts, caches, embeddings and backups can all carry sensitive data. GPU costs and inference operations are not trivial.

## Compare deployment choices against the actual restriction

Write down the restriction precisely: the information involved, permitted location and network path, who may access it and how long copies may remain. Compare provider arrangements and cloud controls against those needs, then assess capacity, response time, model licensing and cost. Assign responsibility for updates, rollback and incidents. A deployment diagram is useful only when its controls and operating duties can be verified.

## The infrastructure and operating plan to inspect

- Data-flow and region controls architecture.
- Model endpoint, secret and network access design.
- A plan for response time, capacity, cost, monitoring and diagnostic records.
- Update, rollback and incident operations.


## More hosting control brings more work

Private hosting offers control but increases engineering burden. Managed providers may meet requirements more economically. Avoid absolute security or data privacy claims.

## Check access, logs, capacity limits and recovery

Test authentication failures, log redaction, model update rollback, rate/usage caps, failover and policy compliance against actual agreements.

## Write down what control must mean here

Define what control means before selecting the hosting topology.

## Related reading

- [Private AI is not one architecture.](https://marmation.com/solutions/private-enterprise-ai/)
- [Choose the model for the job, the data and the operating environment.](https://marmation.com/solutions/model-deployment-and-customization/)
- [How to commission private enterprise AI.](https://marmation.com/review/how-to-commission-private-ai/)