Define what private means
Private can describe several different arrangements: a dedicated environment managed by a provider, a model running in your cloud account, or infrastructure operated by your own team. Ask where inference happens and which organizations can access inputs, outputs, logs, and backups.
A deployment label alone does not establish how data is handled. Review the actual architecture, contract, and settings.
Compare the operational work
List who is responsible for model updates, infrastructure, capacity, monitoring, incident response, and ongoing evaluation. Include that work in the comparison alongside usage and hosting charges.
For a self-managed model, test representative tasks on the hardware and configuration you intend to use. For a hosted service, test the specific model and service tier under consideration.
Check the information lifecycle
Document what enters the system, what is retained, how information is deleted, and whether any content is used to improve models. Confirm the settings that apply to your account rather than relying on general marketing statements.
- Input and output retention
- Training and model-improvement settings
- Access controls and audit records
- Backup, deletion, and export processes
- Subprocessors and deployment locations
Plan for change
Decide how your team would export its work, replace a model, or continue a critical task during an outage. Evaluate a fallback against the same task set before relying on it.
The appropriate arrangement depends on your requirements. Use a pilot and a review by your technical and data owners to make the decision.