AI deployment options
The deployment wizard asks one question about AI, and it asks it once. Private AI puts the model inside your own subscription, Managed AI borrows ours, and No AI leaves the installation without any AI at all.
The question is about where the AI Accelerator gets its model from, not about whether the Accelerator is worth having. Two of the three answers give you a working Accelerator; they differ in whose infrastructure runs it, whose network the prompt travels through, and whose bill the tokens land on. The third answer gives you no Accelerator, on purpose.
Read the whole page before you run the wizard. This setting is fixed for the life of the installation, so it is not a decision you can revisit once the environment is up.
| Option | What gets provisioned | Where an interaction goes | Who is billed |
|---|---|---|---|
| Private AI | A model instance in your own Azure subscription | Stays inside your network | You, on your Azure bill |
| Managed AI | Nothing | Out to the Accelerator hosted by Cerebrum City | Included up to the cap, then per token |
| No AI | Nothing | Nowhere. There is no Accelerator to call. | Nobody |
Private AI creates a new instance of the model inside your Azure subscription, alongside the rest of the installation. From then on, everything you type into the AI Accelerator is answered by that instance. The prompt, the artifacts it generates, and anything you paste into the chat stay within your own network boundary.
Because the model is yours, its consumption is yours too. Token usage appears on your Azure bill under your own subscription rather than being metered by Cerebrum City, and there is no interaction cap: what you can use is what you are willing to pay Azure for.
| Setting | Value |
|---|---|
| Model location | Your Azure subscription, in the deployment region |
| Network boundary | Interactions do not leave your network |
| Interaction limit | None |
| Token limit | None beyond your own model capacity |
| Billing | Your Azure bill |
| Prerequisite | Foundry quota in place before the install, see Azure quota requirements |
Managed AI provisions no AI infrastructure in your subscription at all. Instead the installation calls the AI Accelerator hosted by Cerebrum City. There is nothing for you to size, nothing to bill, and no quota to request, which is the point: the option exists for subscriptions that cannot provision model capacity of their own, student and trial subscriptions in particular.
What you give up is capacity and continuity. The allowance is small and it is counted per installation, not per user, so everybody working in that installation draws from the same pool.
| Setting | Value |
|---|---|
| Model location | Hosted by Cerebrum City |
| Network boundary | Prompts leave your network to reach the hosted Accelerator |
| Interaction limit | 20 per installation |
| Token limit | 1 million per installation |
| Which limit applies | Whichever is reached first |
| Counted per | Installation, shared by everyone using it |
| Beyond the allowance | Additional tokens can be purchased |
| Memory | None. The hosted Accelerator is stateless. |
| Provisioned in your subscription | Nothing |
No AI turns the AI off for that installation completely. No model instance is created in your subscription, and no hosted Accelerator is referenced or reachable either, so this is not a quieter version of Managed AI. There is simply no AI in the environment.
What this removes is the authoring assistant, not any runtime capability. Pipelines, pipeline components, custom functions, and maps that were built elsewhere run exactly as they do anywhere else, because the Accelerator only ever helped you write them. The usual pattern is to develop in an environment that has the Accelerator and to run the result in one that does not.
| Setting | Value |
|---|---|
| Model location | None |
| Hosted Accelerator | Not referenced, not accessible |
| AI Accelerator availability | Unavailable in this installation |
| Provisioned in your subscription | Nothing |
| Billing | None |
| Effect on running flows | None. Existing artifacts behave identically. |
| If this environment | Choose |
|---|---|
| Is where integrations get built, and the organization has model quota | Private AI |
| Must keep every prompt inside its own network | Private AI |
| Is a student, trial or evaluation subscription with no model quota | Managed AI |
| Only needs to show what the Accelerator can do | Managed AI |
| Runs production traffic and hosts no development | No AI |
| Must be able to state that it reaches no AI service | No AI |
Decided before you install
Private AI for the network boundary, Managed AI for a subscription without quota, No AI for production. If you chose Private AI, clear the model quota first in Azure quota requirements, then continue to Azure marketplace deployment. What the Accelerator does with the model once it is there is covered in AI Accelerator.