Shared, dedicated or private AI: choose the operating boundary

Begin with what must stay under your control

The right deployment model follows from the workflow, sensitivity, connectivity and operating capacity. Labels alone are insufficient: ask what is actually shared, isolated and managed.

Shared service

Infrastructure serves more than one customer, with separation implemented through the provider's controls. This may suit a defined workload when the contract, isolation, retention and service terms meet the organisation's requirements.

Ask how tenant separation is tested and what customer data can reach model services, support systems and logs.

Dedicated environment

Some resources are reserved for one customer. Dedicated can refer to a tenant, database, model endpoint or physical hardware; it does not necessarily describe the entire processing chain.

Ask which resources are dedicated and which management, support or upstream services remain shared.

Private deployment

An environment is deployed within an agreed private infrastructure boundary. It may be on premises or in a private cloud, and it may still have approved external connections. Private and air-gapped are not synonyms. Disconnected operation is a separate design with implications for updates, monitoring and support.

Ask who operates the environment, manages keys, applies patches and handles recovery and incidents.

Compare responsibilities before features

For every model, document identity and source permissions, model access, network paths, retention, important action approvals and exit arrangements. Microsoft's sovereignty documentation illustrates that public and private models can have distinct controls. Langfuse's self-hosting guide illustrates the operational choices involved in running application components yourself.

The GDPR does not make one of these labels a blanket compliance certificate. Assess the actual processing and applicable requirements.

Apply this to NeuroCluster

The Cloud & Pods overview describes deployment options. Confirm availability, integrations, sizing and operating responsibilities for the engagement. An Industrial Edge Pod target architecture requires pilot validation; it is not a universal preconfigured hardware promise.

Sources and review date

Checked on 15 October 2026. The model descriptions are explanatory guidance, not vendor benchmark results. The primary sources above support the distinction between product capabilities and deployment responsibilities.