Offline is an operating state
AI without a live cloud dependency - designed, tested and owned.
A private server can run approved models without sending prompts to a hosted AI provider. A genuinely offline or air-gapped environment needs more: controlled transfer, prepared dependencies, update procedures and a recovery plan.
Choose offline AI only when the policy and risk justify the operational cost. A restricted connected or private-first hybrid design is often easier to maintain.
Buyer field note
Let the work and the room reject the wrong machine.
- 01 / Fit
- Start with the accepted task, not a chassis or GPU headline.
- 02 / Site
- Check power, cooling, noise, network and maintenance access.
- 03 / Alternative
- Keep cloud, colocation or a smaller system in the decision.
Prepare everything the workload needs
Models, container images, drivers, licences, document tools and dependencies must be acquired and recorded before the boundary closes.
The acceptance test must prove a cold start and the agreed workload without an external service.
Capacity budget
Memory demand has more than one layer
- Model
- Weights and quantisation.
- Context
- Working input and cache.
- Concurrency
- Simultaneous active work.
- Headroom
- Operations and change allowance.
Connect the category to the physical and operating system.
Platform views and technical plates show what must be tested beyond a product label or aggregate specification.
Updates still need a route
Offline does not eliminate vulnerabilities, model changes or hardware faults. It changes how updates and support evidence cross the boundary.
A signed media process, quarantine system or scheduled controlled connection may be needed.
Site boundary
The room is part of the specification
- Electrical Circuit, protection and UPS policy.
- Thermal Airflow, room heat and cooling route.
- Network Access, bandwidth and isolation.
- Service Rack space, noise and maintenance access.
Logging and deletion matter more, not less
An isolated system still handles user activity and potentially personal or confidential data. Audit, retention, deletion and backup remain required.
Physical access becomes part of the security model.
Procurement record
Make the route to acceptance inspectable
- Fit brief Workload and alternatives recorded.
- Reference test Conditions and results retained.
- Exact build Bill of materials and substitutions visible.
- Handover Owners, versions and exclusions signed off.
Do not confuse local with isolated
A server on the office network may still have internet access, telemetry, package repositories and integrations. The data flow must be documented rather than assumed from physical location.
The order should state the intended connectivity state and who may change it.
Questions answered
Straight answers to common questions
Can local AI work without internet access?
Yes, when the chosen models, runtime and dependencies are prepared for offline use. Updates, licensing and support must also be designed for that state.
Is every private AI server air-gapped?
No. Private describes the controlled deployment. Air-gapped means a deliberately isolated network boundary with operational controls.
Does air-gapping guarantee security?
No. Physical access, removable media, identity, software supply chain, logging, backup and insider risks still exist.
Continue the decision
Useful next steps
Put the claim to work
Turn this guidance into a testable requirement.
The brief asks about workload and operating conditions - not just budget.