On-prem and air-gapped AI
Check whether every component stays inside the intended boundary.
In this lesson · 03 of 7
All lessons in this track →Saved in this browser. View your learning.
Start with the idea
Check whether every component stays inside the intended boundary.
Local inference means the model runs locally. Other components may still communicate remotely. A system diagram needs configuration and observed tests before it proves containment.
Air-gapped system: A system separated from external networks; its full operating and transfer arrangements still need review.
See the example and check your understanding
A made-up teaching example
The laptop uses a local model, but search calls a remote embedding service and traces go to a hosted dashboard. Two data paths leave the machine.
Ask yourself
Does a local model make this whole workflow disconnected?
Show the explanation
No. Replace or approve each external path and test the boundary using invented data.
Start with these checks
- Map models, tools, logs and transfers.
- Do not equate an API with internet access.
- Label capacity calculations as estimates.
Use the different practice case below and check every requirement in its full instructions.
The video explains the main idea. Use the sources, instructions and checks below to complete the practice.
Captions appear here during playback.
On-prem and air-gapped AI
Check whether every component stays inside the intended boundary.
Video playback is unavailable. Read the descriptive transcript below, go directly to practice, or open the video file.
Source-based teaching example · Fictional data · Reviewed 2026-10-03. Microsoft Mark · locally synthesized. On a small screen, use fullscreen for diagram details or read the transcript.
Read the descriptive transcript
0:00 · What you will learn
Check whether every component stays inside the intended boundary.
What appears on screen
Check whether every component stays inside the intended boundary.
0:06 · The idea
Local inference means the model runs locally. Other components may still communicate remotely. A system diagram needs configuration and observed tests before it proves containment.
What appears on screen
Local inference means the model runs locally. Other components may still communicate remotely. A system diagram needs configuration and observed tests before it proves containment.
0:21 · One useful term
Air-gapped system. A system separated from external networks; its full operating and transfer arrangements still need review.
What appears on screen
A system separated from external networks; its full operating and transfer arrangements still need review.
0:32 · A worked example
The laptop uses a local model, but search calls a remote embedding service and traces go to a hosted dashboard. Two data paths leave the machine.
What appears on screen
The laptop uses a local model, but search calls a remote embedding service and traces go to a hosted dashboard. Two data paths leave the machine.
0:44 · Your turn to think
Does a local model make this whole workflow disconnected? Pause and decide before continuing.
What appears on screen
Does a local model make this whole workflow disconnected? Pause the video if you need more time.
1:01 · Why this answer works
No. Replace or approve each external path and test the boundary using invented data.
What appears on screen
No. Replace or approve each external path and test the boundary using invented data.
1:10 · Check 1
Map models, tools, logs and transfers.
What appears on screen
Map models, tools, logs and transfers.
1:17 · Check 2
Do not equate an API with internet access.
What appears on screen
Do not equate an API with internet access.
1:23 · Check 3
Label capacity calculations as estimates.
What appears on screen
Label capacity calculations as estimates.
1:29 · Now try a different case
Open the lesson and try its different practice case. Use every check in the full task. Watching alone does not mark the lesson complete.
What appears on screen
Open the lesson Try the new case Check and revise The full task and its requirements are in the linked lesson.
Watching explains the example. Complete the lesson practice to check what you can do. Explore all visual lessons →
Map the whole deployment. Build a judgment you can explain and a result another reader can inspect.
Prerequisites: know the difference between generated text and source evidence. For configuration vocabulary, use the AI workflow reference. This is public self-study, not official government training, accreditation, a clearance or a professional credential.
Observable outcomes: Trace every destination in a local retrieval workflow; calculate a weights-only estimate and design a containment verification test.
Part 1 Map the whole deployment
Model weights define learned behavior. A runtime serves them; an API is an interface, including a localhost API. A harness supplies instructions, retrieval, tool calls and permissions. Compute is the hardware and resources used. Local/cloud, open/closed weights, model capability and chat/API/agent workflow describe different axes; local and frontier are not opposites.
A local process can call a cloud embedding service, web search, telemetry collector or remote tool. Verify the complete data flow, network configuration, logs, backups and administrators before claiming that data stays inside a boundary. A disconnected system still needs approved imports, signed or hashed artifacts, license review, patching, identity controls and recovery. Open weights do not prove open-source licensing or trustworthiness.
Memory planning includes weights, quantization, context cache, concurrency and runtime overhead. A rough weights-only calculation for 8 billion parameters at 4 bits is 4 billion bytes; actual serving memory is higher. Measure the target workload rather than promising that every model fits a particular laptop.
Part 2 Synthetic worked example
Synthetic architecture: laptop UI → localhost serving API → local model; retrieval → remote embedding endpoint; traces → hosted dashboard. Inference is local, but two data paths leave the laptop. Replace embeddings and traces with approved internal alternatives, document egress restrictions, and test with synthetic canary strings. A drawing alone does not demonstrate containment.
Guided practice: write the tempting conclusion first, then underline the fact that does not support it. Rewrite the conclusion with that limitation visible. Identify one additional observation that would change the decision.
Next step: carry your artifact into role evidence in lesson 04; retain assumptions and unanswered questions.