Local, cloud and frontier: separate the questions
Trace deployment, capabilities and complete data destinations.
Reviewed: 2026-10-03
Illustrated teaching example • Fictional data
Narration: Microsoft Mark · locally synthesized

00:00:00.000 — Three labels do not describe a whole system.

Local, cloud, and frontier are not three mutually exclusive categories. Local and cloud describe deployment aspects. Frontier is a changing capability label. Licensing, connectivity, tool permissions, and organizational authorization need their own checks. Start with a system map rather than a reassuring label.

Visual description:
Capability ≠ Location ≠ Licensing ≠ Authorization

00:00:25.417 — Separate the dimensions.

Ask which capabilities the task needs, where inference runs, what the license allows, and where the supporting services send data. A downloaded model can still have license restrictions. A hosted model may use local tools. A strong model does not automatically make a workflow appropriate for sensitive information.

Visual description:
Capability: Task-specific quality and modalities
Deployment: Where inference and tools run
Licensing: Terms for this model and use
Authorization: Approved data and complete boundary

00:00:49.333 — A local model can have remote dependencies.

In this fictional configuration, the interface and model run locally. The retrieval step calls a remote embedding service, and traces are uploaded to a cloud logging service. The local model label therefore does not establish that all information stays on the device.

Visual description:
Local interface
Local model
Remote embeddings
Cloud traces
Illustrated configuration · not a tested product

00:01:09.333 — Trace every information destination.

Make a record of inference, connectors, embeddings, storage, telemetry, identity, and operator access. Check what information each component receives and retains. Credentials, backups, and administration also matter. If a destination or retention policy is unknown, treat that as an unresolved part of the review.

Visual description:
Inference + tools: Inputs, outputs and destinations
Storage + logs: Retention, backups and access
Identity: Credentials and operator privileges
Unknown destination: Resolve before sensitive use (warning)

00:01:35.208 — Choose for the actual task.

For public fictional practice, a simple chat tool or paper example may be enough. For restricted organizational work, use only a workflow authorized for that data and purpose. Local hosting, a clearance, or a product logo alone does not confer authorization. Keep unresolved sensitive work out of training examples.

Visual description:
Fictional practice: Use supplied examples; minimize setup.
Restricted work: Verify the authorized complete boundary.
Training is not organizational approval.

00:01:59.958 — Find the hidden remote service.

Pause and inspect the illustrated flow. Which parts are local, and which send information elsewhere? Draw the complete boundary, list the unknowns, and identify who must resolve them. A good answer includes embeddings and telemetry, not just the model endpoint.

Visual description:
What must be checked before calling this workflow local-only?
Embedding destination.
Telemetry and storage.
Network reach and authorization.
Your evidence: a complete data-flow record.