ENTERPRISE AI

Local, sovereign, and hybrid

Local means the model and the harness run on hardware you control. Sovereign means the boundary holds. Hybrid routes each job by the data and the risk.

Figure 1: Three places a model can run: on your hardware, inside a boundary you control, or routed by the data.

1. A cloud chat box on a desk is not local

Local means model inference and the harness run on hardware the organization controls. A desktop window that calls a cloud API is not fully local, however private the office feels. The test is simple: does the confidential document leave the building in order to be read by the model?

Sovereign is broader than the address of the server. It means meaningful control of the infrastructure, the models, the data, the identities, the telemetry, and the governance, inside a boundary you can describe. Model location is not enough if embeddings, reranking, logs, or tools still send information out.

2. Open weight is a choice, not a winner

A local or open-weight model expands the choice for private, repetitive, high-volume, or constrained work. It does not automatically replace a managed frontier service. Ambiguous synthesis may still justify a stronger managed model, with the context kept under control. A high-impact action needs an approved model plus a deterministic check and an authorization, because model quality does not replace accountability.

Restricted content stays on a private path unless a written policy explicitly allows otherwise. Write that routing policy during discovery, before anyone builds the agent. A route invented after the demo is how sensitive text ends up in the convenient model.

3. Memory is not a slogan on a spec sheet

The architecture guide gives a rough early estimate, not a Cloudadorn measurement: about half a gigabyte of weight storage per billion parameters at 4-bit quantization, and about two gigabytes at FP16, before runtime overhead. A 7-billion-parameter model is then in workstation range at 4-bit. A 70-billion-parameter model is not. Actual capacity also depends on the runtime, the context length, the cache, the batch, and the architecture.

For a mixture-of-experts model, the total parameters drive how much weight you store, and the active parameters per token drive the speed. A large model with few active parameters can be faster than a dense model of the same headline size, and it can still need the memory for the full weights. One published illustration used about half a megabyte of cache per token. At a long context and several concurrent requests, that illustration becomes tens of gigabytes. Your runtime will differ. Size from a measurement, not from the headline parameter count.

Layman rule: Route by the data and the risk. Do not route by which demo was more impressive.

4. A starting map, not a mandate

Private repetitive classification starts local or private. Ambiguous writing can start on a frontier model with controlled context. A high-impact action starts with an approved model, a check that does not depend on the model, and a named authorization. Restricted content starts private. Each of those is a starting path. Evidence from your own tasks is what keeps it.

5. What to decide next

For one real workload, write where the bytes go: the prompt, the documents, the logs, and the tool calls. If any of those leave the boundary, the design is not sovereign, and it may not be local either.

Wherever it runs, the loop around it has a shape: the loop inside the harness.