On-premise es una arquitectura, no un compromiso

Cloud-native platforms ported to an on-premise deployment carry their assumptions with them. A licence server that must be reached. Telemetry that fails closed. Model weights fetched at runtime. Each is invisible in a connected environment and each is a blocking defect in a field with no outbound route.

Designing for the air-gapped case from the beginning forces different decisions. Updates ship as signed offline bundles. Model registries are local and complete. There is no runtime dependency on anything outside the operator’s network, and that property is tested rather than asserted.

The benefit is not only regulatory. A platform that keeps working when the satellite link drops is a platform the field can trust, and edge inference that survives an outage is worth more than cloud inference that does not.

There is a real cost. Updates are slower. Debugging a site you cannot reach is harder, and it requires diagnostics that can be exported as a file and read offline. Capacity planning is fixed rather than elastic, so the sizing conversation happens up front and must be got right.

Those are the trade-offs. They are worth stating plainly rather than presenting on-premise as free.