Sovereign by design: why industrial AI is moving on-prem

For a decade, the default answer to every analytics question was "send it to the cloud." For industrial AI, that default is breaking — visibly, and for reasons that compound rather than cancel.

Three forces, one direction

Regulation. Gartner expects 65 percent of governments to mandate AI sovereignty requirements by 2028. For manufacturers serving aerospace, defense, and automotive supply chains, data residency is shifting from a preference to a contract clause.

Economics. Deloitte sizes sovereign AI infrastructure spend at roughly $78 billion in 2026, growing at 28 percent annually, while the edge AI market is projected to grow from about $25 billion in 2025 to nearly $119 billion by 2033 (Grand View Research). Streaming high-frequency plant telemetry to the cloud and back is often the most expensive and slowest possible architecture for a decision that must land on the line in milliseconds.

Data gravity. The most valuable industrial data — machine telemetry, quality imagery, process parameters — is heavy, continuous, and competitively sensitive. It resists leaving the building, and increasingly it shouldn't have to.

What "sovereign by design" means in practice

It means the reference architecture assumes on-prem and edge as the default — edge gateways and on-prem compute at the OT layer, a lakehouse inside the perimeter, models trained and retrained where the data lives — with cloud as an optional extension, not a dependency. Built this way, compliance is a property of the system rather than a remediation project, and the cost model scales with the plant instead of with egress fees.

The practical takeaway for operators: ask every AI vendor one question — can this run entirely inside my walls? If the honest answer is no, you are not buying a capability. You are renting one.