Perimeter: decide where it runs
In one breath: the trust anchor is one axis, residency is the other. You draw the boundary the compute lives inside, and models are hosted within it, behind your own endpoint.
Three shapes of boundary
Your own cluster (on-prem)
Inference runs entirely inside your data centre or GPU cluster. Nothing leaves the building, and you still get the same receipts. Run it on hardware you already own, or we help you provision it.
One jurisdiction (regional)
Pin the compute to a single region so data stays inside one legal boundary. This is the simple answer to residency rules: if your regulator says "EU only", the perimeter says it too, and the receipt shows it.
Spread across regions (transcontinental)
The counterintuitive one. Place operator-disjoint nodes in different countries, so that no single site, and no single country, ever holds the whole computation. Paired with the Sovereign tier, this turns geography itself into part of the privacy guarantee.
The endpoint lives inside
Whatever shape you choose, the compiler hands you an OpenAI-compatible endpoint hosted inside that perimeter:
client = OpenAI(
base_url="https://infer.your-perimeter.ai/v1", # yours, inside your boundary
api_key="zk-..."
)
Your engineers keep the SDK they already use. The privacy machinery is compiled in underneath, not bolted on around them.
Perimeters are, by definition, configured per engagement: your hardware, your regions, your rules, agreed in writing. Regional pinning and on-prem deployments follow established patterns; transcontinental operator-disjoint meshes are deployed together with our team. See the Status board.