Architecture
One API. One bill. One audit trail.
Foundation Performance Cloud does not host the models. It is the governance layer in front of them — the control plane that authenticates the request, decides which system may lawfully serve it, proves what happened, and meters the cost.
The path of a request
The control plane sees the request in order to route and meter it. For the most stringent requirements, a future data-plane-bypass mode keeps the content on a direct client-to-model path while the control plane issues only the signed routing decision — see the Trust Center.
Three things you can demonstrate, not just assert
Residency attestation
Every response carries headers — and a stored, exportable record — naming the jurisdiction and ownership of the system that actually processed it, and whether your residency policy was satisfied.
Honest by construction: the attestation reflects what happened, never the policy that was merely requested.
The fail-closed guarantee
If you mark a workload nz_sovereign and no compliant NZ-owned, NZ-located system is available for that model, the request errors. It is never silently sent offshore or to a foreign-owned provider.
A gateway that quietly falls back offshore is worse than none — it makes a false sovereignty claim. We refuse instead.
Verifiable neutrality
We route by residency, health, then cost — on the merits — and the decision rationale is logged. You can see which systems were considered and why one was chosen.
Compatible with what your developers already use
The endpoint speaks the OpenAI API. In most cases adopting Foundation Performance Cloud is a base-URL and API-key change — no SDK swap, no application rewrite. Your spend caps, model allow-lists, and residency policy are configured centrally and enforced on every call.