Private Inference Cloud

Every company gets its own boundary: gateway scope, data plane, weights, and meter.

Your Private Inference Cloud is the boundary everything else in these docs lives inside: a dedicated gateway scope, your own data plane, your own weights, and your own meter. Data, metering, memory, and training are scoped per tenant. Agents are scopes within your tenant, never sibling databases.

Residency

Regions are part of the contract, written the way regions are written into contracts: za-1 · South Africa is the first. A residency tenant's data plane, storage, agent memory, and training all live inside the region you chose. Residency is placement, never a different product.

By default your workloads run on Cloudlight, the low latency plane. Residency and heavy workloads run on Cloudheavy, pinned to your region.

Provider policy

Your organization can hold a provider allowlist, enforced on the completions path itself. A policied tenant's prompts can never reach an unapproved provider, whatever the route or the model field says: the request is refused with a 403 permission_error. If the policy state cannot be verified, the Gateway fails closed rather than guessing about your commitments.

Yours, exportable

Your data and your fine-tuned weights belong to you: stored in your scope, exportable at any time, never pooled with anyone else's. Credentials follow the same rule in reverse: agents and sandboxes never hold provider, tool, or database keys; everything is injected at the platform edge and revoked centrally.