Privacy Policy & Data Retention
Effective 2026-08-27 · v1.0 · Inference Yield ("we")
The short version
- We do not store your prompts or completions. Request and response content is processed in memory and discarded when the response completes.
- We never train on customer data, and we never sell or share request content.
- Per request we retain metadata only: token counts (including cached-token counts), timing (arrival, time-to-first-token, duration), model and configuration identifiers, HTTP status, and a hashed API-key identifier. This is what billing and reliability reporting require.
- Zero-data-retention (ZDR) handling is our default and only mode for content; there is no content-logging mode in production.
Details
Logging
Application logs contain metadata only. Metering records are retained for 90 days for billing reconciliation, then deleted. No log line contains prompt or completion text; this invariant is enforced by an automated content-leak test that runs against our gateway on every change.
Debug buffer
If we are actively investigating an incident with a routing partner, we may enable a short-lived encrypted buffer holding request/response bodies for at most 24 hours, after which it is destroyed. It is off by default, requires an explicit operator action, and its status is disclosed on the status page whenever active.
Subprocessors
GPU capacity is rented from infrastructure providers (currently RunPod, Inc., US datacenters). Request content transits their machines in memory during inference under the same no-retention handling.
Security
TLS in transit. No human review of request content occurs — and after a response completes, none is possible, because content is not stored.
Location
Serving infrastructure is in the United States. Datacenter locations are disclosed per-model in our /v1/models metadata.
Changes
Material changes are versioned and dated on this page, with notice to routing partners before taking effect.