AI gateway data residency in 2026: prove where requests actually travel

Start with a single representative request: client application, gateway ingress, any gateway-side decryption, routing control plane, upstream provider endpoint, inference region, response path and observability systems. Draw a second path for failures and retries. A gateway may terminate TLS in one region and then forward plaintext to a model deployment elsewhere; the location of its public hostname alone does not establish where inference runs.

Make the vendor specify which data categories are covered: prompts, completions, tool arguments, embeddings, uploaded files, credential metadata, billing records, operational telemetry and support tickets. A promise that API payloads remain in a region may explicitly exclude account or billing records. Ask for the exact boundary and exceptions in writing, not a generic badge.

OpenRouter announced US In-Region Routing on September 9, 2026, alongside its existing Europe option. Its documented US endpoint decrypts requests in the United States and routes to US provider endpoints; its European definition covers the EU/EEA, UK and Switzerland rather than the EU alone. The company describes these capabilities for Business and Enterprise customers. These are vendor-documented claims to validate against the purchased plan and contract, not independent certification of a particular customer deployment.

OpenRouter also distinguishes the provider route from operational records that may be handled elsewhere. For a self-hosted proxy such as LiteLLM, hosting the proxy in a US region still leaves the selected upstream model endpoint and logging integrations to evaluate separately. For managed gateways, request the applicable deployment regions and any cross-region observability or support processing before making a purchasing decision.

Configure a workload that may use only an approved region and an explicit list of provider endpoints. Send a normal request and collect sanitized routing evidence identifying the selected provider and deployment region. Then make every eligible endpoint unavailable in a controlled test. The correct result for a strict regional policy is an explicit error or queued request, not a silent fallback to an unapproved geography.

Repeat the test during provider throttling, a gateway restart and a configuration rollout. Include retries initiated by the SDK, gateway and upstream provider so a second layer cannot bypass the rule. Keep a stable request identifier through each hop. A single successful request in the right region is not proof that the policy holds under outage conditions.

Data residency answers where processing happens. Retention answers whether and how long inputs, outputs or metadata are stored. Training restrictions answer whether a provider may use content for model improvement. These are independent dimensions. OpenRouter documents provider-level zero-data-retention and data-collection controls; Cloudflare documents that AI Gateway logs can include request and response payloads and that logging configuration matters.

Ask whether a zero-retention selection covers the gateway, the model provider and every subprocessor. Confirm whether logs are enabled by default, how per-request overrides interact with account policy and whether support personnel can access retained data. A privacy setting on a provider route does not by itself remove gateway-side logs. For regulated workloads, have counsel evaluate the signed terms and the actual deployment rather than relying on marketing labels.

Regional routing narrows the set of eligible endpoints. That can change latency, capacity, model choice and the cost of failover. An in-region deployment might use a different cloud provider, price tier or throughput quota than a global endpoint. Compare equivalent workloads under the same regional restrictions; a global benchmark with a larger pool of providers is not a fair comparator.

Build a total-cost worksheet with gateway subscription, model charges, region-specific processing uplifts if applicable, network transfer, observability retention, operations and the engineering cost of a fallback policy that must refuse service rather than leave the region. Report the percentage of test requests rejected for lack of eligible capacity, not just median latency among successful requests.

Require vendors to complete five rows: gateway payload processing region; upstream inference endpoint region; request/response log location and retention; backup/support/subprocessor processing; and strict fallback behavior. For each row, collect a written statement, configuration reference, observed test result and contractual document where relevant. Mark a missing item as unknown, not as passed.

The purchasing decision then becomes auditable. A managed regional gateway can be a good fit when its boundary and failure behavior match the buyer's requirements. A self-hosted proxy may give more control over one layer but adds operational responsibilities and does not eliminate upstream data transfers. Re-run the evidence matrix whenever a new model, region or logging sink is introduced.