Which LLM Deployment Passes Your Data Residency Rules?
Screen managed inference, customer-run workers, and customer-run model servers against a component-level residency manifest. Cursor Self-Hosted Machines and the OpenAI Agents SDK illustrate why tool, orchestration, and inference locations need separate evidence.
Rank deployments by which restrictions they can satisfy
Rank the architectures against the strictest data-handling rule in the workload. For an approved public-code pilot, a managed environment can be the first candidate to qualify because the vendor can operate both orchestration and execution. If repository checkout and tool execution must stay on customer infrastructure while external inference is approved, put a customer-run worker first. Cursor documents this split for Self-Hosted Machines. For a workload that requires customer-side inference, qualify a customer-run model server and verify each dependency independently. In all three cases, add separate rows for agent state, embeddings, traces, and crash reports. For example, a team may approve sending prompts to a provider but require traces to stay customer-side. A single deployment label cannot express that rule, while a per-component manifest makes it testable.
Choose where Cloud Agents run | Cursor DocsMake the runtime owner prove where session data lands
OpenAI's Agents documentation distinguishes a managed Codex harness from an SDK loop inside your application and direct model calls through the Responses API. Use that distinction to assign an evidence owner to every hop before the pilot. Give a synthetic prompt, retrieval result, and tool result three different marker strings. Exercise request completion and session resume, then inspect the approved request records, session store, and trace export for each marker. Record the destination, component, and observation method. For example, if the tool marker appears in a provider-side session record, the policy needs a row for provider-side session state even when the tool ran locally. Where the provider surface cannot be inspected, log the gap and obtain documentation or contractual evidence for that specific store before treating the placement as verified.
Agents | OpenAI APIDraw the boundary around the model call
Start with one data class and mark where prompts, retrieved passages, tool outputs, credentials, and logs may go. For each destination, record the infrastructure owner, allowed regions, and retention policy. With Cursor Self-Hosted Machines, tool calls run on hardware you control and the agent loop and inference remain in Cursor. Compare that with Cursor-hosted Cloud Agents when your policy permits a managed workspace as well as managed inference.
Illustrative input is an internal repository whose checkout must stay on company hardware, but whose approved excerpts may be sent to a contracted inference service. Test with synthetic files marked for different handling. Verify that the checkout stays on the worker, that permitted excerpts can support the task, and that restricted markers never appear in the request records available for inspection.
Choose where Cloud Agents run | Cursor DocsTest local inference with external access blocked
If external inference is forbidden, qualify NVIDIA AI Enterprise with NIM, Red Hat OpenShift AI, and Ollama on hardware inside the approved boundary. Ask NVIDIA for support for the exact model and GPU, Red Hat for the serving configuration on your cluster version, and the Ollama operator for evidence that identity, logging, and updates meet the policy. Compare them on the same model workload and hardware where supported.
Anthropic's Claude Code deployment guide describes using a proxy for domain allowlists and traffic logging. For an offline acceptance test, also enforce the network boundary outside the application. Start from approved artifacts, restart the service with external routes blocked, and submit a synthetic request. Pass if it completes with local records and no external dependency. For an air gap, repeat the test after importing an update through the approved offline process.
Securely deploying AI agents - Claude Code DocsExercise the hybrid failure path
A hybrid deployment needs a test for what happens when the approved local model is unavailable. Use two synthetic requests, one cleared for a hosted model and one restricted to local inference. Replay both with the local endpoint down. The cleared request can follow its documented route, while the restricted request must fail visibly or queue within the boundary. Inspect request logs and retry behavior for unintended fallback.
For a scoped compute comparison, Runpod lists H100 SXM Clusters at $4.79 per GPU-hour. An illustrative two-GPU, eight-hour trial is $76.64 for that cluster compute line. Verify region, availability, storage, and other charges before comparing it with owned equipment. Use hosted compute for the trial only if the test data and workload are cleared for that environment.
GPU Cloud Pricing | Per-Second H100, A100, RTX | Runpod