Enterprise adoption of a new inference platform is decided as much by the review as by the benchmark. This note summarizes how InfraOptimal dedicated endpoints behave today. The authoritative, dated description is our data handling notice; where this note and that page differ, the page wins.
Where requests go.
Managed inference passes through InfraOptimal’s Cloudflare gateway to the endpoint’s backend; some dedicated endpoints route directly to their serving infrastructure. Each deployment’s connection details identify its API and startup mode.
The startup mode is a choice worth making deliberately. Some endpoints offer an immediate startup that answers requests through a third-party model provider’s API while private GPUs load, then moves to the private replica at the same URL and key. On some models that immediate startup is the default, and the data handling notice names the provider each one uses. Private-only startup waits for the private replica instead. A private GPU does not mean the gateway or management service never handles traffic, and the API guide states what each mode does.
What is kept, and for how long.
The managed proxy’s analytics store numerical usage and outcomes, not prompts, completions or tool arguments. Engine and hardware metrics are numerical and expire on rolling windows. Startup and failure diagnostics keep a small machine and configuration snapshot with the deployment and a few bounded engine-log excerpts that expire after 90 days. Known request-log lines and secrets are filtered from those excerpts, but arbitrary engine exception text is not guaranteed to be content-free, and we say so.
Request content
Routed, not kept in analytics
Prompts, completions and tool arguments are not stored in the proxy’s usage analytics.
Numbers
Bounded windows
Usage counts, timings and hardware metrics expire on published rolling windows.
Records
Kept with the account
Account, deployment and billing records remain after Stop; deletion is a separate request.
Private Hugging Face tokens and registry credentials supplied for a deployment are held in restricted control-plane storage with a 30-day expiry. Permissions accepted for any other service, such as longer retention or use for training and evaluation, are not extended to dedicated traffic.
Keys scoped to the job.
There are two kinds of key, and an application only ever needs the narrower one. An account-wide platform key manages deployments and is stored as a hash with the expiry chosen when it was created. A dedicated endpoint key can start its endpoint, run inference and stop it, and nothing else: it cannot change limits, prices or keys, delete the endpoint or launch anything. Coding agents connect over MCP by signing in, without a platform key.
One account for the team.
An account can be shared with teammates through invitation links. Members manage endpoints using the shared balance, admins also manage billing and invitations, and the owner controls platform keys and admin roles. Spending limits on each endpoint, from rate to total budget, keep shared capacity inside the plan.
Failures that explain themselves.
Every launch shows the step it is on, tied to what the provider is actually doing, with how long it has taken against how long it usually takes. Availability is shown before launch wherever it can be known, and is called unknown where it cannot. A failed launch says whether hardware was obtained, whether the engine started and what was charged, and Launch again retries the same configuration once. A hang should never look like progress.
Agreements.
Signing in to an account and an individually agreed service contract are different things; the data handling notice changes neither. For a contract, a provider or location requirement, data access or a deletion request, write to team@velobyte.ai or book a call. We would rather answer a hard question before launch than after.
The reason to run dedicated endpoints is capacity you control and throughput that improves. The reason to trust them is that each of these answers is written down and dated.