Ask most enterprise AI leads who controls their stack, and they'll point to their own architecture diagram — the orchestration layer, the retrieval pipeline, the agents they've wired together. All real, all theirs.

None of it is the control plane. The control plane is pricing, model deprecation schedules, rate limits, and the terms under which your data can or can't be used — and those switches sit with the vendor, not with you. A provider can deprecate the exact model version your evaluation suite was tuned against, change token pricing structure overnight, or update a usage policy that quietly changes what you're allowed to do with your own outputs.

You don't control the control plane. The important switches still sit largely with the vendors.

That's not an argument against using vendor models — almost nobody is training frontier models in-house, and almost nobody should. It's an argument for treating vendor strategy as a first-class part of AI strategy: a real portfolio view across providers, an actual read on switching costs before you need to switch, and a standing watch on the deprecation and policy calendar instead of finding out from a support email.