Open-weight models are an ops problem

The case for open-weight models is usually made on cost and control. Both are real. But the moment a second team wants an endpoint, you stop having a model and start having a platform, and platforms have requirements that no model card mentions.

Four questions that decide it

Who is allowed to call this endpoint, and how do you revoke that at 17:00 on a Friday? What happens when one team's batch job consumes the capacity another team's checkout flow depends on? Which exact weights served the request a customer is disputing? And who pays for the GPU hour?

None of these are model questions. They are the reason a promising internal deployment stalls at the pilot stage: the platform team cannot answer them, so they cannot sign off.

Treat the endpoint as the unit

The abstraction that works is not the model, it is the endpoint: a named, versioned, quota-bound thing with an owner, a rollout history, and a bill attached. Everything else — which provider, which accelerator, which serving runtime — becomes an implementation detail you can change without telling every caller.