An LLM is a workload, not a feature

Once a model leaves a notebook, it becomes a service: long-lived, hardware-hungry, and shared across users. Kubernetes already runs services like that — databases, queues, payment systems. The job is to teach it the new shape of this one.