must people think it’s just GPU cost. In practice it’s coordination: model latency variance + queueing + retries under load. You don’t scale linearly, you get c

must people think it’s just GPU cost. In practice it’s coordination: model latency variance + queueing + retries under load. You don’t scale linearly, you get cascading slowdowns.