17 March 2026
Idle inference capacity after a campaign week
Hari Raya and year-end campaigns leave replicas running. The invoice in the following month is often a campaign story, not a new product story.
Campaign weeks in Malaysia are visible on inference bills if you know which dates marketing actually pushed. We ask for the campaign calendar at intake, not as an afterthought. Without it, a quiet March invoice that is still high looks like a product problem when it is leftover capacity from a February push.
The leftover is usually a second replica, a larger instance type, or a reserved block bought “just for the week” and never released. Cloud consoles do not send a polite reminder that the campaign ended. Someone has to compare the instance list on the last campaign day with the list two weeks later.
We once sat with a team whose transcription feature had been scaled for a retail weekend. The weekend traffic left. The extra GPU stayed through the next invoice cycle because the scale-down runbook lived in a document nobody opened after launch. The findings memo named the instance, the date it should have come down, and the person who now owns that date.
Idle capacity is not always waste. Some teams keep a warm replica because cold starts break a voice feature. That is a decision. It belongs in the memo as a decision, with a monthly cost, so finance can accept it rather than hunt for a bug.
If you are requesting a usage window briefing after a campaign, send the campaign dates with the invoice. We will spend the session on what remained, not on relitigating whether the campaign was a good idea.