Inference ops on shared orgs

Inference is quieter than training until it is not. A forgotten replica, a canary that never rolled back, or a weekend demo left running will empty a wallet as surely as a long fine-tune.

Org as the unit of ops

Chassis centers work in organizations. Members share context; the wallet is the spend boundary. That makes on-call simpler: you are not hunting which personal account owns the box that is answering traffic.

• Name instances for the service or experiment they belong to.
• Stop capacity when traffic windows close; do not rely on hope.
• Keep API keys for automation separate from human dashboard sessions.

Hot when it must be

Some serving footprints need to stay warm. That is fine — fund the wallet accordingly and treat the burn rate as a first-class metric. Chassis will not invent fake dashboards; your balance and instance list are the source of truth.