2026-01-15 · Leo Zhang
Chargeback sheets GPUs deserve
infra · finance · gpu
Shared clusters rot when teams cannot see cost per experiment. We walk through a chargeback sheet that ties Slurm job IDs to ledger codes, including partial runs that aborted mid-epoch.
Students import their own anonymized numbers; mentors poke holes gently. The emotional arc moves from defensiveness to “oh, that job should have been tagged research, not production.”
We end with a facilitated conversation about fairness: some teams need burst capacity for customer incidents. The policy we draft acknowledges that reality instead of pretending averages solve politics.