Three different clocks
A GPU-hour records one allocated device for one hour. It does not establish how much useful work that device completed. Queue time, data preparation, restart overhead and invalid outputs can all change the relationship between allocation and an accepted result. Even a busy device can be doing work that will later be discarded.
Keep your measurement boundary
A training team might track accepted optimiser steps per wall-clock hour. A serving team might track outputs that meet both quality and latency targets. These measures are workload-specific. Device utilisation, training goodput and accepted-output share are different quantities; the planning tools deliberately do not label them as interchangeable.
What industry platforms are addressing
CoreWeave describes SUNK as a way to connect training workflows, scheduling and infrastructure health. Its April 2026 expansion also addresses consistency across environments. These are examples of the operating work surrounding a cluster, rather than evidence that every deployment achieves the same improvement.
A useful next experiment
Record a small representative run with the model version, dataset, precision, number of devices, storage path and restart policy. Measure total elapsed time and accepted progress. Repeat after changing one constraint. Use the Capacity & Energy tool to document the assumed productive share, then replace that assumption with evidence. A percentage in the tool remains an accounting approximation, not a benchmark.