A port authority data center schedules heavy training at night when municipal power tariffs and cooling are cheaper. What efficiency practice is this?

A port authority data center that schedules heavy training at night when municipal power tariffs and cooling are cheaper is time-shifting AI workloads to lower-cost power and cooling windows. Efficiency here is operational: the same GPU work costs less and stresses cooling less when run off-peak, without changing model architecture. Domain 2.4 recognizes workload scheduling against tariff and thermal windows as a facility efficiency practice.

Geo-shifting AI workloads to low-latency regions for faster inference is geographic placement for latency, not the off-peak scheduling described; it doesn't shift work to cheaper tariff and cooling windows. Time-shifting AI workloads to midday peaks for grid balancing support moves work into the expensive peak window rather than away from it, so it isn't the tariff-and-cooling time-shift described in the stem. Follow-the-sun AI workloads to track peak solar output across zones chases daytime solar availability across regions, not workload scheduling into local night tariffs and cheaper cooling windows.

Exam caveat: Not every inference SLA can shift to night; the stem's heavy training jobs are the elastic workload. Time-shifting complements, rather than replaces, right-sized power and cooling capacity.

Operational check: encode tariff and cooling-cost windows in the port's job scheduler; verify night jobs still meet thermal limits with cooling in normal night mode; report $/epoch and PUE-adjacent savings so efficiency gains are measured, not assumed.