Building a responsible decentralised infrastructure pilot

The college began with a narrow workload: a private assistant over its own course documentation, used by a small group of staff, plus an overnight batch job producing lecture transcript summaries.

Energy data came from an existing rooftop array with an inverter API and no battery. Compute was a single workstation-class node in a ventilated technician's room, on a circuit reviewed by the site's electrical contractor.

Over six weeks, the shiftable batch workload was moved into the middle of the day. The measured local-energy share of node consumption rose substantially, while the always-on assistant continued to draw from the grid outside generation hours.

The report's conclusions were deliberately modest: local AI was viable and useful at this scale; the largest single win was moving batch work into generation windows; and the next step was a professionally designed battery to shift evening load. No savings were guaranteed and seasonal variation was flagged as the dominant uncertainty.

This is an illustrative composite example used for teaching purposes.