Utilities have historically sorted customers into a small number of buckets: residential, commercial, industrial, each with its own tariff structure built around decades of stable, predictable demand patterns. AI training campuses fit none of them, and grid operators are scrambling to invent a category that does.
The problem is variability. A steel mill draws a large, steady load around the clock. A data centre running inference draws a large but comparatively smooth load too. A campus mid-way through training a frontier model can swing its draw by a large fraction of total capacity within seconds, as a training run pauses to checkpoint or a cluster reallocates jobs. Conventional grid infrastructure was not built to absorb swings of that speed at that scale.
Several regional grid operators have begun proposing dedicated interconnection classes for "flexible large load," which would require operators to either smooth their own demand with on-site batteries or accept curtailment during grid stress, in exchange for faster interconnection approval. The trade is explicit: predictability for priority.
The campuses, for their part, are mostly willing to accept the deal. On-site battery buffering was already becoming standard practice to protect training runs from momentary grid disturbances; extending that same buffering to smooth demand for the grid's benefit is a comparatively small additional lift.
What is not yet settled is who pays for the new transmission and substation capacity these campuses require. Utilities want large loads to fund their own dedicated infrastructure upgrades; the campuses argue those upgrades benefit the wider grid too. That argument, playing out in regulatory filings across several jurisdictions, will likely determine how quickly the next wave of training capacity can actually be built.
