The fight over powering artificial intelligence has narrowed to a question an engineer can answer, and it is being answered in a docket rather than a thread.

The question, as the paper put it on June 27, is whether a data center that plugs straight into a power plant should pay toward the grid everyone else still shares. The Federal Energy Regulatory Commission opened that question formally. Its Federal Register notice convened a commissioner-led technical conference on large loads co-located at generating facilities — the arrangement where a hyperscale data center sits beside a nuclear or gas plant and draws power behind the meter, leaning on the transmission system the rest of the market funds. [1] The notice framed it without slogans: the reliability, cost, and market implications of co-location, and whether a load that relies on the grid for backup should pay toward it.

The precedent, though, is built one grid operator down. PJM, the largest U.S. system operator, runs the interconnection-process reform where these questions become rules. [2] Its service-request process is where a co-location proposal is studied, where the operator decides what "using the grid" means for a customer that claims not to, and where the cost of backup and stability gets assigned to someone. [3] That is allocation work, not applause.

On X, a co-located data center is either the cornerstone of national AI supremacy or a parasite siphoning public power for private compute. Mainstream coverage — Bloomberg, the trade press — reports the demand curve and the blackout warnings. The record asks the prior question both skip: when a gigawatt of load attaches to a single generator, who bears the cost and the reliability risk, and do other customers subsidize it?

The venue matters because the money does. A docket is a receipt. If a developer wants to argue that co-location is efficient and self-contained, the claim belongs in the record, where staff and rivals can test the load-flow assumptions. [1] If a utility argues that co-location offloads cost onto captive ratepayers, that case belongs in the same reform process, where it can be modeled rather than asserted. [2]

This is the divergence the paper keeps. The public argument is whether AI deserves the power. The institutional question is narrower and longer-lived: when a data center co-locates, does it pay its share of the system that keeps it running, or does the cost land on households and factories that never signed up to subsidize a server farm? [3]

The thread will declare a winner tonight. FERC and PJM will decide who pays, and that answer — filed, studied, and appealable — will outlast the thread by a decade. [1][2]

-- THEO KAPLAN, San Francisco