Announced vs. Deliverable AI Power Demand

Forecasts of AI-driven electricity demand treat announced capacity as if it were firm on both sides of the meter: the generation announced to serve the load, and the data-centre load itself. This working paper deflates both using the grid operators' own resolved data. Supply side (carried from v0.1): PJM's resolved serial queue (8,253 generation projects, about 837 GW announced) shows only about 16-25 per cent of announced MW reach commercial operation, 22.8 per cent on the mature 2010-2017 cohort, with the steepest attrition at the System Impact Study where each project learns its share of network-upgrade costs. Load side (new in v1): the deflation is performed and published by the operators themselves, so this note packages it rather than re-deriving it. PJM has cut its 2030 large-load request from about 60 GW submitted to about 38 GW accepted, of which only the about 32 GW that is firm (backed by an Electric Service Obligation or Construction Commitment) counts toward capacity; PJM's documented worked example survives at about 48 per cent of requested capacity. In ERCOT, whose about 226 GW queue was padded by low historical entry barriers, the one quantity both operators state in comparable terms is measured rather than assumed, and undercuts PJM: ERCOT's metered per-site peak consumption is 49.8 per cent of requested MW across non-crypto data centres built 2022-2024, against the 70 per cent factor PJM applies forward. ERCOT's commissioned 2030 scenario is lower again (about 35 GW of 226 GW), reported as context on queue padding. Two independent operators, using different methods, show that requested queues far exceed deliverable demand, and both now gate the forecast on an executed agreement. The contribution is the packaging: an independent referee, plus a two-rail deliverable envelope (about 20 per cent energisation floor from generation-queue completion, to about 48-63 per cent operator-accepted ceiling) that names every assumption. The output is a band with its softest input marked, letting a planner treat any announced headline as a scenario, not a baseline, which is a ratepayer-protection argument as much as an analytical one. Methods precedent: the bot-energy paper (concept DOI 10.5281/zenodo.20512703). This is independent analysis, not investment advice. Prepared with AI assistance (Anthropic Opus 4.8-5.0 and Google Gemini 3.7); both vendors maintain commercial relations to the AI data-centre buildout and electricity demand examined, disclosed for completeness. No third party reviewed, funded, or directed this work.