The Bottleneck Beneath the AI Boom: How a Global Energy Average Hides Where Power, Water, and Chips Run Out

A widely circulated September 2026 finding holds that artificial intelligence consumes approximately 0.4percent of global energy today and will reach only 0.7 percent within three years, a figure offered asevidence that executives and the public have badly overestimated AI's environmental and infrastructurefootprint. This brief tests that finding, not by disputing its arithmetic, which appears sound, but byexamining its denominator. Global energy consumption spans transportation, heating, agriculture, andheavy industry across every economy on Earth. It is not the denominator that determines whether ahospital in California can expand its AI enabled diagnostic capacity, whether a semiconductor fab inArizona can secure power, or whether a health system's electricity costs will double inside a single decade. Measured instead against the denominator that actually governs those outcomes, US data center electricitydemand, the picture changes sharply. The same firm's own utilities research projects data centers reaching9 percent of total US electricity by 2030, more than double today's share, requiring a global buildout itsanalysts price above 2 trillion dollars. The Electric Power Research Institute's own 2030 high case runsstill higher, at 17 percent. This brief documents that internal divergence, then extends the analysis in fivedirections a global energy percentage cannot reach: the geopolitical redistribution of compute capacitytoward jurisdictions with the cheapest energy rather than the most advanced regulatory or democraticinstitutions, the second order effect by which cheaper AI models appear likely to increase rather thandecrease aggregate energy demand, the circular financing structure in which the companies forecastingsoaring AI infrastructure demand are, in a growing number of disclosed transactions, also the companiesfunding the customers generating that demand, the water constraint riding silently alongside electricity thatcan halt a fully powered project on its own, and the return of retired nuclear plants, including the site ofAmerica's worst commercial nuclear accident, to service the same demand curve. None of this is offered as an attack on sound, well documented research. It is offered as a demonstration ofa specific analytical failure mode, the substitution of a global average for a regional reality, that thisinstitute believes is already shaping capital allocation, energy policy, and infrastructure investmentdecisions at a scale that deserves the same primary source scrutiny this institute has applied to statemedical boards and healthcare cloud dependency in its published research. Marcus J. Frazier, Author | Koru Health, a Koru Capital Management Company | Published by the Koru Research Institute | Korucm.com ORCID: 0009-0005-5094-9040