Research Note: E075 Demand Outlook and Publication Boundary

Research Note: E075 Demand Outlook and Publication Boundary

Question

The July 2025 recording used dramatic comparisons to make AI infrastructure feel physical. The revised Episode Story needs to preserve Dalton's concern while replacing arithmetic shortcuts and unsourced site claims with current estimates, stated boundaries, and a clear distinction between national scale and local constraint.

Current national evidence

Lawrence Berkeley National Laboratory published its United States Data Center Energy Usage Report: 2025 Update in June 2026. Its reference case estimates 649 terawatt-hours of US data-center electricity consumption in 2030, or 11.8 percent of total US electricity. Its compound uncertainty range is 521 to 843 terawatt-hours, or 9.5 to 15.3 percent.

Those values are modeled scenarios, not a meter reading or promise. The estimate uses planned equipment shipments, per-device energy models, cooling simulations, facility types, and locations. Sensitivity cases change chip shipments, AI-chip lifetime, idle power, and server utilization. The model therefore supports a statement about plausible national demand under explicit assumptions. It does not forecast a particular campus.

The 2024 LBNL report remains useful for history and facility modeling. It estimated that US data centers consumed 176 terawatt-hours in 2023, equal to 4.4 percent of US electricity. Its 2028 outlook has been superseded by the 2026 update for current forward-looking copy.

The International Energy Agency supplies global context and separates data-center demand from total electricity-system growth. IEA scenarios should not be mixed with LBNL's US model as though the geography, forecast year, and method were the same.

Claims that do not survive

The current public episode page says 950 terawatt-hours could power California for nearly 30 years. It compares one forecasted year of global data-center consumption with an erroneous amount of state consumption. The claim should be removed, not softened.

The public page also says a mid-to-large data center uses five million gallons of water every day and that 80 percent of purified drinking water evaporates. Neither is a universal facility value. Cooling design, climate, workload, operating schedule, water source, reuse, and the reporting boundary change the result.

The Georgia household-bill story and its implied cause are not established by a utility bill shown in a video. The Mansfield well-water account does not establish whether construction, groundwater withdrawal, another source, or no data-center activity caused the reported condition.

The Memphis discussion should not be republished as a medical or causal finding. A Shelby County air-permit docket contained serious disagreement about gas-turbine count, classification, and permit scope, but advocacy comments are not an agency enforcement finding. Current public copy can explain why permit identity, equipment inventory, emissions limits, and compliance records matter without declaring that a facility caused a person's illness.

Editorial use

The Episode Story can say that Dalton was directionally right about one thing: digital demand becomes a physical infrastructure decision somewhere. It must also say that the original scale comparisons and causal local claims were not ready for publication.

The revised story should move from the national demand range into the local mechanisms that determine impact. Those mechanisms include interconnection, generation, transmission, substations, rate design, water service, land use, air permits, noise, incentives, emergency services, expansion, and exit risk.

The existing route is https://www.daltonanderson.net/venture-step/the-hidden-costs-of-ai-data-centers-power-water-strain/. It returned HTTP 200 on July 27, 2026. Preserve the route and original July 29, 2025 publication date. Add a visible July 2026 correction note and replace the current article in full only after explicit publication approval.