2026: Energy, AI, and the Risk of Overbuilding. 1920s Echoes

A DeltaSignal river. Published by DeltaSignal (AITrailblazer).

River code TF-RIVER-ENERGY-OVERBUILD · Key K-5863f6c13ae756b2 · Page address https://atlas7.app/rivers/energy-overbuild

Thesis

AI's constraint is physical (energy); overbuild appears before anyone agrees it exists

The river at a glance

Founding article, January 5, 2026October 30, 2026Grade 3 validation, October 7, 2026Grade 2 validation, October 22, 20261818 of 26 flagged items are reviewed members (69%); the rest are candidates no reviewer has read in.18 members, reviewed8 candidates flagged, unreviewed

Founding article

2026: Energy, AI, and the Risk of Overbuilding. 1920s Echoes, January 5, 2026. The founding article is the river's source: its claims below are extracted from it, and its break conditions supply the falsifier they are graded against.

Members

Articles read into this river by a reviewer, in article-date order. The founding article is first. Membership is reviewed by hand under the river's protocol; machine-flagged candidates are counted at the bottom of this page and are not members.

Member notes

Notes read into this river by a reviewer, in date order. A note is listed by its opening line.

Claims

Claims extracted from the founding article and approved for publication. Each carries the condition that would break it.

Claim 1 · empirical
In Virginia, data centers consume roughly 26 percent of statewide electricity.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 2 · thesis
AI infrastructure is being built ahead of realized demand at peak enthusiasm, and the overbuild appears before anyone agrees it exists.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 3 · empirical-forecast
Forecasts pushing US AI-related power demand toward 120 GW by the mid-2030s assume uninterrupted adoption and high load factors.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 4 · empirical
Bitcoin mining adds load stress to the same constrained grids, and miners pivoting to AI chase utilization rather than profitability under post-halving margin compression.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 5 · empirical
Global data-center electricity consumption is on track to exceed 1,000 TWh in 2026.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 6 · empirical
Hyperscale data-center power demand grows more than 20 percent in 2025 and may triple by 2030.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 7 · mechanism
If utilization settles near 60-65 percent rather than the high load factors forecasts assume, a meaningful share of the new infrastructure under-earns.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 8 · stated-weights
As published, the framework weights its three scenarios: Boom 45 percent, Bottleneck 30 percent, Bloodbath 25 percent.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 9 · monitoring
Grid interconnection queues reveal buildout delays before management acknowledges them.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 10 · monitoring
Capex relative to revenue shows stress before prices do.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 11 · empirical
In Ireland, data-center demand approaches one-third of national electricity demand by 2026.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Claim 12 · empirical
US data centers accounted for about 4 percent of US electricity consumption in 2024, with demand expected to more than double by the end of the decade.

What would break it: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Validations

Events graded against the framework on primary sources, graded 1 to 3.

Grade 3 · October 7, 2026
EIA October Short-Term Energy Outlook; first validation of record for the framework.

Source: U.S. Energy Information Administration, Short-Term Energy Outlook, October 2026, 2026-10-07.

Falsifier: Capacity utilization sustains above 80 percent as new supply lands, capex converts to free cash flow, and grid buildout scales on usable timelines; the overbuild does not appear. (Drawn from the founding article's own break conditions for the Bottleneck and Boom cases.)

Grade 2 · October 22, 2026
Google and Voltus expand their Bring Your Own Capacity agreement in PJM to a 1 GW virtual power plant (June 2026 start 100 MW; Oct 22 announcement adds 1,070 MW). A signed tenfold expansion of distributed capacity contracted against data-center load, verified against both companies' primary releases.

Source: Voltus, Google-Voltus 1 GW press page, 2026-10-22.

Source: Google, Virtual power plant 1 gigawatt (blog), 2026-10-22.

Source: GlobeNewswire, Google-Voltus expansion release, 2026-10-22 07:30 ET.

Falsifier: PJM fails to accredit the 1,070 MW; the expansion is cancelled or shrunk; metered delivery misses badly; PJM data-center load growth reverses.

Candidates

The archive linker currently flags 8 items (0 articles, 8 notes) as candidates for this river. Candidates are machine-flagged by keyword, reviewed under the protocol below, and are not members. Measured on October 23, 2026, 29.7 percent of the linker's article edges were confirmed substantive membership on blind review, 41.5 percent counting borderline calls.

How this river is kept

Cadence: weekly scan with the flywheel run; validation events recorded as they occur.
Incorporation: linker candidate edges reviewed by Atlas at the weekly scan; founding and curated edges are human-asserted; candidate edges are never presented as membership until reviewed.
Grade rule: framework validations graded 1-3 on primary sources with a named falsifier, per standing framework practice.
Retirement: no new member and no validation for 12 months, or the thesis falsified on its own named falsifier, marks the river lapsed; the record is kept and labeled lapsed, never deleted.
Owners: surveillance, Atlas; editorial, Rada. Protocol adopted 2026-10-23. Status: living.

On the record

This river's names on the graded record: the AI Capex page.