ZCR-2026-006 · CUSTOM REPORT

Energy for Compute The attention window opened — but not where the headlines put it. Eight terms, one deflated threshold, and the difference between vocabulary that spikes and vocabulary that holds.

OBSERVATION CUT-OFF SEP 2026 · ALL VALUES DEFLATED BY ZBI 1.31 · THREE SOURCES
5/8
terms crossing the structural threshold
14
consecutive months still above it
×22
deflated growth, data center power
1.31
baseline drift factor removed

Is the grid constraint a theme or a moment?

Every investor looking at the compute buildout now asks the same thing: the electricity problem is obviously real, but is it a durable allocation theme or a news cycle that has already peaked?

Attention data can answer that, on one condition — that the measuring instrument has not moved. In September 2026 we established that it had. Every figure in this report is corrected for it first, because without the correction the answer comes out wrong in a specific and flattering direction.

The ruler moved in August 2025

Five semantically inert control terms — iPhone, pizza recipe, banana bread, washing machine repair, driving licence — rise together from August 2025 by a median factor of 1.31. Nothing plausible moves those five in concert. Google Trends reports share of search, not volume, and the share base is contracting.

CONTROL BASKET · 5 INERT TERMS · INDEXED, 100 = 2023–24 AVERAGE
80100120140160 Aug 2025 148 2024-012024-072025-012025-072026-012026-07
control basket2023–24 basestructural break

The basket holds flat around 100 for thirty-one months, then steps up and stays up. It sits at 148 today. That step is not the world becoming more interested in banana bread: it is the denominator shrinking underneath every Google Trends series simultaneously.

Consequence for this report. A term sitting at 19 before the break crosses the 25 threshold after it without anything happening in the world. Every month from August 2025 onward is divided by 1.31 before being compared with the threshold; earlier months are left untouched, because the drift is dated. In this cluster the correction changes one classification — grid interconnection queue reaches twelve months uncorrected and six once deflated — and it lowers every current reading by roughly a quarter.

Independent cross-check, on a sensor that counts volume rather than share: total English Wikipedia traffic fell 7.0% year on year over the same window. One measure of share rising while the measure of volume falls is the signature of a denominator effect, not of demand.

Narrative vocabulary spiked. Procurement vocabulary held.

Zorthex classifies a phenomenon as structural when it holds at or above the threshold for twelve consecutive months. Applied to the eight terms of the grid-and-compute cluster, deflated, the cluster splits cleanly in two — and the split is not where the coverage would put it.

CONSECUTIVE MONTHS AT OR ABOVE THE DEFLATED THRESHOLD
12 months — STRUCTURAL behind the meter14microgrid14onsite power14data center power13stranded silicon13energy certainty7utility sovereignty7grid interconnection queue6
crossed 12 monthsdid not cross

The terms that describe how power is actually bought and connected — behind the meter, microgrid, onsite power — crossed the threshold and are still above it, fourteen months later and counting. The terms the trade press invented to describe the same story — energy certainty, utility sovereignty — reached seven months and fell back below.

TermMonths ≥ thresholdToday, deflatedStatus
behind the meter14 · holding35.1STRUCTURAL
microgrid14 · holding33.6STRUCTURAL
onsite power14 · holding29.0STRUCTURAL
data center power13 · ended22.9STRUCTURAL
stranded silicon13 · ended12.2STRUCTURAL
energy certainty722.1below
utility sovereignty78.4below
grid interconnection queue67.6below

data center power is the instructive case. It grew ×22 after deflation — the largest move in the dataset this revision — crossed the threshold for thirteen months, peaked in May 2026, and has since retraced roughly seventy per cent. The phrase did its work and is being retired. The activity it named has not stopped.

What the volume and community measures say

Wikipedia — the volume measure

Search share tells you what people are asking. It does not tell you whether the subject is consolidating. For that, Zorthex reads a sensor that counts rather than ranks: Wikipedia pageviews, normalised for the 7.0% systemic decline.

The Data center article rises sharply — but by how much depends entirely on what it is measured against, and that choice deserves to be shown rather than buried.

Normalising a technical article against site-wide Wikipedia traffic is too coarse. The 7.0% site-wide decline is not uniform: descriptive articles about physical infrastructure are more exposed to the migration of informational queries than the average page. So the article was also measured against a matched basket — five descriptive articles about physical infrastructure, outside computing, chosen to share its function and its exposure rather than its subject: Sewage treatment, Blast furnace, Wind turbine, Oil refinery, Desalination. That basket fell by a median of 27.6% over the same window, four times the site-wide rate.

NormaliserData center, year on yearDirection
none — raw pageviews+99.6%up
site-wide traffic (−7.0%)+115%up
matched basket (−27.6%, n=5)+176%up

Three denominators, one direction. The direction is robust; the magnitude is not, and this report claims only the first. Note which way the correction runs: the more rigorous normaliser produces the larger figure, which means the site-wide estimate was understating the signal rather than flattering it.

The matched basket carries its own limits and they are not small. Five articles is a thin sample; the dispersion runs from −38.0% to +13.7%, and Desalination at +13.7% plausibly has a cycle of its own. The selection of controls is a judgement, and a different basket would give a different number. It is offered as a second reading, not as a replacement.

Reddit — the community measure

The third source is qualitative by design and is never converted into a number or used to compute L. It answers a different question: is anyone actually discussing this, and who? A phenomenon with high search interest and no community may be a media artefact; one with low search interest and an active community may be early.

Sampled through Reddit's public search feed, which returns the most recent matches rather than a count, the cluster reads as follows.

TermRecent matches spanWhere the discussion sitsReading
onsite power3 monthsbloomenergycorp · EducatedInvestingmature
behind the meter12 monthssolarmature
microgrid12 monthssolar · climatechangemature
data center power12 monthsmixed, incl. investor subsmature
grid interconnection queue4 monthslow-quality / aggregator subsnascent
energy certainty8 matches since 2023scatteredabsent
utility sovereignty2 matches, both 2026—absent

The third source confirms the finding from an angle the first two cannot reach. The procurement terms are not merely searched: they are discussed, in real technical and investor communities, continuously. The two narrative terms have no community at all — it is not that attention faded, it is that a discussion never formed around them. A phrase can generate search interest for seven months without a single durable thread behind it, and that is what these two did.

One term must be flagged. On Reddit, stranded silicon appears almost entirely in hobby-electronics communities discussing silicon as a material, not idle compute capacity. The phrase is ambiguous, which means its Google Trends series is probably contaminated by the unrelated sense. It is reported above for completeness and should not be relied on until disambiguated — a check the first two sources could not have produced on their own.

Sensor agreement is the discriminator between a sustained break and an event-driven one. Where the sources diverge, the attention is narrative. Where they converge, something underneath is consolidating. On the procurement terms all three converge; on the narrative terms all three are absent or falling.

Does the finding survive the range of the correction?

The deflation factor is a median of five control terms spanning 1.25 to 1.35. A conclusion that depends on which end of that range is chosen is not a conclusion. Every term was therefore recomputed at both extremes.

TermMonths ≥ threshold at 1.25 / 1.31 / 1.35Today at 1.25 → 1.35Result
behind the meter14 · 14 · 1436.8 → 34.1stable
microgrid14 · 14 · 1435.2 → 32.6stable
onsite power14 · 14 · 1430.4 → 28.1stable
stranded silicon13 · 13 · 1312.8 → 11.9stable
energy certainty7 · 7 · 623.2 → 21.5stable
utility sovereignty7 · 7 · 78.8 → 8.1stable
data center power13 · 13 · 624.0 → 22.2unstable
grid interconnection queue12 · 6 · 68.0 → 7.4unstable

Six of eight terms hold their classification across the whole range. No current value crosses the threshold as a function of the factor: 29.0 stays above at every factor, 22.9 stays below at every factor. The grey zone one might expect around the threshold does not materialise here.

The two terms that are sensitive are not the ones carrying the conclusion. data center power is already described above as a run that has ended and retraced; grid interconnection queue is already excluded for sparse history. The three procurement terms hold fourteen months at every factor in the range, and the two narrative terms stay below at every factor. The finding is robust to the correction it depends on.

Where the window is, how wide, what type

Where

Not in "AI infrastructure", which is saturated and carries no information. The open window is one layer below, in the language of power procurement: behind-the-meter generation, microgrids, onsite generation. That vocabulary crossed the threshold and is holding.

How wide

Fourteen months and open. Three terms remain above the deflated threshold with no sign of retreat, against a cluster average that has already turned down. In this dataset a structural run that is still extending at fourteen months has historically had further to go, though that is a base rate and not a forecast.

What type

Sustained, not event-driven — on the evidence of sensor convergence, not of magnitude. The distinction matters because the magnitude is exactly what the baseline drift inflates, and what a reader without the correction would have used to reach the opposite conclusion about the retreating terms.

The diagnostic statement. The attention window in energy-for-compute is open in procurement vocabulary and closing in narrative vocabulary. A reader tracking the phrases that appear in headlines is tracking the half of this theme that has already peaked.

What this report does not claim

Replication

Google Trends · worldwide · monthly · 2004–2026 · 13 series downloaded 20 Sep 2026, double-verified
Wikimedia REST API · en.wikipedia pageviews · per-article and total · Sep 2023 – Aug 2026
Control basket, Google Trends · iPhone · pizza recipe · banana bread · washing machine repair · driving licence
Matched basket, Wikipedia · Sewage treatment · Blast furnace · Wind turbine · Oil refinery · Desalination
Reddit · public search feed · qualitative reading · sampled 27 Sep 2026
Dataset · zorthex.com/dataset · 71 phenomena · Zenodo DOI 10.5281/zenodo.20270574
Methodology · zorthex.com/methodology · framework v2.0

All snapshots are locked and timestamped by the author and available on request. The same public data can be downloaded directly from Google Trends and the Wikimedia API by anyone wishing to reproduce these figures.