What does the SIREN infrastructure index read?

Which datacentre regions does Watts cover?The corridors

Nine places where the American build-out actually is, and what each of the three pillars can and cannot see there. 3 of 9 have a free real-time demand feed. The other 6 say why not, in their own cell, rather than leaving a blank that reads as an oversight.

Driest gauge is the lowest of the corridor's gauges as a percentage of that gauge's own long-run median for the same day of the year, named in the cell — the minimum rather than the mean, because a cluster is constrained by its tightest water source and because averaging a creek in flood with a river at normal describes neither. Dev is the same number as a signed deviation. D2+ is the largest share of a corridor state's area the US Drought Monitor puts in severe drought or worse. Signal counts how many of the corridor's measurable series sit outside their band and names which, where the bands are ±25% of the median for a river and 20% of area for drought. Those two numbers are choices, not findings, and they are printed here so a reader can disagree with them. The river band is two-sided: a river far above its median is as unusual as one far below, and usually means rainfall rather than anything to do with a datacentre, so the cell says which side it fell on. A grid series joins the count only once it has a baseline to be unusual against — until then it reads awaiting, which is a different state from normal, and never a zero.

What does overnight grid load show?The grid

Every grid has a daily minimum — the hour when air conditioning, offices and traffic have all gone home. What is still drawing power at four in the morning is the load that never stops, and a datacentre is one of very few large loads that runs flat out at that hour. The overnight floor is the cleanest public proxy for always-on industrial demand that exists without a utility contract. It is also moved by smelters, cold snaps and pumping, which is why New York and California are here as reference regions rather than as subjects.

What does river flow at the gauges show?The water

Cooling a datacentre consumes water, and the argument about how much is one of the loudest live public fights about this industry. Nobody publishes the consumption. What is published, free and keyless, is the state of the water the clusters have to draw on: the flow in the river beside them, measured against the median for that same calendar date over that gauge’s whole record — 127 years at Austin, 116 at Leesburg. The denominator is the Geological Survey’s, not ours.

Which filings mention datacentres?The build-out

Power and water are what a training cluster consumes. This is what it costs, and it is the one part of the story American public companies are legally obliged to write down. A company can decline to say how large its next model is. It cannot decline to tell its shareholders it has committed to buildings full of racks, or to sign a decade of electricity without the contract appearing in a filing.

What state is every Watts source in?

Three states, kept apart everywhere they appear. ●LIVE the feed answered and there is enough baseline to score it. ◐AWAITING BASELINE the feed answered, the number is real and printed, and there is not yet enough history to say where it sits — it is in no average on this page. ✕DARK the fetch failed; the reason is printed under the row and nothing is imputed in its place. Merging the second into the third would claim an outage that is not happening, which is the most common quiet lie in this whole category.

How is the Watts number made?

1. Each source returns one scalar where higher always means more activity. A source that naturally points the other way is negated inside its own adapter, so nothing downstream has to remember which is which.
2. Normalise. Empirical percentile against that source’s baseline, then inverse normal CDF, then S = 50 + 12.5z, clamped to 0–100. This is the same code the main SIREN index uses, imported rather than reimplemented.
3. The baseline is the feed’s own history where the feed has one — five years of weekly Drought Monitor maps, three years of daily USGS values, five years of Federal Register publication dates. Where a feed publishes only today, this collector accumulates one reading per day until it has 30 of them, and the source reads awaiting baseline in the meantime.
4. Pillar score is the mean of its live sources. A pillar with no live source has no score; it is never zero and never carried forward.
5. Composite is 0.7 * mean(live pillars) + 0.3 * max(live pillars) over live pillars. The max term is why one pillar at the top of its record moves the number even when the others are ordinary.
6. Level from the composite on fixed bands, with a 3-point Schmitt deadband, a 6-hour minimum between changes, one step per change, a 2-of-3 pillar quorum, and a freeze on any change while a pillar is genuinely dark.

Two of the main index’s six anti-flap layers are deliberately absent here and are named rather than skipped: the dual-window dwell needs a dense observation history this index does not have — half its sources move once a day and one moves once a week — and the four-hour post-change lock is subsumed by the longer minimum interval.

Every number on this page is in the JSON, including each source’s baseline span, its percentile, and the raw reading behind it. The last build fired the rule none.