Every press release about a new AI cluster comes wrapped in the same fog: a billion-dollar number, a vague "multi-year" timeline, and a map pin that may or may not mean anything. The models are confident. The renderings are gorgeous. And none of it tells you whether the steel is actually in the ground.
Electricity does.
You cannot train a frontier model on a slide deck. AI data centers are, before they are anything else, an enormous and growing electrical load — racks of accelerators that pull power 24 hours a day, weekends included, weather be damned. That makes the power grid one of the few places where the AI buildout has to show up physically, in megawatt-hours, whether or not a company wants to talk about it. The grid doesn't lie. It just meters.
So we started watching the meter.
Why electricity demand is a leading indicator
Most of the signals investors use to track the AI boom are lagging. Earnings land a quarter late. Capex guidance is a promise, not a measurement. Sell-side models inherit their assumptions from the same press releases everyone else read.
Electricity demand sits upstream of all of it. Before a data center books revenue, before it shows up in a hyperscaler's depreciation schedule, it has to be energized — and energizing a few hundred megawatts of compute is not subtle on a regional grid. The load arrives, the utility serves it, and the federal government records it. That record is public, daily, and effectively impossible to spin.
It's the difference between a team's press conference and the box score. One is narrative. The other is what actually happened on the floor.
How we measure it (in plain English)
Raw electricity demand is noisy. It spikes in a July heat wave, sags on a mild Sunday, and swings with the seasons — none of which has anything to do with AI. If you just watched the raw line, you'd mistake every hot week for a data-center boom.
So we don't watch the raw line. We do two things first.
Weekday adjustment. Grids run hotter on a Tuesday than a Sunday. We account for the day-of-week pattern so a normal weekly rhythm doesn't masquerade as growth.
Residual z-score. For each region we build a trailing baseline of what its demand should be, then measure how far today sits above or below that baseline — expressed as a z-score, a count of standard deviations from normal. A z of 0 is a perfectly ordinary day. A z of 2 means the grid is running about two standard deviations hotter than its own recent normal. We flag any region at |z| ≥ 2 as an anomaly and |z| ≥ 3 as extreme.
The point of all this: a high z is the grid running hotter than its own normal without a weather or weekday excuse. That residual heat is the closest thing we have to a fingerprint for new, always-on industrial load — the kind a data center brings. It's not proof. It's a tell.
The underlying data is the EIA-930 Hourly Electric Grid Monitor, the US Energy Information Administration's feed of regional demand, rolled up to daily totals across 13 grid regions.
What the grid shows right now
As of June 13, 2026, the Lower-48 as a whole is already flashing: total US demand is up 6.0% over 30 days and 14.1% over 90 days, with a 90-day z of 2.8 — the national grid is running meaningfully hotter than its own baseline.
But the national number hides the story. The interesting part is where the heat is concentrated. Here are the regions currently registering as anomalies, ranked by how far above baseline they're running:
| Region | 30d trend | 30d z | 90d trend | 90d z | Flag |
|---|---|---|---|---|---|
| Carolinas | +25.4% | 3.3 | +41.3% | 4.0 | Extreme |
| New England | +29.3% | 3.2 | +34.3% | 4.0 | Extreme |
| California | +20.1% | 3.8 | +25.5% | 3.9 | Extreme |
| Southwest | +28.3% | 3.1 | +41.6% | 3.4 | Extreme |
| New York | +21.1% | 2.3 | +28.2% | 3.4 | Anomaly |
| Mid-Atlantic (PJM) | +13.2% | 2.0 | +21.2% | 3.1 | Anomaly |
| Northwest | +6.8% | 2.5 | +8.3% | 2.1 | Anomaly |
Seven of thirteen regions are running hot, and four are extreme — three-plus standard deviations above their own normal. The Carolinas and New England are both posting a 90-day z of 4.0, with the Carolinas up 41.3% over 90 days and the Southwest right behind at 41.6%. These are not rounding errors. A grid does not casually run four standard deviations above baseline.
A few of these names will surprise no one. The Mid-Atlantic region is PJM — literally nicknamed "Data-Center Alley" for the Northern Virginia corridor that hosts the densest concentration of server farms on earth — and it's flashing a 90-day z of 3.1 on a +21.2% trend. That's the control group lighting up exactly where you'd expect.
The more interesting tells are the ones that aren't obvious. The Carolinas and the Southwest, both up roughly 41% over 90 days, are precisely the kind of cheap-land, cheap-power, tax-friendly geographies where the next wave of buildout has been rumored to land. The grid is registering it before the ribbon-cuttings.
What it means for the buildout
If you're trying to map where AI infrastructure dollars are actually flowing — as opposed to where they've been announced — this is a map. The regions running hottest are the regions absorbing new load fastest, and new always-on load at this scale has a short list of plausible culprits, with compute near the top.
That has obvious gravity for the broad power-and-grid theme: the regulated utilities that have to serve this load, the independent power producers selling into it, the data-center REITs leasing the buildings, and the grid-equipment makers selling the transformers and switchgear to connect it all. We're not naming tickers and we're not telling you what to do. We're telling you where the electrons are going, which is a fact, and facts are a better starting point than narratives.
The signal is also a useful skeptic. When a region's grid isn't heating up despite a splashy announcement, that gap is worth a second look. Texas (ERCOT), for all its data-center headlines, is sitting at a 30-day z of just 1.7 — below our anomaly threshold. Big plans, ordinary grid, so far. The meter hasn't caught up to the marketing yet. Maybe it will. Maybe the marketing was early.
The caveats (because the grid tells the truth, not the whole truth)
Read this honestly or don't read it at all.
Demand is not solely attributable to data centers. A hot region could be absorbing a new factory, an electrification push, population growth, reshored manufacturing, or a genuine weather quirk our model didn't fully scrub. High residual demand is consistent with a data-center surge; it doesn't prove one. It's a flashlight, not a courtroom.
Weather adjustment is good, not perfect. We adjust for weekday patterns and measure against a trailing baseline, but we don't model temperature directly. An unusual stretch of heat or cold can leak into the residual. Treat single-region spikes with more suspicion than broad, multi-region, multi-window heat — which is exactly why the national z of 2.8 and the cluster of four extreme regions together are more convincing than any one number alone.
It's a daily snapshot, not a verdict. These z-scores move. A region can cool off next week. The value isn't the single reading; it's the regime — and right now the regime is hot, broadly, across windows.
The honest summary: the grid is the best near-real-time, hard-to-fake leading indicator we have for the AI buildout, and right now it's running hot in more places than the headlines would suggest. That's worth knowing before the earnings calls catch up.
See it for yourself
We track all 13 regions live and flag the anomalies as they cross threshold. The map updates as the EIA does.
→ See which grids are running hot right now at standardpoorly.com/grid
Not investment advice.