A large data center can pull as much electricity as a small town, and almost every watt of it ends up as heat. We argue about the power these buildings draw and the water they drink. We rarely ask a simpler question: does a data center measurably warm the neighborhood around it? I built DCTII — the Data Center Thermal Impact Index — to answer that with evidence instead of adjectives.
Measuring it from space. Public satellites have quietly recorded the temperature of the ground beneath them for more than a decade. DCTII pulls that record — from NASA's MODIS sensors and the Landsat missions — for 42 data-center sites across six regions: Phoenix, Houston, Central Texas, Northern Virginia, Toronto, and Montreal. The trick is what you compare against. A warehouse district in the Arizona desert is hot all on its own, so I match each site to nearby control zones and subtract what the area would look like with no data center there. What remains is the building's own thermal fingerprint, distilled into a single score from 0 to 100.
Why the night matters. The score leans on nighttime, not daytime. After dark the sun is gone and the extra heat a facility sheds stands out clearly; in daylight it is drowned out. So the index weights four things: the night-time temperature difference (55 percent), the waste heat the building throws off (25 percent), how far that warm halo spreads (10 percent), and how many people live inside it (10 percent).
Predicting a building that doesn't exist yet. Scoring a site that is already running is useful. The more valuable question comes earlier: if we build a 150-megawatt facility here, cooled this way, how much will it warm the area? A companion model — DCTII-Predict — answers before the concrete is poured. Give it a location and a few specs (capacity, cooling type, efficiency, footprint) and it returns the expected night-time warming with an honest range around it.
Keeping the model honest. It is easy to build a model that dazzles on paper and fails in the field. Two guardrails push back. First, I held out an entire region — Montreal, the most climatically unusual — and graded the model on a city it never saw while training. Second, every prediction carries a calibrated confidence interval: when the model claims 90 percent certainty, it is right about 96 percent of the time, and on that held-out city it lands within roughly a quarter of a degree of the truth. Ask it about a place unlike anything it has seen, and it says so instead of bluffing.
What actually drives the heat. The model explains itself, and its reasons match the physics. The strongest driver of night-time warming is the raw waste heat a facility sheds, then how paved-over its surroundings are, then the local climate. Trees and vegetation pull the other way. None of that is a surprise — and that is exactly why I trust the numbers a little more.
What it can't do yet. Daytime warming is still hard. The signal is buried in noise, and I would rather ship nothing than ship a number I don't believe, so the daytime model stays experimental. The night-time model is the one in production.
Why it matters. I recently wrote that the data-center business ultimately comes down to power. This is the other side of that coin. All of that power becomes heat, and the heat lands on real streets and real people. As these buildings grow larger and edge closer to where we live, "how much does this warm my neighborhood?" stops being a rhetorical question. DCTII is one attempt to put a defensible number on the answer — and to hand it to a planner before anything is built.
The whole thing is open source, from the satellite pipeline to the interactive map.
Project links