The AI boom has a heat problem nobody budgeted for.
A new preprint study out of the University of Cambridge has put a number on something researchers suspected but couldn’t fully quantify: AI data centers aren’t just consuming energy at a historically unprecedented rate — they’re physically warming the land around them. Land surface temperatures rise by an average of 2°C after a hyperscale AI data center starts operations, creating what the researchers call the “data heat island effect.” That’s not a global average getting nudged. That’s a neighborhood getting cooked.
The 6-Mile Radius Nobody Told You About
The researchers filtered out seasonal trends, global warming trajectories, and other environmental noise — isolating the data center footprint specifically. In extreme cases, nearby temperatures climbed as high as 16.4 degrees Fahrenheit. The heat doesn’t stay put, either: temperature increases were recorded up to 6.2 miles from the facilities.
Picture a quiet street in Santa Clara, three blocks from a hyperscale campus. The hum of industrial cooling fans rolls in around midnight and never quite stops. In summer, the backyard feels a few degrees warmer than it used to — not dramatically, not unmistakably, but persistently. The air conditioning runs a little longer. The electricity bill nudges up. Nobody can point to a single cause, because nobody told them to look at the warehouse-sized building with no windows down the road. That’s the data heat island in practice. Diffuse, cumulative, and invisible on any official map.
The study focused on more than 6,000 hyperscale data centers globally, deliberately filtering out dense urban cores where manufacturing and residential heat would muddy the picture. What remained was a clean signal — AI infrastructure, rewriting the microclimate of its immediate surroundings.
The Tradeoff Industry Doesn’t Want to Talk About
There’s a reason the heat island effect is getting worse, and it runs directly through the industry’s preferred sustainability metric: water.
Data centers face a fundamental tradeoff between air cooling and evaporative cooling. Evaporative cooling is more energy-efficient but consumes large volumes of water — up to 85% of which evaporates and never returns to the local supply. Air cooling eliminates direct water consumption, but the tradeoff is significantly higher electrical energy demand, especially during hot weather.
In water-stressed regions, operators deliberately choose air cooling to avoid depleting local water resources. But air cooling pushes more heat into the surrounding environment rather than absorbing it through evaporation.
So the industry’s response to one environmental complaint — water depletion — quietly amplifies another: localized warming. Communities in arid regions like the Bay Area get the worst of both worlds. Less cooling water means more heat is displaced into the air. The data heat island effect isn’t a bug. It’s a predictable consequence of the tradeoffs being made right now, at scale.
340 Million People, One Footnote
The researchers estimated that more than 340 million people worldwide could be affected by the data heat island effect. That figure is buried in a preprint paper. It hasn’t entered any regulatory framework, no federal siting guidance, no environmental impact assessment standard.
The study’s lead researcher, Giorgia Marinoni, was direct about what she sees happening: “The ‘rush for AI-gold’ appears to be overriding good practice and systemic thinking, and is developing far more rapidly than any broader, more sustainable systems.”
She’s not wrong. Capital expenditures for data center facilities are projected to reach $760 billion in 2026, up from $450 billion the year before. Alphabet alone plans to invest $185 billion in AI infrastructure — spending that exceeds the entire GDP of Sweden. Every dollar of that spend is another facility, another heat source, another 6-mile radius with no disclosure requirement attached. The Property Market Hasn’t Priced This In Yet
The real estate industry is only beginning to grapple with data center proximity — and mostly from the wrong angle. Industry-commissioned studies have repeatedly argued that proximity to a data center doesn’t statistically suppress residential property values. Those studies measured noise and aesthetics. None of them measured a 2°C temperature increase that extends over six miles and amplifies heat wave mortality risk.
In Chandler, Arizona, residents near a data center have reported constant low-frequency humming that caused headaches, vertigo, and sleep disturbances — and the city council ultimately voted against a new proposed facility in 2025 after years of community pushback. Noise is visible, recordable, and litigable. Heat creep is none of those things — yet. Once insurance actuaries start incorporating localized temperature elevation data into their models, the liability picture changes fast.
What “Fixing It” Actually Requires
The research team proposed a two-track solution. The honest assessment: neither track moves fast enough.
Hardware: Advances in semiconductor efficiency — next-generation ASICs and post-Blackwell GPU architectures — could reduce heat output at the source. Liquid and immersion cooling systems are becoming standard for AI workloads in 2026, with cooling already accounting for roughly 40% of total energy use in modern facilities. Immersion reduces operational heat displacement, but the physics of gigawatt-scale compute still puts enormous thermal load into the surrounding environment.
Software: AI developers could treat energy efficiency as a core design constraint rather than a compliance checkbox. Today, it’s neither. The incentive structure rewards performance per dollar, not performance per BTU.
The deeper fix is situating governance. Where planners build these facilities—and whether they account for the cumulative thermal impact on surrounding communities—remains a regulatory gap that existing environmental review processes fail to close. Environmental Impact Assessments did not anticipate data heat islands. Zoning boards cannot model 6-mile thermal radii. And the communities living inside that radius didn’t get a vote on the infrastructure expansion — they’re just living with the temperature.
The AI gold rush has a shadow. It’s 6.2 miles wide, it’s warming, and nobody’s mapped it onto a planning document yet.
Related: Industrial Policy for the Intelligence Age: Compute vs. Cash Debate
