Are Data Centers Running Servers Hot to Save Water — and Burning Out the Chips?

Are they really cranking the heat to save water?

Here's what sent me down this hole: a comment claiming operators shove intake air up to 95–105°F to cut their water bill, and that it's quietly chewing through hardware. So I went looking for who actually does that.

The recommended ceiling for the air going into a server — set by ASHRAE, the engineering group that writes the thermal rules — tops out around 80.6°F. That's a warm office. The gear is *allowed* to run much hotter; the top equipment class is rated to 113°F. But that higher number is meant for short stretches, not a permanent setting.

What I couldn't find was a named operator confirming they run intake air that hot around the clock. What I did find is the real version of the story, and it's about water, not air. Schneider Electric modeled a big AI site and found that raising the cooling *water* from about 90°F to 113°F cut its water-use number by roughly two-thirds in Dallas and more than 80% in Paris. Warmer coolant lets large dry radiators shed heat straight into the outside air — the way your car radiator handles a traffic jam, instead of you cracking a cooler of ice over the engine — so the site barely evaporates any water at all.

Does running hot actually wear out the chips?

This is the part the comment got half-right. Heat does age electronics — nobody argues that. An ASHRAE engineer in that same report said it plainly: silicon and the power parts around it can age faster when it runs hotter. The old rule chip people quote is brutal — every 18°F warmer, roughly, halves a part's lifespan.

But when researchers went looking at real failure data, the doom didn't show up the way the rule predicts. A University of Toronto team dug through records from Google, Los Alamos, and others, and below about 122°F inside the device, failures climbed with heat in a straight line — not an exploding curve. For memory and for whole-node outages, they found basically no link to temperature at all.

So the chips do wear faster when you run them warm. Just nowhere near the 'you're melting your own investment' speed the viral version implies. Which left me with a bigger question: if it's not really about the hardware, why do it?

The catch nobody puts on the brochure

Here's the trade the water-saving pitch skips over. Almost every comment I read eventually circled the same question: if they're not evaporating water, aren't they just burning more electricity to chill it instead? Mostly, yeah. You can push a site's on-paper water use down to nearly nothing — but you've shoved the cost onto the power bill, and onto whatever power plant is feeding it, which usually drinks water of its own.

And the raw numbers stay staggering either way. Google reported using about 7.2 billion gallons of freshwater in 2024. A researcher at UC Riverside estimated that a single gigawatt-scale site leaning hard on evaporative cooling at peak could need 5 to 10 million gallons a day — the kind of draw that makes a town's water pressure sag when the servers get thirsty.

Running hot doesn't erase that water. It mostly relocates it somewhere that never shows up on the facility's own report.

So where does that leave the rest of us?

Which leaves me right back where I started — a stranger's comment, a number that sounded insane, and a weekend of reading to figure out which part was true. Turns out the chips are mostly fine, the water largely just moves, and the real fight is over who pays for the power.

That's the whole reason Byte Bungalow exists: grab one of these threads and pull until the boring, buried, actually-true part drops out. The full breakdown — the turbines, the substations, the grid math under all of it — lives on the channel. Come watch, and subscribe so the next thread finds you.

Common questions

Keep reading

Sources

• TechTarget — Schneider Electric modeled a 100 MW AI facility and found raising the cooling supply water from 32°C (90°F) to 45°C (113°F) cut water-usage effectiveness by about 66% in Dallas and 82% in Paris; ASHRAE's Thomas Loxley said 'Silicon and other power distribution components can experience accelerated aging at higher temperatures'; UC Riverside's Shaolei Ren estimated a 1 GW facility relying heavily on evaporative cooling at peak could need roughly 5 to 10 million gallons of water per day.
• National Renewable Energy Laboratory (NREL) — ASHRAE's recommended data center inlet air range is 18–27°C (64.4–80.6°F); allowable dry-bulb limits run by class up to 45°C (113°F) for class A4.
• Axios — In 2024 Google used about 7.2 billion gallons of freshwater and replenished roughly 4.5 billion gallons, about 64%.
• USENIX ;login: (El-Sayed et al., University of Toronto) — Using failure data from Google, Los Alamos and SciNet, the study found that below about 50°C internal device temperature errors tend to grow linearly with temperature rather than exponentially, and found no correlation with temperature for DRAM failures and node outages.