Byte BungalowByte Bungalow

← Stories

Immersion Cooling Looks Like the Obvious Fix. So Why Won't Data Centers Switch?

Updated October 18, 2026

You can sink an entire server into a tank of clear oil, and it runs cooler and quieter than one in a screaming air-cooled rack — and it sips a fraction of the water. So why are the biggest AI builders mostly walking past that and clamping tiny radiators onto their chips instead? I spent an afternoon chasing that exact question. The cooling part works great. It's the tank, the warranty, and one discontinued chemical that keep getting in the way.

So where does the water even come from?

First thing I had to unlearn: a tank of oil isn't touching any water. The server sits in what's called dielectric fluid — liquid that carries heat but won't short out electronics. One commenter on a cooling video put it bluntly: the loop is closed, so where's all this water waste coming from? He's half right.

The water doesn't live in the tank. It lives at the far end of the loop, the same way the coolant in your car never actually touches the air cooling it — the radiator hands the heat off. A data center hands its heat off too, usually to evaporative cooling towers that boil away water by the thousands of gallons, or back to the power plant that's burning water to make the electricity in the first place.

So here's where immersion actually wins. Pull the heat straight into oil and you get to skip a lot of that evaporation step. Measured by water used per unit of computing, a sealed liquid setup can run at a sliver of what a big air-cooled hall burns. That part is real. Which is exactly why the question gets loud — if it uses way less water and has no fans to scream, why isn't everyone doing it?

The two-minute problem

Picture the most boring task in a data center: a drive dies, a tech walks over, pops it out, slots a new one in. Two minutes, done, never even shut the machine off.

Now picture that same drive in a tank. The tech has to hoist the whole server up out of the fluid on handles, let it drip-dry on a rack like last night's dishes, work on it while it's still slick, then lower it back in. A guy who said he'd spent a career in data centers left a comment that stuck with me — he couldn't even picture how you'd service these things once they're submerged. Neither could I, until it clicked that the answer is just: wet and slow, and it needs trained hands and a place to drain.

That's the human on the other side of the shiny efficiency number. Every repair in a building full of thousands of servers just got messier. For an operator racing to switch on capacity, messier is a dealbreaker. And it gets worse, because the hardware itself often isn't even allowed in the tank.

The warranty, and the fluid that got pulled off the shelf

Most server warranties are written around air or a cold plate. Drop that same box in fluid and a lot of manufacturers call it out of spec — so now you're risking the coverage on machines that cost more than a house. Which means you need special immersion-rated servers and a separate support deal, another wall between you and the tank.

Then there's the chemical problem, and this one surprised me most. The version of immersion with the absolute best numbers is called two-phase: the fluid actually boils off the hot chips and recondenses above them, like a tiny indoor rainstorm carrying heat away. The catch is that the fluids that pull off that trick are PFAS — the forever chemicals. 3M, the dominant maker, announced it's exiting PFAS production by the end of 2025, and Europe's chemical regulators are moving to restrict the whole family. A commenter on a Meta data-center video flagged the same fear: that servicing one of these closed systems can vent those chemicals into the air.

So the best-performing flavor of the obvious fix is losing its magic ingredient right as AI demand goes vertical. Single-phase oil sidesteps the forever-chemical fight, but it doesn't hit the same jaw-dropping numbers. I couldn't pin down a clean replacement that matches two-phase and dodges the regulators — as far as I can tell, nobody's fully there yet.

Why the boring fix won

So what are the AI giants reaching for instead? Mostly a thing called direct-to-chip: a little liquid-cooled plate sits right on top of the hot chip and pipes coolant through it, while the rest of the server stays dry and normal. Nvidia's big AI systems ship with this as the default. It drops into a regular rack, keeps the warranty intact, and a tech can still swap a part without going fishing in a tank.

That's the whole answer, and it's almost annoyingly plain. The tank isn't losing on physics. It's losing because the buildout is moving too fast to redesign the entire room, retrain the crew, and bet on a fluid that might get regulated off the shelf. The obvious fix lost to the one that fit through the door.

If you like pulling a thread like this — the water, the fans, the forever chemicals, the strange economics of the machines behind every AI prompt — that's the whole point of Byte Bungalow. Come watch the full breakdown on the channel and hit subscribe, so you catch the next rabbit hole the minute I go chase it down.

Common questions

Does immersion cooling actually use less water than a normal data center?
On-site, yes. The tank is a closed loop, so the savings come from skipping a lot of the evaporative cooling towers that boil off water. The catch is that the heat still has to go somewhere, and the power plant feeding the building is often the bigger water user. So less water is true, but a tank by itself doesn't make a data center water-free.
How do you repair a server that's submerged in fluid?
You lift it out of the tank on handles, let it drain, service it while it's still coated in oil, then lower it back in. It works, but it's slower and messier than the two-minute hot-swap you get with an air-cooled server, and it needs techs trained for that workflow. That servicing friction is one of the biggest reasons operators hesitate.
What are the 'forever chemicals' people mention with data center cooling?
The best-performing immersion setups, called two-phase, use fluids in the PFAS family that boil and recondense to carry heat. Those are the forever chemicals regulators are cracking down on. 3M said it would exit PFAS production by the end of 2025, which pulls the main supply out from under that approach. Single-phase oil cooling avoids the PFAS issue but gives up some performance.
Why don't they just build data centers somewhere cold or next to a lot of water?
Some do chase cold climates and big water sources for exactly that reason. But a data center also needs cheap, reliable power, fat fiber lines, roads, and people to staff it — and those usually aren't sitting at the North Pole. Picking the location solves part of the heat problem, not all of it.
Is direct-to-chip cooling the same as immersion?
No. Direct-to-chip puts a small liquid-cooled plate right on the hot chip while the rest of the server stays dry in a normal rack. Immersion dunks the whole server in fluid. Direct-to-chip is winning the AI buildout mostly because it fits into existing racks and keeps hardware warranties intact.

Keep reading

Sources

  • akcp.com — The spine source ("Immersion Cooling Has the Best Numbers, So Why Isn't It Winning the AI Buildout?"). Key facts taken: warranties/OEM support are written around air-cooled or dire

Byte Bungalow

I do the homework out loud.

If this is the kind of thing you want checked against the documents instead of the hype, the video's above — and there's a new one every week.

Subscribe on YouTube →

By Byte Bungalow. Home power and home tech, checked against the documents instead of the hype. Independent commentary; not affiliated with any manufacturer, utility, or builder named here. Not professional electrical advice.