Data Centers Are About to Melt — and Cooling Just Became the AI Trade's Biggest Sleeper Market

Data Centers Are About to Melt — and Cooling Just Became the AI Trade's Biggest Sleeper Market

Let me tell you something that's been sitting with me all week. Every time we talk about the AI buildout, we talk about chips. We talk about power. We talk about gigawatts and transformer lead times and grid queues. Nobody talks about the fan.

That's about to change. This week a Danish industrial giant most people can't pronounce — Danfoss — published half-year numbers that should make every founder running infrastructure sit up straight. The company grew 15% organically in the first half of 2026 on sales of EUR 5.3 billion. Data centers were 7% of their global business last year. This year, they expect that number to hit 14% to 16%. Double. In twelve months.

I watched a video breakdown this week walking through ten data center cooling stocks, and the thesis is simple: the chips the hyperscalers are buying run so hot that air conditioning doesn't work anymore. That's not a marketing problem. That's physics. And when physics says you need a new way to cool a rack, an entire industry gets rebuilt underneath you.

The Physics Wall Nobody Can Talk Their Way Around

Here's the number that matters: Nvidia's GB200 NVL72 rack draws 120 kilowatts. One rack. The kind of density we used to spread across an entire server room now sits inside a single cabinet the size of a fridge. And air cooling — the stuff every data center on earth was built around for thirty years — tops out around 20 to 30 kilowatts per rack before it becomes a joke in terms of efficiency.

You don't need to be a thermal engineer to see where this ends. You just need to have paid an electricity bill. At 120kW per rack, you're not cooling that thing with fans and cold aisles. You're cooling it with liquid, full stop. Industry analysts put the liquid cooling market somewhere between $870 million in 2024 and $10.7 billion by 2030 — a 52% compound annual growth rate. Other forecasts are even more aggressive: roughly $4 billion this year to nearly $19 billion by 2032. Pick your number, the direction is the same.

And here's the part that nobody in the capex headlines mentions: cooling is already the second-largest capital expense in a data center after power, and the single biggest non-IT operating cost. The AI trade spent the last two years arguing about GPUs. The next argument is about the bill to keep them from melting.

The Supplier Stampede Is Real — and It's Not Just the Hyperscalers

Danfoss isn't the only one smelling the heat. This week Trane and Eaton announced a strategic collaboration to integrate thermal management and electrical architecture — the whole point being that when your cooling system and your power system talk to each other, you stop wasting both. It's built around Nvidia's Omniverse data center blueprint, which tells you exactly whose roadmap these guys are betting on.

Samsung got in on the act too, with its FläktGroup subsidiary opening a plant in Pune, India, with capacity for up to 6,500 HVAC units a year aimed squarely at AI data center cooling. That's a company with phones in everyone's pocket deciding cooling infrastructure is worth factory floor space.

Why should an independent hosting guy care? Because when suppliers like these start ramping dedicated production lines, it means the demand is real, it's structural, and it's already priced into their order books. The hyperscalers can outbid everyone for GPUs. The cooling supply chain, on the other hand, has to serve every colo, every regional provider, every two-rack operation that suddenly finds itself hosting an inference workload that draws more heat than its building was designed for. That's the opportunity. And that's the squeeze.

The Water Fight Nobody Wants to Have

Here's where this story gets weird, and I mean that literally. This week Liquid Death — the canned water company — teamed up with former Eagles star Jason Kelce for a marketing campaign joking that people should pee on data centers to cool them. It's funny. It's also accidentally describing a real tactic.

Data centers are already using recycled wastewater for cooling, and the practice is booming. TechCrunch talked to executives at Ecolab, the WateReuse Association, and engineering firm CDM Smith, and the message was consistent: treating sewage and wastewater to industrial grade and piping it into cooling towers isn't a gimmick, it's a decades-old practice that's now scaling precisely because AI demand is blowing past the potable water supply in drought-stressed regions. The UN projects that by 2030, AI data centers will consume water equivalent to the daily needs of 1.3 billion people.

I did a whole breakdown on the water crisis earlier this summer — 264 billion gallons a year, communities fighting transparency battles with Google over The Dalles and other facilities hiding their usage. Nothing about that math has gotten better. What's changed is that now the cooling supply chain itself is having to engineer around water scarcity, and that cost is going to land somewhere. It always does.

The Retrofit Trap — the Secondary Bottleneck Nobody's Pricing In

Here's the part that keeps me up at night, and it's the part the Danfoss headlines don't cover: the new-build stuff is fine. Nvidia's liquid-cooled racks in new hyperscale facilities — great, they're designed for it. The problem is the installed base. Every existing air-cooled data center in the world was built on the assumption that 10 to 15 kilowatts per rack was a lot. Most of them can't take liquid cooling without a retrofit that costs more than the building is worth.

That's the secondary bottleneck, and it's the same shape as every other bottleneck in this buildout: the announced capacity looks glorious, the delivered capacity is a different story. Retrofits take months of planning, new piping, new power distribution, and downtime you can't afford if you're running customer workloads. Meanwhile the equipment lead times are stretching the same way transformer lead times did — five years for a big transformer, remember? The cooling gear is heading down the same road.

Add water disclosure laws tightening in drought-prone states, and you've got a perfect storm: the cooling bill is coming due for the old facilities at the exact moment the suppliers are fully booked building for the new ones.

What This Means for Independent Hosting Providers

So what do you actually do with this? I'll give it to you straight, the way I'd tell a client:

First, budget for the cooling line item now. If you're in a drought-prone region, expect 10% to 15% power-cost creep as operators shift to waterless and liquid cooling that draws more electrical load for pumps and chillers. That's not a maybe. That's the trajectory.

Second, lock your colo contracts early in secondary markets. The big players are already hoovering up capacity in places with water and power to spare. If you wait until the disclosure laws tighten and the supply dries up, you're paying whatever's left.

Third, demand water numbers. If your provider can't tell you their WUE — water usage effectiveness — that's an answer in itself. Favor operators who disclose it, because they're the ones who've actually thought about the constraint.

Fourth, get ahead of equipment lead times. If you're planning any cooling upgrade, order now. The same dynamic that made transformers a five-year wait is forming in cooling gear, and the suppliers are booked building for hyperscalers first. You're not going to jump that queue from behind.

The Bottom Line

Here's the truth. We spent two years treating cooling like a utility line item — something the facilities guy handles, boring, no margin in it. That era is done. Cooling is now a market, with its own supply chain, its own stock plays, its own bottlenecks, and its own political fights over water. The hyperscalers are going to be fine; they have the balance sheets to build whatever physics demands. The independents who treat cooling as an afterthought are going to get cooked.

Stop watching the GPU announcements. Start watching the thermals. The AI trade just found its next bottleneck, and it's not a chip — it's a radiator.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Bloomberg, Danfoss, TechCrunch, Data Center Dynamics, EE Times, Arizton, BMAG Consulting.