AI Industry Structure, Layer 3: Cooling

When people talk about data centers, they usually think of servers and power, but one of the hottest topics in the industry right now — literally — is cooling. When tens of thousands of GPUs run non-stop in one place, a paradoxical situation emerges: performance actually drops because the heat can’t be dissipated fast enough. In this article, we’ll cover why cooling suddenly became so important, how air cooling differs from liquid cooling, and which companies are leading this market.

Revisiting Cooling

GPUs generate a tremendous amount of heat when they compute. If that heat isn’t dissipated in time, the chip either throttles itself or shuts down entirely. That’s why designing cooling equipment has become just as central to building a data center as designing for servers and power.

Cooling itself is actually old technology. It’s just that conditions have changed to the point where the old method — cooling with air — simply can’t keep up anymore.

Why It Suddenly Became a Problem

The answer is that the heat a single GPU generates has multiplied several times over in just a few years. As competition over AI performance intensified, chipmakers packed more and more compute units into a single chip, and power consumption (and heat output) grew right along with it. The heat generated by a single server rack has climbed to a level that would have been unimaginable by old server-room standards.

On top of that, AI servers are built to cram GPUs as tightly together as possible, so the heat concentrates in one spot. With enough space, air cooling is fine — but once heat sources are packed into a tight space, air alone hits its limits.

Air Cooling vs. Liquid Cooling

The latest high-performance GPU servers have reached a point where they can’t even be designed without liquid cooling. Circulating liquid through a cooling plate mounted directly on the chip has become the standard, and more and more new data centers are being designed around liquid cooling from the ground up.

Liquid Cooling Also Splits Into Two Approaches

Liquid cooling itself splits into different methods. Direct-to-Chip mounts a cooling plate on the chip and cools just that spot with liquid, while Immersion Cooling is the more radical approach of submerging the entire server in a special liquid. Direct-to-Chip is more widely used right now since it doesn’t require major changes to existing server design, while immersion cooling has higher cooling efficiency but requires redesigning the server itself, so it’s still being adopted mainly in a handful of the newest data centers.

Cooling’s Two Costs — Electricity and Water

PUE, the Metric for Measuring Efficiency

The industry standard metric for comparing how efficient a data center is, is PUE (Power Usage Effectiveness). It’s total power consumption divided by the power actually used for the servers (computation), and the closer it is to 1, the better the data center. The more power that gets siphoned off for cooling, the worse this number gets, so newer liquid-cooled data centers tend to post lower PUE figures. That’s also why cloud companies frequently mention this metric in earnings calls.

A Problem Just as Big as Electricity: Water

Cooling doesn’t just use electricity. Evaporative cooling dissipates heat by evaporating massive volumes of water — a single large data center can use up to 5 million gallons a day, roughly the same amount a city of 50,000 people would use. Since this water comes from local water supplies, resident opposition citing water shortages has grown significantly. In Q1 2026 alone, 75 data center projects worth $13 billion were delayed by local opposition.

So alternatives that use little to no water at all are emerging quickly. Closed-loop cooling continuously recirculates the same coolant, drawing in almost no fresh water. NVIDIA recently unveiled a system that circulates coolant as roughly 45-degree “warm water” to dissipate heat without evaporation, stating that “water consumption inside the data center has effectively been solved.” That said, this only covers what happens inside the data center building — the power plants (especially fossil-fuel plants) that generate the electricity still use a lot of water, so it’s more accurate to say the water problem is shifting location than to say it’s disappeared entirely.

What Investors Should Watch For

Cooling doesn’t generate as much buzz as GPUs or cloud, but it has a defining trait: it’s a “hidden necessity.” As long as AI data centers keep growing, this demand follows almost automatically.

The thing to check is “how quickly liquid cooling’s share of this company’s revenue is growing.” For a company where air-cooling revenue still dominates, the shift to liquid cooling could actually be a growth opportunity. Conversely, a company that’s already a liquid-cooling leader is structured to benefit directly as the market grows.

Closing Thoughts

Cooling isn’t as glamorous as GPUs or cloud, but it’s the quietest, most reliably growing layer in the AI industry structure. Every time another AI data center gets built, cooling demand follows almost automatically, and the shift from air cooling to liquid cooling has already become an irreversible trend.

That said, this layer is tricky because it requires weighing two resources — electricity and water — at the same time. Different data centers make different choices between methods that save electricity by using more water (evaporative cooling) and methods that save water by using more electricity (air cooling, closed-loop), and that choice can lead to unexpected variables like local resident opposition or environmental regulation. So investors need to look not just at these companies’ revenue growth, but also at where they’re building their data centers — and whether that’s a water-stressed region.

But solving the cooling problem is pointless if a company can’t secure the power to run those GPUs in the first place. In the next article, we’ll take a close look at the third topic in Layer 3: power. We’ll cover why big tech companies have started signing direct deals with nuclear power operators.