Air has run out of road. As AI chips pack more power into smaller spaces, the fans that cooled data centres for decades can no longer keep up — and the industry is switching to liquid at remarkable speed. According to Goldman Sachs projections, liquid-cooled AI servers are expected to rise from roughly 15% of the market in 2024 to about 76% in 2026.

Why air stopped working

The problem is density. A traditional server rack might draw 5 to 10 kilowatts of power, all of which eventually becomes heat. Modern AI racks packed with accelerators can draw many times that, and next-generation designs are pushing toward and beyond one megawatt per rack.

Air is a poor medium for moving that much heat. It has low thermal capacity, so cooling a dense rack by air requires enormous volumes of fast-moving cold air — which means bigger fans, more noise, more energy spent on cooling rather than computing, and hard physical limits on how tightly equipment can be packed. Past a certain density, air simply cannot remove heat fast enough to keep chips within safe temperatures.

How liquid cooling works

Liquid carries heat far more efficiently than air. Three approaches dominate:

  • Direct-to-chip (cold plate). Coolant flows through metal plates mounted directly on the hottest components. This is the most common approach in new AI deployments because it can be retrofitted into fairly conventional rack designs.
  • Immersion cooling. Entire servers are submerged in a non-conductive fluid. Extremely effective, but it requires purpose-built facilities and changes how hardware is serviced.
  • Two-phase systems. The coolant boils and condenses, absorbing large amounts of heat during the phase change. Industry roadmaps describe these as the successor to today's single-phase systems as densities climb.

Beyond the rack, coolant distribution units manage flow and temperature, and heat is ultimately rejected outside the building. Increasingly, chipmakers are integrating microfluidic channels directly into chip packages, capturing heat closer to its source than any external plate can.

The numbers behind the shift

Market forecasts underline how fast this is moving. The AI data-centre liquid cooling market is projected at roughly $3.7 billion in 2026, with longer-range estimates pointing toward around $17.8 billion by 2036. Modular, pre-assembled cooling units starting around two megawatts are becoming a default building block for high-density sites.

These are projections rather than results, and long-range market forecasts are frequently revised. The near-term adoption figures, however, are corroborated across multiple industry analyses.

Background: cooling is an energy story

Cooling has always been one of the largest non-computing energy costs in a data centre — a point we explored in our explainer on why AI needs so much power. Efficiency is measured with PUE, the ratio of total facility energy to energy reaching the computers. Because liquid moves heat more efficiently, it can meaningfully lower that overhead.

That matters commercially as well as environmentally. When AI infrastructure deals are negotiated in gigawatts — as in the AMD–Anthropic partnership — every percentage point of energy spent on cooling instead of computing is capacity that cannot be sold.

The trade-offs

Liquid cooling is not free of complications. It introduces plumbing into rooms full of electronics, requires new maintenance skills, and raises the cost of retrofitting existing facilities. Leaks, while rare in well-engineered systems, are a failure mode air cooling never had. Many operators are therefore running hybrid facilities — liquid for dense AI racks, air for everything else.

Why it matters

Cooling has quietly become a constraint on how fast AI can scale. A site's practical capacity depends not only on the power it can secure but on the heat it can remove. As we noted in our analysis of who profits from the AI boom, this is exactly the kind of unglamorous infrastructure layer where real money is being made.

Key takeaways

  • Liquid-cooled AI servers are projected to rise from about 15% of the market in 2024 to roughly 76% in 2026.
  • Air cooling fails at modern AI rack densities, which are pushing toward a megawatt per rack.
  • Direct-to-chip cold plates dominate new deployments; immersion and two-phase systems handle higher densities.
  • Chipmakers are building microfluidic cooling channels directly into chip packages.
  • Liquid cooling lowers energy overhead but adds plumbing, maintenance, and retrofit complexity.

The bottom line

The switch to liquid cooling is one of the fastest infrastructure transitions in modern computing, and it is happening because there is no alternative. If AI is going to keep scaling, the heat has to go somewhere — and air can no longer carry it.