The global data center liquid cooling market is on track to reach $141.19 billion by 2036, according to a new industry analysis published this week, as AI training workloads push rack densities past what air cooling can handle.
Air cooling has dominated data center design for four decades. It worked because servers sat at 5 to 8 kilowatts per rack and a raised floor with chilled water could keep them below 27 degrees Celsius. That assumption broke when GPU clusters for training large models started landing at 40 to 60 kilowatts per rack, and some liquid-cooled configurations now run above 100 kilowatts.
The numbers tell the story. Grand View Research valued the liquid cooling market at $6.7 billion in 2025 and projects $8.2 billion this year, climbing to $29.5 billion by 2033. A separate forecast cited by Yahoo Finance puts the 2036 figure at $141.19 billion. The spread between those two endpoints reflects different assumptions about how fast hyperscalers retrofit existing sites versus building new ones from scratch.
The reason this matters now is timing. Microsoft, Google and Amazon have all announced liquid-cooled data center expansions in the past two quarters, and several colocation providers have started retrofitting air-cooled halls rather than waiting for new construction. The retrofit question is the one that will decide whether the smaller number or the larger number turns out right.
Why air cooling hit the wall
Air has a specific heat capacity of about 1.005 kilojoules per kilogram per Kelvin. Water sits at 4.18. That gap is the whole story. Moving the same amount of heat out of a rack takes roughly four times less mass flow with liquid than with air, which is why liquid cooling can handle densities that air simply cannot move enough volume through.
The practical limit showed up in 2023 and 2024, when operators started seeing hotspots in GPU clusters even with upgraded CRAC units. Raising the airflow rate adds fan power, which adds heat, which raises the cooling load further. Operators I spoke to described it as chasing their own tail.
Two approaches have split the market. Direct-to-chip cooling runs coolant through cold plates bolted to CPUs and GPUs, and it is the easier retrofit because the rest of the rack stays air-cooled. Immersion cooling submerges whole servers in dielectric fluid, and it is more efficient but harder to service and harder to retrofit into an existing building.
The retrofit problem
Retrofitting is where the money actually gets spent. A new liquid-cooled hall can be designed around manifolds, leak detection and raised containment from day one. An existing air-cooled hall has to be drained, re-piped and re-certified without taking production offline.
Dell'Oro Group analyst Baron Fung told Reuters that roughly 60 percent of the addressable market through 2028 is retrofit rather than greenfield. That is the figure that should worry operators, because retrofit projects run 18 to 24 months longer than greenfield builds and carry higher contingency budgets.
The other constraint is water. Liquid cooling still needs a heat rejection path, and in water-stressed regions that path increasingly means dry coolers or adiabatic systems that use less water but more energy. Operators in Phoenix and Singapore have told me they are already hitting permitting friction on new cooling water allocations.
What the vendors are selling
Vertiv, Schneider Electric and CoolIT dominate the direct-to-chip segment. Vertiv reported a 22 percent year-over-year increase in liquid cooling orders in its most recent quarterly filing. Schneider has been bundling CDU, manifold and monitoring software as a single SKU, which matters for retrofit buyers who do not want to integrate three vendors.
Immersion has its own leaders. Submer and LiquidStack compete on single-phase and two-phase systems, and both have signed supply agreements with two of the three largest hyperscalers in the past year. Neither company discloses volumes, so the market share figures floating around are estimates.
The pricing structure has shifted too. Five years ago liquid cooling was sold as a premium option at roughly 30 percent above air-cooled equivalents. Today the gap has narrowed to under 10 percent on new builds, and on retrofits the comparison is meaningless because air cooling cannot hit the density target at all.
The edge angle
Edge sites are the awkward middle child of this story. They are small, they are often in buildings not designed for data center loads, and they cannot justify a full liquid loop. But edge AI inference is the workload growing fastest, and inference at the edge is where latency requirements are tightest.
The compromise most operators have landed on is rear-door heat exchangers and in-row liquid loops that top out around 30 kilowatts per rack. That covers most inference workloads today. It will not cover the next generation of models, and nobody I have spoken to has a clean answer for what happens then.
What to watch next
Three things will tell us which forecast is closer to right. First, whether hyperscalers actually retrofit or build new, which shows up in construction permits and CDU order books. Second, whether immersion gains share against direct-to-chip, which shows up in service contract terms. Third, whether water permitting tightens in the regions where the buildout is concentrated.
I am skeptical of the $141 billion figure. It assumes retrofit economics improve faster than they have in the past five years, and it assumes water permits stay available in the regions where the buildout is concentrated. Neither assumption is obviously right. The $29.5 billion Grand View number looks more defensible to me, and even that requires retrofit projects to accelerate.
The more interesting question is not the market size but the operational one. Operators who get liquid cooling right will run denser, cheaper-per-watt compute. Operators who get it wrong will spend two years re-piping a hall and still miss their density targets. The gap between those two outcomes is where the next five years of this market will be decided.
For more on related infrastructure trends, see data center coverage.
Source: Reuters data center reporting.
One more data point worth noting: the Fortune Business Insights cooling market forecast puts the 2026 global figure at 21.0 billion dollars, which implies the liquid share is still a minority of total cooling spend. That ratio is the one I would track quarter to quarter.