How big is the data center cooling market as AI densities rise?

Why do estimates of the data center cooling market vary so much?

The data center cooling market is currently worth roughly $20 billion globally, but a realistic range is closer to $15 billion to $25 billion because researchers count different parts of the cooling stack.

The spread between published estimates is unusually wide. Mordor Intelligence puts the 2026 market at about $12.4 billion, MarketsandMarkets at $13.2 billion, Fortune Business Insights at $21.0 billion, The Business Research Company at $22.8 billion and Grand View Research at $31.4 billion. The median of those five estimates is almost exactly $21 billion.

Most of the gap comes from definitions rather than radically different views of demand. A narrow definition focuses on chillers, computer-room air handlers, heat exchangers, coolant distribution units, direct-to-chip equipment and other thermal hardware. Broader studies include installation, maintenance, controls, containment and other cooling-related services.

AI makes that boundary even messier. A 120-kW AI rack can require cold plates, manifolds, CDUs, secondary coolant loops, pumps, controls, leak detection and changes to the facility water system. Depending on the researcher, some of those components sit inside “liquid cooling,” some appear under broader data center infrastructure and some fall into services.

For our purposes, cooling means the equipment and specialist services whose main job is to move heat away from servers and ultimately out of the facility. General construction, GPUs, servers and electrical equipment stay outside that definition.

Using that boundary, we would put the core equipment market around $15 billion to $20 billion today. Once installation and specialist thermal services are included, roughly $20 billion to $25 billion is more useful. Estimates above $30 billion generally use a wider definition.

Current market estimate	2026 value	What it appears to count
Mordor Intelligence	$12.4B	Relatively narrow cooling market
MarketsandMarkets	$13.2B	Cooling equipment and selected services
Fortune Business Insights	$21.0B	Broader cooling equipment market
The Business Research Company	$22.8B	Wider equipment and service exposure
Grand View Research	$31.4B	Broad cooling solutions and services
Five-estimate median	~$21.0B	Useful central reference
Range we would use	~$15B-$25B	Best representation of the current market

If you want more recent data on this point, please see our latest data center cooling market report.

How big is the data center cooling market today compared with the rest of AI infrastructure?

Data center cooling is already a roughly $20 billion market today, yet it remains a relatively small slice of the hundreds of billions now being spent on AI infrastructure.

IDC estimated worldwide AI infrastructure spending at $318 billion in 2025, more than twice the $153 billion recorded one year earlier. Even if we use the upper end of our cooling range, thermal equipment and services still represent only a fraction of the full AI infrastructure bill.

Cooling looks much bigger when we compare it with the physical infrastructure surrounding servers rather than with GPU spending. Dell'Oro reported that power, cooling and related data center physical infrastructure generated $10.9 billion of manufacturer revenue in the fourth quarter of 2025 alone. Annual spending on the physical layer is therefore already measured in tens of billions of dollars.

The next leg of growth comes from capacity. McKinsey estimates that global data center demand could rise from roughly 82 GW in 2025 to around 220 GW by 2030. If that path is even close, the industry could add more than 130 GW in five years.

Every additional megawatt of IT load eventually becomes heat. Cooling vendors do not capture a fixed percentage of that spending, but the physical relationship gives the market an unusually strong underlying demand floor.

So a $20 billion market can still look small beside GPUs while becoming a very large industrial equipment category in its own right.

Have AI rack densities really changed enough to transform data center cooling?

Yes. Frontier AI racks now use roughly ten times the power of a typical data center rack, which has pushed cooling into a completely different engineering regime.

Uptime Institute's surveys show how unusual the AI buildout still is compared with the installed base. Typical rack densities remain around the high single digits or low teens in kilowatts, and its 2025 survey found that more than 80% of operators had no racks above 30 kW.

NVIDIA's rack-scale AI systems are several levels above that. The GB200 NVL72 runs around 120 kW per rack. The GB300 NVL72 can reach roughly 142 kW. These systems connect 72 GPUs inside a single rack-scale architecture and depend on direct liquid cooling for the main compute components.

That is an order-of-magnitude jump from the ordinary racks still operating across most enterprise and colocation facilities.

And densities are still climbing. Current NVIDIA roadmaps and data center designs are preparing for rack densities well above today's Blackwell systems, while infrastructure suppliers are already discussing several-hundred-kilowatt racks. ASHRAE has also begun treating those densities as a real design problem rather than a distant scenario.

The gap between an ordinary 10-kW rack and a 120-kW-plus AI rack explains most of the disruption happening in cooling today. The average facility has changed slowly; the most valuable new compute deployments have changed extremely quickly.

Computing environment	Typical or approximate rack density	Cooling approach
Legacy enterprise IT	5-10 kW	Conventional air cooling
Modern enterprise/cloud	10-20 kW	Air cooling remains practical
Dense enterprise/HPC	20-30 kW	Liquid cooling becomes attractive
High-density HPC	30-60 kW	Hybrid or liquid increasingly useful
Dense AI systems	60-100 kW	Liquid cooling becomes increasingly hard to avoid
NVIDIA GB200 NVL72	~120 kW	Direct liquid cooling
NVIDIA GB300 NVL72	Up to ~142 kW	Direct liquid cooling
Next-generation AI racks	Hundreds of kW increasingly discussed	Facility design increasingly built around liquid

Is liquid cooling already mainstream in AI data centers?

Liquid cooling is already becoming standard for frontier AI infrastructure, even though most of the world's data center racks still run on air.

TrendForce estimates that liquid cooling penetration among AI chips rose from roughly 14% in 2024 to about one-third in 2025 and is reaching around 53% in 2026. Its forecasts put penetration close to 60% in 2027.

That is a fast transition for data center infrastructure, where equipment can stay in service for well over a decade.

The leading AI platforms are pulling the market forward. NVIDIA designed its GB200 and GB300 NVL72 systems around liquid cooling. Google has spent years building liquid-cooled TPU infrastructure, and TrendForce estimates that more than 80% of Google's AI servers now use liquid cooling.

At the same time, Uptime's data shows why it would be wrong to call liquid cooling universal. Most existing facilities still operate at rack densities where air works perfectly well.

So the market is split. Air remains dominant across the installed base, while liquid is quickly becoming the normal choice for the highest-density AI deployments. Because those AI deployments attract a disproportionate amount of new capital, liquid cooling can become a very large business long before it reaches most racks.

How big is the liquid-cooling market right now?

The data center liquid-cooling market is probably around $5 billion to $7 billion today under a broad definition, while direct manufacturer revenue is still closer to the low single-digit billions.

Dell'Oro gives us one of the cleaner benchmarks because it tracks manufacturer revenue rather than adding every related service. The firm estimated that liquid-cooling manufacturer revenue roughly doubled during 2025 and ended the year close to $3 billion.

Broader market studies now produce much larger figures. Current 2026 estimates include approximately $4.1 billion from MarketsandMarkets, around $6.0 billion from Global Market Insights, $6.3 billion from 360iResearch, $6.4 billion from The Business Research Company, $6.8 billion from Mordor Intelligence and more than $8 billion from Grand View Research.

Across those broader estimates, the center of gravity sits around $6 billion.

We would therefore separate two numbers rather than force them into one. Roughly $3 billion tells us what liquid-cooling manufacturers were already generating recently. Around $5 billion to $7 billion better captures the current ecosystem once a wider set of products and services is included.

The growth rate is more striking than the exact boundary. Dell'Oro's data suggests manufacturer revenue roughly doubled in a year, while the wider cooling market is growing much more slowly. Liquid cooling is taking an increasingly large share of each new dollar spent on thermal infrastructure.

Liquid-cooling estimate	Current 2026 estimate
MarketsandMarkets	~$4.1B
Global Market Insights	~$6.0B
360iResearch	~$6.3B
The Business Research Company	~$6.4B
Mordor Intelligence	~$6.8B
Grand View Research	~$8.2B
Practical range we would use	~$5B-$7B

If you want more recent data on this point, please see our latest data center cooling market report.

Is AI creating more cooling spending or simply replacing air cooling?

AI is creating additional cooling spending because high-density racks need hardware that conventional air-cooled servers never required.

Some substitution is obviously happening. A liquid-cooled AI hall can use less room-level air-moving equipment per unit of compute than a facility designed entirely around fans and air handlers.

The new liquid loop adds another set of products, however. High-density racks need cold plates, manifolds, pumps, CDUs, piping, monitoring, fluid controls and often changes to the building's heat-rejection system.

Air also remains part of many AI designs. NVIDIA's GB200 NVL72 uses direct liquid cooling for its major heat-producing components while other parts of the rack still rely on airflow. The facility then has to move all that captured heat outside the building through another thermal system.

Dell'Oro has been tracking this overlap for several years. Its forecasts show liquid cooling growing far faster than conventional thermal equipment while air-cooling revenue can still increase because the total volume of installed computing capacity keeps expanding.

That is a pretty good setup for cooling suppliers: new liquid products are entering the bill of materials at the same time as the overall data center market is getting larger.

Does every AI data center need liquid cooling?

No. Plenty of AI workloads can still run in air-cooled data centers, particularly when operators spread servers across more racks instead of pushing each rack toward 100 kW or more.

Inference is the main reason this distinction matters. Many inference deployments use fewer accelerators per rack, lower-power chips or architectures optimized for performance per watt. Those systems can stay inside a 10-kW, 20-kW or sometimes considerably higher air-cooled envelope.

Even NVIDIA still sells powerful air-cooled AI servers. A DGX B300, for example, draws roughly 14.5 kW at the server level. Operators can decide how many systems to place in a rack rather than automatically building a 140-kW rack-scale NVL72 environment.

Uptime Institute has repeatedly pointed out that densification involves an economic choice. A company with plenty of floor space may prefer more racks at lower density if that avoids an expensive mechanical retrofit.

The economics start changing around the 20-30 kW range. Purpose-built air systems can go higher, but airflow, fan energy and physical space increasingly become a headache. Once racks move toward 60, 100 or 140 kW, direct liquid cooling becomes much easier to justify.

So AI expands the liquid-cooling market without making liquid mandatory for every AI server. The strongest demand comes from training clusters and dense rack-scale systems where spreading hardware across a larger room would undermine the architecture or economics.

Are old data centers going to need expensive cooling retrofits?

Many older data centers will struggle to host dense AI racks without major cooling upgrades, which creates a second market alongside all the new AI facilities being built.

A facility designed around 5-10 kW racks was never built for a 120-kW GB200 NVL72. The problem reaches beyond adding a few pipes. Operators may need new water loops, CDUs, pumps, heat exchangers, electrical distribution, floor reinforcement and additional heat-rejection capacity outside the building.

That explains why greenfield AI projects can move more cleanly to liquid cooling. Engineers can design the mechanical and electrical systems around high-density racks from day one.

Existing data centers have more options than a complete rebuild. Liquid-to-air CDUs can let operators deploy liquid-cooled servers without extending facility water all the way to the rack. Dedicated high-density zones can also concentrate upgrades in one part of a building while ordinary workloads continue running elsewhere.

The scale of the installed base makes this important. Synergy Research counted 1,360 hyperscale data centers by the end of 2025, and that figure excludes a much larger universe of older enterprise and colocation sites.

We should therefore expect two cooling markets to grow side by side: purpose-built infrastructure for new AI campuses and retrofit products designed to squeeze denser compute into facilities that were built for another era.

Is direct-to-chip cooling winning over immersion cooling?

Direct-to-chip cooling is clearly winning the first large wave of AI liquid cooling because it handles high heat loads without forcing operators to completely redesign how servers are built and maintained.

Immersion cooling remains technically impressive. Servers sit inside dielectric fluid, which can remove large amounts of heat and reduce reliance on server fans.

The operational change is much bigger, though. Hardware maintenance, component compatibility, fluid handling and rack design all work differently in an immersion environment.

Cold-plate systems create less disruption. Liquid runs directly across plates attached to CPUs and GPUs while much of the rest of the server still looks familiar to operators and technicians.

NVIDIA's GB200 and GB300 rack-scale systems helped settle the near-term contest. The industry's highest-volume frontier AI architecture adopted direct-to-chip cooling, and an ecosystem of suppliers quickly formed around cold plates, manifolds, quick disconnects and CDUs.

Dell'Oro's latest work also describes single-phase direct liquid cooling as the dominant architecture in current AI clusters.

Immersion still has room in specialized deployments, and two-phase technologies could become more useful as rack heat loads climb further. For the current AI buildout, though, cold plates have become the path of least resistance.

If you want more recent data on this point, please see our latest data center cooling market report.

Who is actually buying all this high-density cooling equipment?

Hyperscalers and large colocation operators are driving most of the high-density cooling transition because they control the giant AI clusters where liquid cooling makes the most economic sense.

Synergy Research estimates that hyperscale operators already account for roughly 48% of worldwide data center capacity. Their share could approach two-thirds by 2031 as the number and average size of hyperscale facilities continue to rise.

The hyperscale data center count reached 1,360 at the end of 2025. More important for cooling vendors, quarterly hyperscaler capital expenditure reached about $142 billion in the third quarter of that year, almost 180% higher than roughly three years earlier.

A small number of companies can therefore move an entire equipment category. When Google, Microsoft, Meta, Amazon or a large GPU cloud standardizes a particular cooling architecture across hundreds of megawatts, suppliers suddenly have repeatable volume rather than one-off engineering projects.

Colocation operators extend that demand to companies that want dense AI infrastructure without building their own campuses. Those providers increasingly sell the ability to support 50-kW, 80-kW or 120-kW racks as part of the product.

The construction pipeline shows how much capacity is coming. Cushman & Wakefield counted 50.3 GW of operational data center capacity across the Americas in the first half of 2026 and another 37.7 GW under construction, with more than 90% already precommitted. Asia-Pacific had another 26.5 GW in its development pipeline.

Only part of that pipeline will use dense liquid-cooled AI hardware. Even so, cooling architecture increasingly has to be decided before a tenant's GPUs arrive. That gives hyperscalers and colocation companies enormous influence over which technologies become standard.

Is air cooling going away as AI grows?

Air cooling is going to remain a large business for years because most data center racks still run at densities where liquid cooling would add cost without solving a real problem.

Uptime Institute's 2025 survey found that 82% of operators had no racks above 30 kW. Ordinary enterprise servers, storage, networking and a large amount of cloud infrastructure continue to fit comfortably inside air-cooled environments.

There is also plenty of non-AI construction ahead. In McKinsey's continued-momentum scenario, global data center demand reaches roughly 219 GW by 2030, with about 156 GW tied to AI and another 64 GW coming from other workloads.

Even liquid-cooled AI racks retain some air-cooled components, so air handling does not disappear when cold plates arrive.

The shift shows up in the growth rates instead. Air cooling is attached to an enormous installed base and can keep growing as total data center capacity expands. Liquid cooling starts from a much smaller base but is taking a much larger share of high-density additions.

Today, the useful question for suppliers is less about when air cooling disappears and more about how much of the incremental thermal budget moves toward liquid. That transfer is already well underway.

Are cooling companies actually making much more money from AI?

Yes. The AI cooling boom is now visible in revenue, orders, backlogs and long-term capacity contracts rather than mainly in product announcements.

AAON gives us one of the cleanest examples. Its BASX data center business generated about $345 million of revenue in the second quarter of 2026, up 216% from a year earlier, while backlog reached roughly $1.43 billion. Earlier disclosures also showed trailing liquid-cooling sales growing close to 187%.

Modine's Airedale business gives us a different kind of evidence. The company signed an agreement with one strategic customer covering more than $4 billion of cooling products scheduled for 2027 through 2029. The customer paid $165 million upfront to help fund the manufacturing capacity needed to fulfill the order.

The scale is striking compared with Modine's recent data center business. Its data center revenue was around $700 million in fiscal 2025, so one customer's future commitment represents several years of that earlier annual sales base.

Vertiv's broader infrastructure results point in the same direction. The company reported fourth-quarter 2025 organic orders up 252% year over year as hyperscale and colocation customers accelerated their buildouts.

These figures give us more confidence than market forecasts alone. Companies are already converting high-density cooling demand into hundreds of millions of dollars of quarterly revenue and multibillion-dollar backlogs.

Why are companies paying billions to buy data center cooling businesses?

Data center cooling assets have become expensive because large infrastructure companies now want control of the thermal chain from the chiller all the way to the chip.

Schneider Electric completed its acquisition of a 75% controlling stake in Motivair for €850 million in 2025. Motivair brought CDUs, cold plates, heat exchangers and chillers into a company that already had a huge data center power business.

Eaton went much further with its $9.5 billion agreement to acquire Boyd's thermal business. Boyd was expected to generate around $1.7 billion of 2026 sales, including approximately $1.5 billion from liquid cooling. Eaton was therefore willing to pay several years of current revenue for a stronger position in thermal infrastructure.

Vertiv has concentrated on the service layer as well as hardware. It completed the roughly $1 billion acquisition of PurgeRite, whose specialists flush, purge and filter liquid-cooling loops. More recently, Vertiv agreed to acquire Ireland-based King Environmental Services, extending similar fluid-management and commissioning capabilities into Europe.

That latest acquisition shows a pattern rather than another isolated deal. Vertiv first bought a major North American fluid-management specialist and is now adding the same type of expertise in EMEA.

Liquid systems create more work after the hardware has been installed. Coolant chemistry, contamination, trapped air, filtration, leaks and flow rates all affect performance. A badly commissioned loop can compromise expensive compute hardware even if the CDU and cold plates themselves work perfectly.

Cooling has consequently become interesting to industrial buyers at several layers at once: equipment, integration, commissioning and ongoing maintenance. The recent acquisition prices show how much strategic value large vendors now attach to owning more of that chain.

If you want more recent data on this point, please see our latest data center cooling market report.

Could higher rack densities actually reduce spending on cooling?

Higher rack densities can lower cooling cost per unit of compute, but they are unlikely to shrink the overall cooling market while total data center power demand is rising this quickly.

A rack drawing 120 kW produces roughly 120 kW of heat. Packing that compute into fewer racks saves floor space and can reduce some room-level infrastructure, but the heat still has to travel from the chips to the outside environment.

Liquid cooling can actually make that process more efficient. Capturing heat close to the chip reduces the amount of fan energy required, and warmer coolant temperatures can sometimes support free cooling or reduce the need for compressor-heavy chillers.

That changes the mix of spending. Operators may buy fewer air handlers per megawatt of compute while spending more on cold plates, CDUs, pumps, heat exchangers, controls and liquid-loop services.

Efficiency therefore puts some downward pressure on cooling cost per megawatt. Data center capacity growth is currently moving much faster in the opposite direction.

If global demand rises from roughly 82 GW in 2025 toward 200 GW or more around the end of the decade, cooling suppliers can sell less thermal infrastructure per unit of compute and still end up with a much larger market.

What could slow down the AI data center cooling boom?

Power shortages are probably the biggest near-term threat to AI cooling growth because a data center that cannot secure electricity never reaches the stage where it needs a full cooling system.

Grid connections, transformers, transmission capacity, permitting and local opposition are already delaying data center projects in several major markets. A cooling vendor can have strong underlying demand while still seeing orders move to the right when a customer's facility cannot get energized.

Rack density is another uncertainty. Semiconductor roadmaps are pushing toward extremely dense systems, but operators do not always have to deploy hardware in the most concentrated configuration available. Some inference workloads can be spread across more racks and stay on air cooling.

Better chip efficiency also changes the equation. If AI accelerators deliver much more useful computation per watt, AI usage can grow faster than the physical thermal load.

Pricing could become another pressure point. The number of companies offering CDUs, cold plates and related components has increased quickly. As interfaces become more standardized, some hardware categories will probably become less differentiated and margins could fall even while unit volumes rise.

These are meaningful risks to growth rates and profitability. They look much less threatening to the existence of the market itself. The industry is still adding enormous amounts of computing power, and almost every watt that reaches those servers eventually becomes heat that somebody has to remove.

How big could the data center cooling market become by 2030?

A global data center cooling market of roughly $30 billion to $50 billion around 2030 looks realistic, while forecasts materially above that level usually depend on broader definitions or very aggressive growth assumptions.

Current research supports that order of magnitude. MarketsandMarkets expects its relatively narrow market definition to grow from about $13.2 billion in 2026 to roughly $37.6 billion by 2033. Fortune Business Insights starts from a larger base and expects the market to exceed $50 billion by 2034. Broader research firms naturally arrive at still larger numbers.

We can test those forecasts against physical capacity instead of simply accepting a CAGR.

McKinsey's continued-momentum scenario takes global data center demand from roughly 82 GW in 2025 to about 219 GW in 2030. AI capacity alone climbs from approximately 44 GW to 156 GW.

That would mean around 137 GW of additional total data center demand in five years, including more than 110 GW of additional AI capacity.

At the same time, a growing share of that AI capacity is moving toward cooling architectures with more thermal hardware per rack. TrendForce's estimates take liquid-cooling penetration among AI chips from around one-third in 2025 to more than half in 2026, with further gains expected afterward.

Those two trends reinforce each other. The industry is adding a huge amount of capacity, while the fastest-growing part of that capacity is adopting a more expensive cooling stack.

We would therefore be comfortable with a cooling market reaching several tens of billions of dollars annually by 2030 without using the most bullish forecasts available.

So how big is the data center cooling market as AI densities rise?

The data center cooling market is roughly a $20 billion global business today, and our best estimate is that it can grow toward $30 billion to $50 billion around 2030 as AI pushes more racks beyond the practical limits of conventional air cooling.

The current market is easiest to understand as two overlapping businesses.

The first is a large and mature air-cooling market serving most of the world's installed data center capacity. Uptime's surveys still show that the vast majority of facilities have no racks above 30 kW, so conventional thermal equipment has years of demand ahead.

The second is a much smaller liquid-cooling market growing at a completely different speed. We estimate that ecosystem at roughly $5 billion to $7 billion today under a broad definition, while Dell'Oro's manufacturer-only data was still closer to $3 billion in 2025.

AI rack density is the reason the gap is closing so fast. A typical data center rack still sits around the high-single-digit or low-teens kilowatt range. NVIDIA's GB200 NVL72 runs at roughly 120 kW, and the GB300 can reach about 142 kW. TrendForce now estimates that liquid cooling reaches roughly 53% of AI chips.

The commercial evidence has also become hard to ignore. AAON's BASX business more than tripled year over year in its latest reported quarter. Modine secured more than $4 billion of future cooling commitments from one customer. Eaton agreed to spend $9.5 billion on Boyd's thermal business. Vertiv has kept buying fluid-management capabilities after its roughly $1 billion PurgeRite acquisition, most recently adding another specialist in Europe.

Taken together, the market has already moved beyond a speculative AI-infrastructure theme. Cooling companies are shipping equipment, expanding factories, building billion-dollar backlogs and buying specialist capabilities years before the densest AI architectures have reached broad deployment.

The biggest mistake would be to assume that every data center is suddenly moving to liquid. Most are not. What matters is where new money is going. The most power-hungry and expensive AI systems increasingly require liquid cooling, and those systems sit at the center of today's data center investment cycle.

That is enough to reshape a $20 billion market even if ordinary air-cooled racks remain the majority for many years.

If you want more recent data on this point, please see our latest data center cooling market report.