Which data center cooling business models will capture AI demand?

Why is AI changing the data center cooling business so much?

AI has turned data center cooling from a supporting infrastructure category into one of the things that can determine whether a new GPU cluster can be deployed at all.

NVIDIA's GB200 NVL72 puts 72 Blackwell GPUs and 36 Grace CPUs into a liquid-cooled rack-scale system. GB300 NVL72 follows the same basic direction. At the facility level, Equinix now discusses AI deployments above 200 kW per rack and future systems that could approach 1 MW. That is far beyond the roughly 5–10 kW rack densities around which a large part of the existing data center estate was built.

Cooling vendors are already designing around that gap. Schneider Electric's Motivair portfolio includes CDUs reaching 2.5 MW per unit, with coordinated systems designed to scale beyond 10 MW. Vertiv is qualifying multi-megawatt CDUs for NVIDIA AI-factory infrastructure. Airedale by Modine has introduced chillers above 3 MW specifically for large AI facilities.

Higher chip power is forcing changes from the server rack through pumping, water loops, controls, chillers and heat rejection. For the densest AI clusters today, thermal infrastructure has moved onto the critical path of the project.

Is liquid cooling really taking over data centers now?

Liquid cooling is taking over the densest AI racks, but most data centers are still nowhere close to becoming fully liquid cooled.

Uptime Institute's latest global work is a useful reality check. Typical rack densities are still moving upward gradually, even though more operators now report peak racks above 30 kW. Uptime has also found that direct liquid cooling remains concentrated in the applications where air cooling is becoming impractical.

That leaves two markets running in parallel. Frontier GPU clusters are moving rapidly toward direct liquid cooling, while most enterprise servers, storage, networking equipment and older racks still run comfortably with air.

The latest products increasingly assume that both will coexist. Schneider recently launched a CDU designed specifically to work with fan-wall air cooling so operators can change the balance between liquid and air without redesigning the entire data hall. Munters' roughly SEK 2 billion US AI cooling order combines high-capacity CDUs with over-rack computer-room air handlers. Digital Realty and Vertiv are also designing high-density environments around mixed cooling.

What does it actually mean to capture AI cooling demand?

A cooling company captures AI demand when higher GPU densities increase how much customers spend with it and that relationship survives beyond the first equipment order.

Rapid unit growth alone does not necessarily create a great business. A cold-plate supplier can benefit enormously during a new GPU generation, yet still face redesign risk whenever server architectures change. A company supplying CDUs, chillers, controls, commissioning and maintenance can earn money at several stages of the same deployment.

Recent acquisitions show where large industrial companies think the valuable territory sits. Eaton paid $9.5 billion for Boyd Thermal, which had forecast roughly $1.5 billion of liquid-cooling sales for 2026. Schneider bought control of Motivair. Trane acquired LiquidStack. Vertiv has added cold-plate engineering, heat-rejection technology and fluid-management services through several transactions.

For this article, the strongest business model means more than having the fastest-growing technology. We are looking for companies that can capture a large amount of spending per megawatt, keep reasonable pricing power and continue selling into the installed base after the AI servers go live.

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

Will direct-to-chip liquid cooling become the default for AI servers?

Direct-to-chip liquid cooling currently has the clearest path to becoming the standard cooling architecture for very dense AI racks.

The attraction is practical. Cold plates take heat directly from GPUs and CPUs while the rest of the server or rack can still use air where needed. Operators can therefore handle far more heat without putting the entire machine into dielectric fluid or completely changing normal server servicing.

The hardware ecosystem is moving the same way. NVIDIA's GB200 and GB300 rack-scale systems are designed around liquid cooling. Schneider, Vertiv, Eaton through Boyd, Trane through LiquidStack, Munters and several specialists are all building direct-to-chip infrastructure around the same demand.

It also fits the retrofit market better than more radical architectures. Operators can add CDUs, manifolds and liquid loops to selected rows while leaving lower-density equipment alone.

Direct-to-chip looks strongest where chip heat is high enough to justify the added plumbing and operational complexity. That category is growing quickly because of AI accelerators, but it still represents only part of the broader data center installed base.

Cooling approach	Density it can handle	Changes to existing servers	Retrofit fit	Commercial position today
Conventional air cooling	Low to medium	Minimal	Excellent	Still dominant outside dense AI
Rear-door heat exchangers	Medium	Minimal	Very good	Useful transition technology
Direct-to-chip liquid cooling	High	Moderate	Good	Strongest AI growth market
Liquid-to-air CDU systems	High	Moderate	Very good	Strong where facility water is limited
Liquid-to-liquid CDU systems	Very high	Moderate	Good	Core large AI-cluster architecture
Single-phase immersion	Very high	Major	Weak	Specialized
Two-phase immersion	Very high	Major	Weak	Technically strong, adoption still limited
Hybrid air plus direct liquid	High	Moderate	Excellent	Likely broadest deployment model

Will immersion cooling beat direct-to-chip cooling for AI?

Immersion cooling should win some extreme-density applications, but direct-to-chip looks much more likely to capture the larger AI cooling market.

Immersion has genuine technical advantages. Putting computing hardware directly into dielectric fluid can remove large amounts of heat in a compact space. The harder part is getting customers to change server formats, maintenance procedures, component handling and sometimes warranty arrangements.

Direct-to-chip asks customers to add a liquid loop around familiar servers. Immersion changes much more of the operating model.

Commercial behavior reflects that difference. LiquidStack, one of the best-known immersion companies, expanded into direct-to-chip systems before being acquired by Trane. Trane now markets both technologies inside a much broader thermal portfolio. Large data center operators also tend to present immersion as one available architecture while direct-to-chip receives the heavier emphasis for mainstream high-density deployments.

Uptime Institute's research still points to operational complexity as one of the main things slowing direct liquid cooling outside the highest-density applications. Immersion adds another layer of change on top of that.

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

Are coolant distribution units the cleanest way to bet on AI cooling?

CDUs are probably the clearest equipment play on the move toward liquid-cooled AI, although competition should get much tougher as the category matures.

A coolant distribution unit sits between the facility cooling system and the liquid loop serving the servers. It controls temperature, pressure and flow while keeping the sensitive IT-side loop separate from the building's water system. Once a data center shifts a large amount of heat into liquid, CDUs become difficult to avoid.

Product capacities show how quickly the category is moving. Schneider's Motivair line now reaches 2.5 MW per CDU and can coordinate systems above 10 MW. Vertiv has a 2.3 MW CDU qualified for NVIDIA's DSX infrastructure program. Munters' roughly SEK 2 billion US order includes custom high-capacity CDUs for a large AI build-out.

There is a fairly direct relationship between liquid-cooled megawatts and the amount of distribution capacity required. That gives CDU vendors strong exposure to AI infrastructure spending.

The catch is already visible: almost every serious cooling platform wants to sell them. Schneider, Vertiv, Munters, Airedale, LiquidStack and smaller specialists are competing in the same category. NVIDIA qualification programs and industry standards should also make it easier over time to compare systems from different suppliers.

Could chillers make more money from AI than cold plates and CDUs?

Yes. Chillers and heat-rejection systems could capture more absolute AI cooling spending than some of the liquid-cooling products getting far more attention.

Every watt of GPU heat ultimately has to go somewhere. Cold plates move heat away from the processors and CDUs transfer it between loops, but large facilities still need major equipment outside the server room to reject that heat.

Recent orders are already much larger than what we usually see in specialist rack-level cooling. Modine signed a long-term agreement giving one strategic customer access to more than $4 billion of Airedale cooling products over 2027–2029. The customer paid $165 million upfront to support the capacity needed to fulfill that commitment.

Munters has also reported billion-krona data center cooling orders, including large chiller contracts, while its Data Center Technologies backlog reached SEK 15.2 billion after more than doubling year over year in its first-quarter reporting. Vertiv is currently doubling chiller production capacity at its Tognana site in Italy.

Those numbers make the economics easier to see. AI is creating a rapidly growing market for CDUs and cold plates, but it is also pushing huge amounts of spending into industrial-scale chilled-water and heat-rejection equipment.

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

Why are hybrid air-and-liquid cooling systems winning now?

Hybrid cooling is becoming the practical default for many AI data centers because most facilities do not need liquid cooling on every rack.

A modern AI hall can contain liquid-cooled GPU systems beside networking, storage and less dense compute that still works perfectly well with air. Existing data centers make this mixture even more obvious because operators rarely replace thousands of functioning air-cooled racks simply to install a few rows of new accelerators.

Recent product development is lining up around that reality. Schneider's newest WCDU is designed to work directly with fan-wall air systems and lets operators change the balance between air and liquid cooling later. Munters' large AI order pairs CDUs with over-rack air handlers. Vertiv offers modular systems combining direct-to-chip liquid cooling with air cooling.

Uptime Institute's latest data adds another reason to be cautious about an all-liquid future. Typical rack densities across the industry are still far below the extreme numbers associated with the newest GPU systems. Its current view is that liquid cooling should remain concentrated in high-density applications over the next several years.

Is retrofitting old data centers a real AI cooling business?

Yes. Retrofitting existing data centers should create a large market for cooling engineering and integration because most older facilities were never built for today's AI racks.

The gap can be enormous. A facility designed around 5–10 kW racks can require substantial changes before it can host racks drawing many tens or hundreds of kilowatts. The work can involve piping, pumps, CDUs, electrical distribution, chillers, heat exchangers, controls and new operating procedures.

That makes retrofit cooling very different from simply shipping equipment.

Vertiv's work installing NVIDIA DGX GB300 infrastructure at the US Naval Postgraduate School provides a useful example. The project combined power protection, racks, liquid cooling, fluid management, installation, testing and commissioning inside an existing environment. The value came from making the whole deployment work together.

Colocation operators are solving the same problem on behalf of customers. Equinix has been adding liquid-cooling capability across its footprint so companies can deploy dense AI hardware without rebuilding their own facilities first.

A standardized CDU can eventually face intense price competition; an engineering team that knows how to put new liquid infrastructure safely inside an awkward 15-year-old facility is harder to replace.

Are chip-to-chiller cooling platforms the strongest AI cooling business model?

Yes. Companies that can cover most of the path from the server to heat rejection currently have the strongest overall position in AI cooling.

Look at where the large industrial buyers have been spending. Schneider combined its power and cooling portfolio with Motivair's CDUs and liquid infrastructure. Eaton completed the $9.5 billion acquisition of Boyd Thermal, adding cold plates and other liquid-cooling technology to a huge data center power business. Trane acquired LiquidStack to connect direct-to-chip and immersion cooling with chillers, heat rejection and controls.

Vertiv has followed a similar route through multiple smaller transactions. It added Strategic Thermal Labs for cold-plate and server-side engineering, ThermoKey for heat rejection and more recently agreed to acquire King Environmental Services for fluid management and commissioning in Europe, the Middle East and Africa.

The pattern has become even clearer lately. Trane's newest 250 MW AI-factory reference designs combine its own thermal equipment with LiquidStack and Stellar Energy technologies. Schneider's newest hybrid CDU links liquid distribution directly with fan-wall air cooling.

Software fits naturally into the same model. Schneider connects liquid cooling into EcoStruxure, while Vertiv is developing monitoring, digital-twin and infrastructure-management tools around its hardware. Controls become more valuable as pumps, CDUs, chillers and variable AI loads need to react together.

Business model	What it sells	Main advantage in AI	Main weakness	Current position
Cold-plate specialist	Cold plates and manifolds	Very close to GPU demand	Server-design and qualification risk	Fast-growing
CDU specialist	Distribution, pumping and controls	Scales with liquid-cooled MW	More vendors entering	Very strong near term
Chiller specialist	Chillers and heat rejection	Huge facility-level spending	Less unique to AI	Underappreciated winner
Integrated thermal platform	Rack-to-facility cooling	Captures several spending layers	Harder to execute	Strongest overall model
Retrofit integrator	Design, piping and commissioning	Legacy facilities are difficult	Labor intensive	Attractive fragmented market
Fluid-management provider	Flushing, filtration, testing	Recurring installed-base demand	Smaller initial contract size	Growing quickly
AI-ready colocation	High-density capacity	Sells cooling capability as a service	Extremely capital intensive	Strong enterprise model
Immersion specialist	Immersion tanks and fluids	Handles extreme density	Bigger operational change	Valuable niche

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

Can specialist liquid-cooling companies still beat Schneider, Vertiv or Trane as cooling hardware standardizes?

Specialist liquid-cooling companies can still build valuable businesses, but they need stronger technology or customer lock-in as CDUs and cold plates become easier to compare.

Specialists often move faster technologically. Cold plates, dielectric fluids, immersion systems and new distribution architectures can come from focused engineering teams long before they become priorities for large industrial companies.

The problem appears when a hyperscaler wants thousands of units across several continents. Qualification, manufacturing capacity, spare parts, service engineers and integration with the rest of the thermal plant suddenly matter almost as much as the component itself.

The acquisition market shows how this usually plays out. Schneider bought Motivair. Eaton bought Boyd Thermal. Trane bought LiquidStack. Vertiv bought Strategic Thermal Labs rather than relying entirely on internal development.

Boyd shows how valuable a specialist can become before being absorbed. Eaton's $9.5 billion purchase price represented 22.5 times Boyd Thermal's estimated 2026 adjusted EBITDA. Boyd was forecasting $1.7 billion of sales for the year, including roughly $1.5 billion from liquid cooling.

At the same time, standardization is starting to reduce some of the protection around individual products. Programs such as NVIDIA DSX Ready give customers clearer qualification requirements and make it easier to compare suppliers. Interoperability should help liquid cooling spread faster, but it also gives purchasing departments more bargaining power once several vendors can perform the same job.

Cold plates could eventually face similar pressure if interfaces and thermal requirements become more standardized across server generations.

The specialist businesses with the best chance are the ones that become difficult to replace before the category commoditizes: companies with deep server-OEM relationships, qualified designs, genuinely better thermal performance or enough manufacturing scale to serve very large deployments.

Are cooling maintenance and fluid services the best recurring-revenue business?

Cooling maintenance, commissioning and fluid management may become the most attractive recurring-revenue layer created by AI liquid cooling.

Liquid systems introduce pumps, filters, valves, hoses, seals, heat exchangers and coolant chemistry close to servers worth millions of dollars. Contamination, corrosion, trapped air, poor flow balancing or bad fluid quality can cause performance problems that operators have little appetite to experiment with.

The market is already moving beyond installation. Vertiv acquired PurgeRite to expand fluid-management capabilities in North America and recently agreed to acquire King Environmental Services, which provides flushing, filtration, water treatment, commissioning, load testing and related services in Europe.

The timing is important. The first large wave of direct-liquid infrastructure is now creating an installed base that will need attention for years. Every additional megawatt deployed adds future maintenance rather than simply another one-time hardware sale.

Munters and other large cooling suppliers are also attaching startup and commissioning work to equipment contracts. Retrofit projects naturally create even more service work because older facilities require custom testing and integration.

Margins will vary, and this market is smaller in upfront dollars than multi-megawatt chiller orders. Its quality is potentially better, though. Once a provider is trusted to maintain the fluid system surrounding very expensive GPUs, changing contractors to save a small amount of money becomes less appealing.

Will water shortages change which AI cooling companies win?

Yes. Water availability is becoming important enough to influence cooling design, especially for giant AI campuses in places where water use already creates permitting or community problems.

The opportunity is more complicated than a race to claim “waterless” cooling. Operators still have to manage electricity use, capital costs, reliability and outside-air conditions.

Large suppliers are now designing around those trade-offs. Trane recently introduced two 250 MW AI-factory reference designs built around zero-water cooling. The designs combine technologies from Trane, LiquidStack and Stellar Energy and are meant for NVIDIA DSX-based facilities. Johnson Controls has also developed large zero-water reference architectures for AI data centers.

Dry cooling and air-cooled chillers can cut water use dramatically, although they may carry an energy or equipment penalty under certain weather conditions. Direct-to-chip systems can also use closed loops while changing how heat is ultimately rejected outside the building.

If access to water becomes one of the reasons a 500 MW campus cannot be permitted, a cooling architecture that avoids that constraint can be worth far more than the difference in equipment cost.

Is manufacturing capacity becoming part of the AI cooling business?

Yes. For large AI projects today, being able to guarantee thousands of cooling units on time is becoming a product in its own right.

Modine provides the clearest example. One strategic customer agreed to secure more than $4 billion of future Airedale cooling capacity for deliveries over 2027–2029 and paid $165 million upfront to help Modine build the capacity required.

That arrangement says something unusual about the market. The customer was willing to finance part of its supplier's expansion to reduce the risk that cooling equipment becomes a bottleneck later.

Vertiv is making the same constraint visible from the supplier side. The company is doubling chiller production capacity at its Tognana campus and has opened additional manufacturing capacity in Johor, Malaysia, for power, cooling and integrated infrastructure serving Asian AI demand.

Munters' growing backlog adds another piece. Its data center business has collected several very large multi-year orders, including the roughly SEK 2 billion AI cooling project combining CDUs and air handlers.

These orders are no longer just about who has the best thermal specification. Customers also need confidence that a supplier can build, test, ship and service equipment across a multi-hundred-megawatt rollout.

Which AI data center cooling business models have the best economics?

Integrated thermal platforms look strongest overall, while fluid services offer the cleanest recurring revenue and CDUs provide some of the highest direct exposure to near-term liquid-cooling growth.

The differences come from how each model gets paid. Cold-plate and CDU suppliers benefit immediately when more GPU racks move to liquid. Chiller vendors capture large facility-level orders. Integrated platforms can sell several of those products into the same project. Service companies collect smaller amounts repeatedly after installation.

Immersion has attractive technical economics in specific environments but still carries more adoption risk. Conventional air cooling remains a large business, particularly inside hybrid facilities, although its revenue is less directly tied to the most extreme AI density growth.

Business model	Exposure to AI demand	Recurring revenue	Pricing protection	Capital needed	Our assessment
Integrated thermal platform	Very high	High	High	High	Best overall model
Direct-to-chip components	Very high	Medium	High for qualified designs	Medium	Excellent growth
CDU platforms	Very high	Medium	Medium	Medium	Major near-term winner
Chillers and heat rejection	Very high	High	Medium-high	High	Underestimated winner
Retrofit engineering	High	High	High	Low-medium	Very attractive fragmented market
Fluid management and service	High	Very high	High	Low-medium	Best recurring-revenue layer
AI-ready colocation	Very high	Very high	High	Very high	Strong but infrastructure-heavy
Immersion cooling	Medium	Medium	Potentially high	Medium	Valuable specialist market
Air-only cooling	Medium	Medium	Medium	Medium	Large market, weaker direct AI leverage

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

Which data center cooling business models will capture the most AI demand?

Integrated chip-to-chiller platforms should capture the largest and most durable share of AI cooling demand, with CDUs, large chillers and lifecycle services emerging as the strongest individual businesses underneath them.

The evidence has become unusually consistent. Eaton paid $9.5 billion for Boyd Thermal to connect power infrastructure with liquid cooling. Schneider added Motivair and is now linking CDUs, air cooling and controls. Trane bought LiquidStack and is already using it inside 250 MW AI-factory thermal designs. Vertiv has expanded in both directions, adding server-side engineering, heat rejection and fluid-management services.

Direct-to-chip should capture the fastest structural growth inside the rack. CDUs sit in an excellent position immediately behind it. Large chillers and heat-rejection systems may collect even more dollars at the facility level, especially as AI campuses move from tens of megawatts toward hundreds.

Hybrid cooling should also capture more demand than an all-liquid thesis suggests. Uptime Institute's latest work still shows a much less dense broader data center market, while Schneider's newest hardware and Munters' recent AI orders explicitly combine liquid and air cooling.

Services become increasingly attractive as that installed base expands. Fluid management, commissioning, maintenance and monitoring generate revenue long after the first rack has been delivered.

Manufacturing scale provides the final advantage. Modine's $4 billion capacity agreement and the factory expansions underway at Vertiv show that AI developers increasingly value guaranteed delivery alongside cooling performance.

The conclusion is clear today: selling a clever cooling component can build a fast-growing company, but controlling more of the path from the GPU to the outside environment creates the stronger business.