Which data center cooling startups are winning the AI buildout?

Why is AI suddenly forcing data centers to switch to liquid cooling?

AI racks are now getting so hot that liquid cooling has moved from an efficiency upgrade to basic infrastructure for the highest-density systems.

NVIDIA's GB200 NVL72 pushed roughly 120 kW into a single rack, several times what a conventional data hall was built to handle. Successive AI systems are pushing the problem further because chip power is rising alongside the number of accelerators packed into each rack.

The market has reacted quickly. Dell'Oro estimated that the liquid-cooling market roughly doubled to close to $3 billion in 2025 and now expects manufacturer revenue to approach $7 billion by 2029. More importantly, Dell'Oro says single-phase direct liquid cooling already accounts for the vast majority of liquid-cooled capacity coming online.

A clever cold plate or impressive demonstration is no longer enough. Customers need thousands of components produced consistently, qualified against expensive AI hardware and supported inside facilities that may ultimately consume hundreds of megawatts.

The heat problem has become a manufacturing and infrastructure problem as much as a thermal-engineering one.

What does “winning” the AI cooling race actually mean?

The cooling companies winning today are the ones getting designed into real AI infrastructure at scale, especially when those deployments turn into repeat orders, major hardware partnerships or acquisitions by large industrial groups.

That is a more useful test than comparing the maximum number of kilowatts each product claims it can cool.

LiquidStack's 300 MW multi-site CDU order tells us far more about commercial position than an isolated laboratory demonstration. Submer says it has passed 500 MW of deployed liquid-cooled infrastructure. ZutaCore says its technology has reached more than 75 environments. Accelsius says customer deployments cover hundreds of megawatts of new capacity.

Another group has already reached a different kind of finish line. Flex bought JetCool. Trane Technologies bought LiquidStack. Daikin bought Chilldyne. Schneider Electric took control of Motivair. Those acquisitions brought startup-origin technology into companies that already manufacture, install and service infrastructure around the world.

There are really two races happening at once. One is between independent specialists trying to become large cooling companies. The other is between technologies being absorbed into much larger chip-to-chiller platforms.

Has direct-to-chip cooling already beaten immersion cooling for AI?

Direct-to-chip liquid cooling is currently winning the architecture battle by a wide margin.

Dell'Oro says single-phase direct liquid cooling represents the vast majority of liquid-cooled capacity coming online and expects it to remain the dominant approach through most of the decade. Immersion is still being deployed, but much more selectively.

The reason is fairly practical. Direct-to-chip lets operators keep racks, boards, cables and much of the normal server-service model while replacing the thermal path around the processors. Immersion asks them to operate servers inside dielectric fluid and changes much more of the workflow around maintenance and hardware handling.

NVIDIA's own rack-scale systems have accelerated this preference. Once a dominant AI platform is built around cold plates and coolant loops, OEMs, CDU suppliers and data-center operators have a strong incentive to standardize around the same architecture.

The behavior of former immersion specialists is revealing too. LiquidStack built its reputation around immersion but has pushed aggressively into direct-to-chip CDUs. Submer still sells immersion systems while increasingly offering direct-to-chip infrastructure and complete AI deployments.

Immersion remains useful where extreme density, space constraints or full heat capture justify the operational changes. It just isn't where most of the hyperscale AI money is going today.

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

Which AI cooling companies have the strongest evidence of real scale?

LiquidStack, Submer and Motivair have some of the strongest public evidence of scale, while ZutaCore and Accelsius are further behind in installed volume but moving quickly in technologies aimed at hotter future chips.

The figures need to be read carefully because companies disclose different things. Submer's 500 MW-plus number refers to deployed liquid-cooled infrastructure. LiquidStack disclosed a 300 MW multi-site customer order. Accelsius talks about customer deployments spanning hundreds of megawatts of new capacity, although it has not publicly separated commissioned systems from capacity still being built.

ZutaCore gives another kind of figure: more than 75 deployed environments rather than megawatts. That makes a precise ranking impossible, but it still gives us enough evidence to separate companies with meaningful field exposure from companies living mostly in pilot projects.

Manufacturing capacity adds another layer. CoolIT Systems says it has multi-gigawatt production capability and has worked with NVIDIA on liquid cooling for years. Motivair now offers individual CDUs reaching 2.5 MW and architectures that can scale beyond 10 MW.

The current leaders are being tested less by whether the cooling works and more by how much identical equipment they can deliver without disrupting a very expensive AI deployment.

Company	Strongest public evidence	Main approach	Position today
Submer	500 MW+ of liquid-cooled infrastructure reported deployed	Immersion and direct-to-chip	Major independent platform
LiquidStack	300 MW multi-site CDU order	Direct-to-chip and immersion	Acquired by Trane
Motivair	Large AI deployments plus CDUs scaling to multi-MW systems	Single-phase direct-to-chip	Controlled by Schneider Electric
Accelsius	Customer deployments spanning hundreds of MW of new capacity	Two-phase direct-to-chip	Fast-rising independent
ZutaCore	75+ deployed environments	Two-phase direct-to-chip	Leading independent contender
CoolIT Systems	Multi-gigawatt manufacturing capacity	Single-phase direct-to-chip	Large specialist owned by KKR
JetCool	Dell server integration and Broadcom XPU work	Single-phase direct-to-chip	Acquired by Flex
Chilldyne	Long-running production GPU deployments	Negative-pressure direct-to-chip	Acquired by Daikin

Is LiquidStack winning the AI cooling race right now?

LiquidStack has the strongest claim to being the clearest startup-origin commercial winner right now.

The 300 MW multi-site CDU order remains the company's strongest piece of evidence. To understand the scale, 300 MW is equivalent to the thermal load of roughly 2,500 racks consuming 120 kW each. That does not mean the customer is installing exactly that configuration, but it puts the order in the right order of magnitude.

LiquidStack has also followed the market rather than clinging to the technology that first made it famous. The company expanded heavily from immersion into direct-to-chip cooling and developed GigaModular CDU systems that can scale into double-digit megawatts.

Then Trane Technologies bought the company.

What happened next makes the acquisition more convincing. Trane has already incorporated LiquidStack technology into new 250 MW AI-factory reference designs based on NVIDIA's DSX platform. LiquidStack is moving from selling specialist cooling equipment to sitting inside complete designs for enormous AI campuses.

That position is stronger than simply having a large order book. Trane can combine LiquidStack's chip and rack cooling with chillers, heat rejection, controls and worldwide service.

The obvious downside is that LiquidStack is no longer an independent startup. From a technology and commercialization perspective, though, very few cooling startups have travelled further.

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

Is Submer still a winner if immersion cooling is losing share?

Submer is still one of the strongest independent companies in AI cooling because it has stopped making its future depend entirely on immersion.

The company says it has surpassed 500 MW of deployed liquid-cooled infrastructure, giving it one of the largest publicly claimed installed bases among specialist vendors. Its roughly $100 million of cumulative reported funding has also given it more room to expand than many smaller cooling startups.

What makes Submer interesting these days is the change in what it wants to become.

The company increasingly sells modular AI infrastructure, GPU systems, data-center development and direct-to-chip solutions alongside immersion. One of its newer edge-AI offerings can scale from smaller deployments into megawatt-class containerized infrastructure holding hundreds of GPUs.

That broadening looks sensible given what has happened to immersion adoption. Submer can still use immersion where it gives customers a clear advantage without requiring immersion to become the standard cooling method for every AI data center.

The trade-off is that Submer is becoming harder to compare with companies such as ZutaCore. It increasingly looks like an AI infrastructure company with deep cooling expertise rather than a pure cooling startup.

That may ultimately be a better business.

Is ZutaCore the strongest independent two-phase cooling startup?

ZutaCore currently has the strongest combination of funding, deployed systems and heavyweight industrial backers among independent two-phase cooling companies.

Its latest major financing brought in more than $100 million and included Mitsubishi Electric, Carrier Ventures and Samsung. Those names are unusually relevant. They bring manufacturing, cooling infrastructure and electronics relationships rather than just capital.

ZutaCore says its waterless two-phase technology is now deployed in more than 75 environments around the world. The company is also expanding its product range beyond chip-level components, including end-of-row systems rated at 1.2 MW and 2 MW.

Its basic bet is that conventional water-based cold plates eventually become awkward as individual accelerators move deeper into multi-kilowatt power consumption. ZutaCore circulates dielectric refrigerant directly through a sealed system at the chip, where phase change carries heat away without putting water close to electronics.

The problem is timing.

Dell'Oro currently expects single-phase cooling to remain dominant because today's operators already know how to build and maintain it. Two-phase adoption should become more attractive as heat flux rises, but reaching that crossover point could take several hardware generations.

ZutaCore therefore has a stronger position in the future cooling race than its current market share alone would suggest. If chips keep getting substantially hotter, the market starts moving toward the problem ZutaCore has spent years solving.

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

Is Accelsius catching ZutaCore?

Accelsius is now close enough to ZutaCore that we should treat the two companies as the main independent two-phase contenders rather than placing ZutaCore in a category of its own.

Accelsius raised a $65 million Series B led by Johnson Controls with participation from Legrand. It has since brought its MR250 row-level CDU into general availability and launched the IR150, an integrated rack capable of handling 150 kW of cooling.

The company also says its customer deployments span hundreds of megawatts of new capacity. We should be cautious with that figure because Accelsius has not publicly given a clean split between equipment already operating and larger projects still under development.

Even with that caveat, the commercial trajectory is unusually fast for a young cooling company.

Johnson Controls and Legrand make the funding round more useful than its dollar amount suggests. Both already sell deeply into the physical infrastructure around data centers. They can introduce Accelsius to customers that a standalone startup would otherwise spend years trying to reach.

ZutaCore still has the clearer public deployment history. Accelsius has the stronger recent acceleration.

If two-phase direct-to-chip begins moving from early adoption into mainstream AI facilities, we would expect these two companies to be among the first specialists operators seriously compare.

Has JetCool already won by getting bought by Flex?

JetCool has already achieved a strong commercial outcome because Flex is turning its cooling technology into something that can be manufactured alongside AI servers at global scale.

Flex acquired JetCool for roughly $53 million in estimated consideration, which looks modest compared with some later cooling transactions. The important part came afterward.

JetCool's technology moved into Dell PowerEdge systems. More recently, JetCool worked with Broadcom on cooling for next-generation AI XPUs capable of sustained multi-kilowatt operation at heat fluxes reaching 4 W/mm². Flex provides the mass-production infrastructure behind that work.

JetCool has also launched a turnkey liquid-cooled Dell PowerEdge system that bundles the server, cooling, commissioning, maintenance and warranty into one deployment.

That last step tells us where AI cooling is going.

Enterprise customers usually do not want to assemble a cooling system from six vendors and then work out who is responsible when something fails. A tested server-and-cooling package with one accountable supplier removes a large part of that friction.

JetCool is more interesting today as evidence that cooling is moving directly into the server-manufacturing supply chain than as a startup acquisition story.

Did Chilldyne solve one of liquid cooling's biggest practical fears?

Chilldyne built a genuinely different answer to coolant leaks, and Daikin decided the technology was valuable enough to buy.

Most direct-to-chip systems push coolant through equipment at positive pressure. Chilldyne operates its water loop below atmospheric pressure, so damage to a line tends to pull air inward rather than force water outward.

For data-center operators, that addresses a simple fear: liquid sitting close to processors worth tens of thousands of dollars each.

Chilldyne also has a longer operating history than many newer cooling startups. Cirrascale has used its technology on GPU clusters since 2014, while Sandia National Laboratories has operated Chilldyne equipment for years.

Daikin acquired the company after previously buying DDC Solutions, whose high-density cabinets had already been integrated with Chilldyne's cooling technology.

The combination gives Daikin an increasingly complete route from chip cooling to room and facility infrastructure.

We still do not know whether negative-pressure cooling becomes a widely adopted architecture. Daikin does not need that to happen for the deal to work. Chilldyne gives it an unusual technology that can now be offered through a global cooling company to customers nervous about putting water near expensive AI hardware.

Is Iceotope keeping up with the AI cooling leaders?

Iceotope is still technically credible, but today we do not see enough public commercial evidence to put it alongside the leaders.

The company has raised another $26 million and continues to develop precision liquid cooling that surrounds electronics inside sealed chassis rather than placing entire racks into large immersion tanks.

Iceotope has also built a respectable validation network involving server vendors and industrial companies. Its technology has been tested around systems from companies such as Lenovo and HPE, while fluid and thermal work has involved groups including Shell, PETRONAS and Valvoline.

The missing piece is scale.

We can point to a 300 MW LiquidStack order, more than 500 MW reported by Submer, dozens of ZutaCore environments and large Accelsius customer programs. Iceotope has not disclosed anything similarly convincing about current AI deployment volume.

That does not make the technology weak. It makes Iceotope harder to call a present-day winner.

The company could still become strategically attractive to an OEM or cooling incumbent that wants chassis-level liquid-cooling intellectual property. Based on what is public now, though, we would keep Iceotope in the contender group rather than the leading group.

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

Are NVIDIA, AMD and Broadcom partnerships deciding who wins AI cooling?

Chip and server partnerships are increasingly deciding which cooling technologies reach volume because thermal engineering now begins before the AI system leaves the drawing board.

CoolIT is one of the clearest examples. Its relationship with NVIDIA stretches back many years, and the company now supplies cooling components around Blackwell-class infrastructure while operating manufacturing capacity it describes in gigawatt terms.

JetCool is moving closer to the same point with Broadcom by designing thermal hardware around next-generation XPUs rather than waiting for finished chips and adapting later.

Motivair has also moved deeper into the hardware roadmap. Schneider Electric and AMD have developed a validated reference design for AMD's Helios rack-scale platform using Motivair cooling technology.

That kind of integration is difficult for smaller competitors to displace later. Once cold plates, pumps, flow rates, mechanical clearances and CDUs have been engineered around a specific rack, changing supplier creates new testing and reliability work.

It also explains why the industry giants keep buying cooling startups. Chip-level cooling is moving upstream into product design while facility cooling remains downstream in chillers and heat rejection. Owning both ends lets Schneider, Trane, Daikin and others participate across the whole thermal system.

Cooling startups that secure early design work with chip and server companies therefore have a much better chance of surviving the consolidation ahead.

Will two-phase cooling replace today's water-based systems?

Two-phase cooling should gain ground as AI chips get hotter, but single-phase water-based direct cooling is likely to remain the workhorse for years rather than disappear suddenly.

Current deployments strongly favor single-phase systems, and operators now have an expanding base of engineers, spare parts, qualified suppliers and operating procedures built around them.

The technology is improving as well. Larger cold plates, better internal flow paths, warmer-water operation and more capable CDUs keep pushing the practical limit higher.

That forces ZutaCore, Accelsius and other two-phase suppliers to clear a demanding bar. Better heat transfer alone will not make customers switch. Their systems need to be easier or cheaper to operate once conventional systems begin struggling.

The physics may eventually help them. Dell'Oro expects leading-edge GPU TDP to move above 4,000 W before the end of the decade. At some point, continuously pushing more water through increasingly hot chips becomes unattractive.

The interesting question is how quickly that point arrives.

If single-phase engineering keeps pace with semiconductor power, two-phase cooling stays relatively specialized. If chip heat rises faster, two-phase adoption could move quickly because operators will have fewer alternatives.

For now, we would bet on coexistence, with single-phase dominating current deployments and two-phase gaining relevance first in the hottest systems.

Are cooling startups really beating Vertiv and the industrial giants?

Cooling startups are shaping the technology, but industrial incumbents are increasingly capturing the commercial upside.

Dell'Oro currently places Vertiv at the top of the liquid-cooling market and identifies established specialists such as CoolIT, nVent and Boyd among the stronger suppliers.

Meanwhile, the acquisition list keeps growing. As seen above, Trane bought LiquidStack, Flex bought JetCool, Daikin bought Chilldyne and Schneider took control of Motivair. Eaton has also moved deeper into thermal management through Boyd.

This consolidation makes sense because AI customers increasingly want complete systems.

A hyperscale operator may need cold plates, manifolds, CDUs, controls, chillers, heat rejection, commissioning and maintenance across several countries. A startup can be excellent at one of those layers and still struggle to provide everything surrounding it.

The large infrastructure companies have the opposite problem. They already know how to build and service data centers globally but historically stopped well before the processor.

Buying liquid-cooling specialists closes that gap quickly.

We therefore expect more cooling technology to originate inside startups while an increasing share of the eventual revenue flows through much larger infrastructure groups.

Which data center cooling startups are actually winning the AI buildout?

LiquidStack is the clearest startup-origin winner today, Submer is the strongest broad independent platform, and ZutaCore and Accelsius are the two independent companies we would watch most closely as AI chips get hotter.

LiquidStack has the best combination of hard commercial evidence and strategic adoption. Its 300 MW order already put it into serious deployment territory, while Trane's acquisition has now carried LiquidStack into 250 MW NVIDIA-based AI-factory reference designs. It has effectively graduated from cooling startup to part of a global thermal platform.

Submer remains the most interesting large independent because it has substantial reported deployed capacity and has adapted its strategy as immersion lost the race to become the standard hyperscale architecture. Its broader move into modular AI infrastructure gives it more ways to grow.

ZutaCore currently leads the independent two-phase group on accumulated proof. More than $100 million of fresh capital, backing from Mitsubishi Electric, Carrier and Samsung, and more than 75 reported deployment environments give the company a credible commercial base.

Accelsius is closing quickly. Its $65 million funding round, Johnson Controls and Legrand relationships, commercially available systems and claimed customer programs spanning hundreds of megawatts make it the company most capable of changing this ranking over the next stage of the buildout.

JetCool and Chilldyne belong in the winner group too, even though neither remains independent. Their acquisitions by Flex and Daikin show that useful cooling technology can create strategic value long before a startup becomes a billion-dollar standalone business.

Motivair is further ahead commercially than most companies on this list, although calling it a startup now stretches the term. Schneider has turned it into the liquid-cooling core of a much broader AI infrastructure portfolio, and its recent AMD Helios work reinforces that position.

The biggest conclusion from our review is that the AI cooling race is already becoming less startup-like. Scale, chip partnerships, manufacturing and worldwide service are taking over from pure thermal novelty as the main competitive advantages.

That favors LiquidStack's path through Trane and Motivair's path through Schneider today. Among companies still standing on their own, Submer has the broadest platform, while ZutaCore and Accelsius have the most interesting exposure to the next cooling problem.

Company	Our view now	What puts it there	What could change the ranking
LiquidStack	Clearest current startup-origin winner	300 MW order, Trane acquisition, NVIDIA AI-factory integration	Execution at Trane-scale deployments
Submer	Strongest broad independent	500 MW+ reported deployments and wider AI infrastructure strategy	Direct-to-chip growth versus immersion exposure
ZutaCore	Leading independent two-phase company	$100M+ funding, 75+ environments, major strategic investors	Large hyperscale two-phase deployments
Accelsius	Fastest-rising challenger	$65M round, Johnson Controls and Legrand, large customer programs	Converting announced capacity into operating installations
Motivair	Major commercial winner	Schneider scale, multi-MW CDUs, AMD Helios integration	Already better viewed as part of Schneider
JetCool	Strong strategic winner	Flex manufacturing, Dell systems, Broadcom XPU work	Volume achieved through OEM channels
Chilldyne	Strong specialist outcome	Negative-pressure technology and Daikin ownership	Wider adoption of its architecture
Iceotope	Credible contender	Distinctive chassis cooling and strong validation work	Needs clearer evidence of deployment scale

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