As artificial intelligence pushes computing power and density to unprecedented levels, traditional air cooling solutions are rapidly becoming obsolete. The future of high-performance computing, particularly for AI, hinges on innovative liquid cooling technologies that can efficiently manage extreme heat, enabling the continued advancement and scalability of our digital infrastructure.
Step into almost any traditional data center, and the first thing you notice is the pervasive hum of thousands of fans. These diligent machines tirelessly push hot air away from sensitive computer chips and towards air-conditioning units. For decades, this air-cooling paradigm served the industry well. However, with the advent of next-generation artificial intelligence (AI) and high-performance computing (HPC), this familiar sound is becoming a relic of a bygone era.
The new generation of AI servers has dramatically increased its power consumption and density. As single-cabinet power continues its ascent, traditional air cooling is unequivocally approaching its limits in terms of energy efficiency and heat dissipation. This foundational shift necessitates a strategic embrace of liquid cooling technologies across the data center landscape.
The Unmanageable Heat: AI’s Power Explosion
The primary catalyst for this cooling revolution is the exponential growth in the power density of advanced computer chips, particularly Graphics Processing Units (GPUs) vital for AI workloads. In just a few years, the wattage consumed by these chips has skyrocketed. For instance, Nvidia’s V100 GPU in 2017 drew 300 watts; by 2020, the A100 reached 400 W; and the popular H100, released in 2022, consumes up to 700 W. Looking ahead, the newest Blackwell GPUs, unveiled in 2024, push this figure to an astonishing 1,200 W, with industry road maps projecting chips exceeding 2,000 W, and even 5 kilowatts in the foreseeable future, as reported by IEEE Spectrum.
This power explosion is a direct consequence of AI’s voracious appetite for computation. All these extra calculations, consuming vast amounts of power from advanced chips, translate into immense amounts of heat. As Josh Claman, CEO of Accelsius, highlights, the average power density in a rack has jumped from around 8 kW to an astounding 100 kW for AI applications, an order of magnitude increase that demands an urgent re-evaluation of cooling strategies. It’s this rapid AI adoption that is creating the real urgency for better cooling solutions.
The superiority of liquid over air for heat transfer is undeniable. Water, for example, possesses roughly four times the specific heat of air and is approximately 800 times as dense. This means it can absorb around 3,200 times more heat than a comparable volume of air. Furthermore, water’s thermal conductivity is 23.5 times higher than air, allowing heat to transfer much more readily. As Seamus Egan of Airedale by Modine eloquently puts it, “The liquid transfers heat much, much, much, much more quickly.”
The data center industry overwhelmingly concurs: liquid cooling is the future, especially for AI-centric data centers. Karin Overstreet, president of Nortek Data Center Cooling, states, “As AI has made racks denser and hotter, liquid cooling has become the de facto solution.” However, the journey to implement liquid cooling is not uniform, with a variety of methods emerging, each with its own intricacies and advantages.
Diving Deep into Liquid Cooling Technologies
The innovation in liquid cooling has led to several distinct approaches, ranging from targeted chip cooling to full server immersion. Understanding these methods is crucial for building and operating next-gen data centers.
1. Single-Phase Direct-to-Chip Cooling
This is arguably the most technologically mature approach. It involves placing metal blocks, known as cold plates, directly on top of the hottest components—primarily GPUs and some CPUs—within a server. These cold plates contain channels through which a coolant, typically a mixture of water and glycol (to prevent bacterial growth, stabilize temperature, and protect against corrosion), circulates. This mixture efficiently whisks heat away from the source.
The glycol-water mixture operates in a closed loop, circulating from the cold plates to a heat-exchange unit, which then cools the liquid back down using a separate loop of “facility water.” This facility water can be cooled by an electrically powered chiller or a more energy-efficient dry cooler (though dry coolers are limited to cooler climates). Companies like Mikros Technologies are at the forefront of single-phase direct-to-chip solutions. This approach often results in a hybrid-cooling solution, with liquid cooling handling about 80 percent of the heat load and existing air cooling managing the remaining 20 percent for less power-intensive components like memory units and power supplies.
2. Two-Phase Direct-to-Chip Cooling
As GPU power densities continue to surge, single-phase water cooling faces limitations. To overcome this, two-phase direct-to-chip cooling leverages the physics of latent heat—the energy absorbed or released during a phase change. In this method, a specially formulated dielectric liquid circulates through cold plates atop high-energy chips and boils into vapor, absorbing significant latent heat without a drastic temperature increase. The vapor is then fed back to a heat exchanger, where it condenses back into liquid, cooled by facility water.
My Truong, CTO of ZutaCore, describes it as “boiling to cool.” Unlike water, which boils at 100 °C, the specialized dielectric fluids (from suppliers like Honeywell and Chemours) can boil at temperatures as low as 18 °C, adjustable by pressure. This property allows facility water to be kept 6 to 8 degrees warmer than with single-phase systems, leading to significant energy savings and often eliminating the need for chillers, especially in cooler climates. Furthermore, these dielectric fluids are non-conductive, mitigating damage risk in case of a spill. Companies like Accelsius and ZutaCore are leaders in this space, noting that two-phase systems require significantly lower liquid flow rates—about one-fifth of single-phase—reducing pumping energy and equipment wear.
When liquid boils on top of a hot chip, the chip is cooled not only through contact with the cooler liquid, but also through the latent heat it takes to induce a phase change. Source: Accelsius
3. Single-Phase Immersion Cooling
This method takes liquid cooling a step further by submerging entire computer servers directly into a dielectric fluid, typically an oil. By bypassing cold plates altogether, all components are uniformly cooled. Data centers adopting this approach are designed around immersion tanks rather than traditional racks, with each tank roughly the size of a refrigerator.
Rachel Bielstein of Baltimore Aircoil Co. emphasizes that the fluid must be non-conductive, possess strong thermal transfer properties, and exhibit long-term stability with low environmental and fire risks. After the oil absorbs heat, various methods are used to cool it, such as circulating facility water through heat exchangers within the tank and then to an outside cooler. This process can yield significant energy savings, with Baltimore Aircoil claiming up to 51 percent less energy consumption compared to traditional designs.
Companies like Sustainable Metal Cloud (SMC) specialize in modifying servers for immersion cooling, including removing fans and swapping out thermal-interface materials that might degrade in the oil. Oliver Curtis, co-CEO of SMC, highlights the benefits: “There’s no dust, no movement, no vibration, because there’s no fans. And it’s a perfect operating temperature.” However, for the highest power densities, some chips might still require additional cold plates to enhance fluid flow, as pointed out by Seamus Egan of Airedale by Modine, potentially introducing the complexity of two separate cooling loops.
4. Two-Phase Immersion Cooling
Considered by some as the “moon-shot technology” of data center liquid cooling, two-phase immersion combines the benefits of full immersion with the enhanced heat transfer of phase change. Brandon Marshall, global marketing manager of data-center liquid cooling at Chemours, believes this is where the industry is headed, arguing that boiling liquids offer 10 to 100 times the cooling capacity of single-phase liquids due to latent heat.
In this system, high-power servers are submerged in tanks filled with a proprietary, non-conductive, and non-corrosive dielectric fluid designed to boil at the precise operating temperature of the chips. The heat from the components causes the fluid to boil, and the resulting vapor rises and condenses on a cooled surface (either at the top or back of the tank), returning to liquid form to repeat the cycle. This condenser is cooled by circulating facility water. Marshall notes that they only need water about 6 degrees lower than the boiling point (around 43 °C), often eliminating the need for chillers entirely, which simplifies mechanical infrastructure and reduces costs.
The Chemours team demonstrates their two-phase immersion cooling fluid in action, where the entire server is dunked into a tank, and heat causes the liquid to boil, providing efficient cooling. Source: Chemours
While critics raise concerns about the expense and evaporation of specialized fluids, companies like Airedale by Modine are developing solutions, such as their EdgeBox tanks, designed to minimize vapor loss during maintenance periods by maintaining the vapor layer lower in the tank. A case study by Chemours researchers suggests two-phase immersion can be more cost-effective over 10 years due to lower power and simplified mechanical systems.
Industry Landscape and Adaptable Solutions
The imperative for liquid cooling has spurred innovation across the data center industry, leading to a variety of practical solutions designed for diverse deployment scenarios.
Aligned Data Centers, for instance, has launched its DeltaFlow™ liquid cooling technology, a universal, patent-pending platform. This solution supports high-density compute requirements for next-generation applications, including AI, machine learning, and HPC. DeltaFlow™ extends Aligned’s ExpandOnDemand™ capabilities, offering customers the flexibility to scale seamlessly. Their DeltaFlow™ and Delta 3™ cooling technologies simplify the transition from air-cooled to liquid-cooled systems, or allow for hybrid deployments within the same data hall, eliminating the need for new AI-dedicated build-to-suit data centers or complete facility retrofits. Aligned Data Centers is capable of supporting densities ranging from three to 300 kW per rack and beyond, while maintaining a commitment to sustainability with waterless, closed-loop designs. This turnkey solution integrates with various liquid cooling technologies, including direct-to-chip, rear-door heat exchangers, and immersion cooling, ensuring compatibility with current and future advancements, as detailed in their press release.
Similarly, Fourier Cooling Solutions has entered the market with scalable prefabricated data centers tailored for the AI and HPC era. Their containerized modules are engineered to meet Tier III+ reliability standards, providing redundant power and cooling paths. These solutions significantly reduce construction cycles from years to months and can scale cooling capacity per rack from 40 kW to over 200 kW. Fourier leverages proven immersion and hydro cooling technologies, deployed globally, to ensure infrastructure aligns with rapid semiconductor innovation. Their commitment is to enable AI and HPC data center infrastructure to scale sustainably, efficiently, and globally, according to their PR Newswire announcement.
Even entry-level liquid cooling options are becoming more accessible. Rear Door Heat Exchangers (RDHX) offer a compelling way to introduce liquid cooling architecture without overhauling an entire data hall. These systems, taking up no additional floor space, are fitted with configurations using refrigerant-based, chilled water, or glycol mediums. RDHX solutions can remove 70-75% of the heat generated by equipment, with the remaining 25-30% managed by air cooling, making them an excellent gateway for organizations looking to scale into higher densities.
The market reflects this growing demand. According to the Dell’Oro Group, the liquid cooling market revenue is projected to approach $2 billion by 2027, with a 60% compound annual growth rate (CAGR) for the years 2020 to 2027, driven by cloud services and AI adoption.
Challenges and the Path Forward
Despite its myriad benefits, liquid cooling is not without its challenges. The primary concern often cited is the risk of leaks or other failures that could potentially damage critical hardware. However, with careful design, robust safety protocols, and well-thought-out implementation, these risks can be minimized, ensuring reliability and maximizing the benefits.
The adoption of these sophisticated liquid cooling technologies fundamentally alters the blueprint for data center builds. It demands a new level of collaboration and specialized expertise from all stakeholders across design, construction, and operation phases. Designers must pivot from HVAC-centric models to intricate architectures integrating complex piping, fluid containment, and advanced leak detection. Construction professionals are tasked with meticulous execution, requiring exact pipe fitting, hermetic seals, and seamless integration with power, networking, and security infrastructure. Specialized trades and highly skilled technicians are becoming indispensable, demanding rigorous quality control and adherence to stringent safety standards.
As Brandon Marshall of Chemours asserts, “Unless the floor falls out from under AI and everybody stops building these AI clusters, and stops building the hardware to perform training for large language models, we’re going to need to keep advancing cooling, and we’re going to need to solve the heat problem.” The rapidly evolving nature of data centers ensures that the need for innovation will continue unabated.
Which specific cooling technology will ultimately dominate in the coming AI factories remains an open question. Each method presents unique advantages, and hybrid approaches are becoming increasingly common. What is clear, however, is that building the liquid-cooled data centers of tomorrow is a collective endeavor, requiring integrated strategies and cutting-edge approaches to enable the continued advancement of AI and pave the way for more sustainable and resilient digital infrastructure. As Drew Matter of Mikros Technologies aptly summarizes, “There’s not only a great market for liquid cooling, but it’s also a fun engineering problem.”
