Executive Summary

• Almost all the electrical energy used by an AI chip eventually becomes heat. A 2 kW processor is effectively a 2 kW heater that happens to perform computations.
• As chip and rack power densities rise, air becomes increasingly difficult to use as the main heat-transfer medium.
• Liquid transports heat far more effectively than air. That is why direct-to-chip liquid cooling is becoming standard in high-density AI systems.
• The coolant circulating inside a server does not automatically get consumed. Modern direct-to-chip systems can keep it in a closed loop.
• Much of the water consumption happens at the final heat-rejection stage, especially when cooling towers evaporate water to release heat into the atmosphere.
• This creates a basic trade-off: evaporative cooling can reduce electricity use, but it consumes water.
• That trade-off is beginning to change. Warm-water liquid-cooling systems can use dry coolers to release heat, reducing cooling-water consumption dramatically—and, in suitable conditions, potentially eliminating it. NVIDIA's 2026 Rubin architecture is one example of this direction.

AI chips are getting more powerful—and producing more heat than ever. The surprising part is that the water used in many data centers does not need to touch the chip. Instead, water is often consumed at the final stage, where it helps transfer large amounts of heat into the atmosphere efficiently.

The rapid growth of AI has created a thermal-engineering challenge that is easy to underestimate. A modern accelerator can release around 1,000 watts of heat, and the latest high-performance systems are moving toward chips that dissipate several kilowatts. NVIDIA has discussed processors exceeding 2 kW and AI racks drawing more than 1 MW of power.

Modern AI accelerators already produce substantial amounts of heat. NVIDIA's H100 and H200 GPUs have a maximum TDP of 700 W, while the Blackwell B200 reaches 1,000 W, GB200 GPUs reach 1.2 kW, and GB300 GPUs reach 1.4 kW. At rack level, the figures become even more striking. A GB300 NVL72 contains 72 GPUs, giving the GPUs alone a theoretical thermal load of roughly 101 kW; NVIDIA specifies approximately 135 kW for the complete rack. A modern AI rack is, in effect, an industrial heater that also performs computation. The table below compares the heat loads of selected NVIDIA products in individual-GPU and rack configurations.

Every watt an AI processor consumes eventually becomes heat. From a cooling perspective, a 1 MW AI rack therefore behaves much like a 1 MW heater running continuously. As computing power becomes more concentrated, removing that heat fast enough to keep the electronics within safe operating temperatures becomes increasingly difficult. How do you remove a megawatt of heat from a cabinet without overheating its components? Increasingly, the answer involves liquid—and, depending on the facility design, that can also mean significant water consumption.

1. First Principles: Where Does the Heat Come From?

The principle is straightforward: almost all the electrical energy consumed by an AI processor eventually becomes heat. The energy is temporarily used to perform calculations, move data and maintain electrical states, but computation does not create a significant long-term energy output. For thermal-design purposes, a 1 MW AI system therefore behaves approximately like a continuous 1 MW heater. This does not mean the chip is simply wasting energy; heat is the final physical form of the energy required to compute.

A processor consuming 1,000 W produces approximately 1,000 W of thermal power. At 2,000 W of electrical consumption, it produces roughly 2,000 W of heat.

Direct-on-chip liquid-cooling designs are already targeting thermal loads of around 1,000 W in compact AI-chip systems. Meanwhile, high-power AI applications can reach heat fluxes of about 200 W/cm². NVIDIA has discussed chips above 2 kW and future racks exceeding 1 MW. The challenge is not just the total energy involved; it is the fact that this energy is concentrated in very small physical areas.

2. Why Is AI Creating a New Cooling Problem?

Traditional servers may use CPUs with relatively modest power demands. AI servers, by contrast, can contain several high-power GPUs or accelerators connected by extremely high-bandwidth networks. Their power density continues to rise. NVIDIA notes that hyperscale data centers have moved from historical rack power levels of around 20 kW to more than 135 kW for today's high-density systems, with the next generation expected to go considerably further. A rack can resemble a large electric heater packed into a refrigerator-sized cabinet—but unlike a heater, its components cannot simply be allowed to get hot. Semiconductor devices must remain within tightly controlled operating-temperature limits, so the heat has to be removed continuously.

A rough normalization suggests that a typical frontier-model response may require around 0.1–0.3 Wh per kilobyte of generated text. This estimate varies considerably with the model, prompt length and reasoning workload, so it should not be treated as a fixed physical metric.

One AI response does not necessarily use a huge amount of electricity. Google's measured median Gemini text prompt, for example, is 0.24 Wh. The bigger challenge is scale: AI services handle billions of requests, and models are increasingly used for long-context tasks and complex reasoning. The same trend appears in infrastructure. Representative high-density racks have grown from roughly 20 kW to more than 100 kW, and NVIDIA specifies 135 kW for its GB300 NVL72 rack. AI infrastructure is not merely adding more computers; it is concentrating more computation—and therefore more heat—into less space. Cooling has consequently become a core part of AI-system design. A 1 MW AI facility behaves thermally like a 1 MW heater, while an 800 MW campus would need to reject roughly 800 MW of continuous heat at full IT load. At that scale, the challenge is industrial heat rejection that can operate reliably and economically around the clock.

Scale adds another dimension. ABI Research forecasts around 8,821 operational data centers worldwide by the end of 2026, while other current databases count more than 10,000, with some estimates reaching 11,000–12,000.

The International Energy Agency estimates that global data-center electricity use was about 485 TWh in 2025 and could reach roughly 950 TWh by 2030 in its base case—almost double in five years. Electricity demand from AI-focused data centers is expected to grow even faster, roughly tripling over the same period. The IEA projects that data centers will account for around 3% of global electricity consumption by 2030. The global share may seem modest, but demand is unevenly distributed: in some countries and regions, data centers are already a major source of additional electricity demand.

It remains difficult to measure global AI water use precisely because transparent, standardized historical data for AI specifically are limited. Even so, available studies point to substantial growth. Global data-center water consumption was estimated at about 0.56 billion m³ in 2023, while AI-specific Scope 1+2 consumption is projected at 0.38–0.60 billion m³ in 2027. The IEA projects total global data-center water consumption of around 1.2 billion m³ by 2030. Scope 1+2 covers water used directly by AI and data-center operations, plus the indirect water footprint of the electricity they consume. It does not include wider upstream impacts, such as semiconductor and equipment manufacturing.

The available evidence is enough to show the scale and direction of the issue, but the industry still lacks transparent, standardized reporting of AI-specific water use. That gap is important in itself. A 2026 paper in Water Research estimates that AI's global water footprint could reach 4.2–6.6 billion m³ of water withdrawal by 2027, while Scope 1+2 consumption is estimated at 0.38–0.60 billion m³. The paper also highlights how little facility-level water-use data are disclosed.

One distinction is essential:

Water withdrawal is not the same as water consumption.

The 4.2–6.6 billion m³ figure refers to water withdrawn, while the 0.38–0.60 billion m³ estimate refers to water consumed. Much of the consumed water evaporates and is therefore not immediately returned to the local water system.

3. Why Not Just Use Air?

This is where the physics matters. Air is convenient, inexpensive, does not leak and can be moved with fans. However, it transfers heat less effectively than liquids. As power density rises, more air must be moved across the heat source. Imagine trying to cool a chip producing several kilowatts using room-temperature air. You would need to move far more air or move it much faster, and neither option scales particularly well. With air cooling, heat travels along this path:

Chip → heatsink → air → room → cooling system → outside environment

Every step adds thermal resistance and makes heat transfer more difficult. Liquid cooling shortens the path:

Chip → cold plate → liquid → heat exchanger → environment

The coolant is brought much closer to the heat source. Guidance from the U.S. Department of Energy notes that water is a much more effective heat-transfer medium than air, helping warm-water cooling support higher-density computing. One major reason is water's much greater heat capacity compared with air.

The table is generated for the case of a 10°C coolant temperature rise

Water is not automatically better simply because it is water. Its advantage comes from its much higher volumetric heat capacity and much higher heat-transfer coefficients. These properties allow the cooling system to collect heat closer to the silicon. For the same heat load and coolant temperature rise, the required volumetric flow of water is roughly 3,500 times lower than the required airflow.

The difference becomes especially clear when we look at heat transfer from a small chip. Consider a 1 kW AI processor with an effective heat-transfer area of approximately:

30 mm × 30 mm = 900 mm² = 0.0009 m²

That is roughly the size of a large semiconductor package. Even with an optimistic forced-air heat-transfer coefficient, the temperature difference required to remove 1 kW would be:

Even if we assume a very aggressive value for air:

the result is still:

That is clearly impractical. This is the difficulty of removing kilowatts of heat from a tiny silicon area using air. Now suppose a liquid-cooled cold plate achieves:

The required temperature difference is still substantial, but the result is at least within a physically meaningful range.

If improved microchannel or jet-impingement designs reach:

the chip could potentially be cooled with a much more manageable temperature difference. This is why direct-to-chip liquid cooling matters: the coolant does more than carry heat away; it can also reduce the thermal resistance between the silicon and the coolant.

The energy needed to move the cooling medium also differs substantially between air and water.

In this illustrative example, the power required to move the coolant differs by about 80×. The exact result depends on pressure drop and other efficiency assumptions, but the gap still makes a strong case for water.

4. The Surprising Part: Water Doesn't Necessarily Need to Be "Used Up"

Discussions about AI water use can be misleading if they do not distinguish between circulating coolant and water that is actually consumed. In a modern direct-to-chip system, coolant flows through a cold plate attached directly to the processor:

GPU → cold plate → coolant → heat exchanger → facility cooling loop

The same coolant can circulate continuously; it does not disappear or need to be replaced after each computation. NVIDIA's Blackwell systems, for example, use closed-loop liquid cooling. The coolant captures heat at the chip and carries it through coolant distribution units and heat exchangers.

So where does the water consumption actually occur?

The answer is heat rejection. Cooling towers are often the main source of water consumption. Suppose the IT equipment produces 1 MW of heat. A closed liquid loop can carry that heat away from the GPUs, but the facility still has to release the 1 MW into the environment. One common solution is an evaporative cooling tower, where the process follows this path:

AI chips → liquid cooling loop → heat exchanger → facility water loop → cooling tower → water evaporates → heat enters the atmosphere

This is why cooling towers consume water. They reject heat primarily through evaporation. As water evaporates, it carries heat into the atmosphere. Additional water is lost through blowdown and, to a smaller degree, drift. This gives us a crucial distinction: Water inside the cooling loop can be recirculated. Water used by evaporative heat rejection is consumed. That distinction is essential when discussing AI's water footprint.

5. Why Is Evaporation Such a Powerful Cooling Mechanism?

The answer is thermodynamics. When water evaporates, it takes a large amount of energy with it in the form of latent heat.  A water molecule must absorb a significant amount of energy to escape from the liquid into the vapor phase. This energy is called the latent heat of vaporization.  Evaporative cooling exploits this thermodynamic property of liquids: changing one kilogram of water from liquid to vapor can remove roughly 2.4 megajoules of heat—far more energy than simply warming that kilogram of water. At around room temperature, the latent heat of vaporization of water is roughly: 2.4 MJ per kilogram. That means evaporating only about: 0.42 kg of water can theoretically carry away approximately: 1 MJ of heat. Now scale that up. A system rejecting: 1 MW = 1 MJ/s would theoretically need on the order of: 0.42 kg of water per second if evaporation were the only heat-rejection mechanism. That's approximately: 36,000 liters per day for every megawatt of continuously rejected heat under idealized conditions. Real systems differ because of temperature, humidity, heat-exchanger efficiency, cycles of concentration, partial dry cooling and other factors. But the order of magnitude illustrates why the numbers can become large. This is the fundamental reason water is attractive: Evaporation is an extraordinarily effective way to get rid of heat.

6. Why Not Simply Use Air Outside the Building?

Using outside air is possible, and more data centers are doing exactly that. The trade-off is that hot outdoor air makes heat rejection harder. Evaporative cooling has a thermodynamic advantage: it can cool the heat-rejection stream toward the ambient wet-bulb temperature, rather than only toward the dry-bulb temperature. This can reduce the need for mechanical refrigeration. In other words, water can help reduce electricity consumption.

The trade-off is that cooling which uses little electricity can consume water.

The optimum solution therefore depends heavily on:

The U.S. Department of Energy notes that water and energy use in cooling towers are closely linked, and that the best approach depends on climate and operating conditions.

At 25°C and 50% relative humidity, for example, the outdoor wet-bulb temperature is roughly 18°C—well below the 25°C dry-bulb temperature. An evaporative cooling tower can take advantage of this difference, allowing a water-cooled chiller to reject heat against a colder effective heat sink than a purely air-cooled system.

For a 1 MW cooling load, large air-cooled chillers typically use around 244–270 kW of compressor power. A water-cooled chiller, under representative full-load COP assumptions, uses approximately 189–217 kW. That means the water-cooled system needs roughly 20% less chiller electricity, before adding the power used by pumps and cooling-tower fans.

This is one of the central trade-offs in data-center cooling: air can reduce water consumption, while water-based evaporative cooling can reduce electricity use. New AI cooling designs are trying to avoid that trade-off by combining direct-to-chip liquid cooling with warm-water loops and dry coolers.

7. But AI Cooling Is Changing Again

One of the most interesting recent developments is that the cooling approach itself is changing.

The progression is moving from air cooling to direct-to-chip liquid cooling, then to warm-water liquid cooling, and potentially to fully liquid-cooled systems that reject heat dry.

The key insight is that coolant does not necessarily need to be cold. If a processor can operate safely with relatively warm coolant entering its cold plate, the facility may be able to release heat to the outside air much more efficiently.

So what does warm-water cooling mean in practice? Conventional liquid-cooling systems may supply water at around 10–20°C to maintain a large temperature difference between the processor and the coolant. But if the processor can safely operate with coolant entering at 40–45°C, the facility may not need to chill the water back down to 10–20°C. Instead, it can send the warm liquid outdoors through a large radiator-like unit called a dry cooler. The outdoor air does not need to bring the liquid down to 15°C; it may only need to cool it from, for example, 55°C to 45°C. That is much easier when the outside temperature is 30–35°C. In favorable conditions, this can mean no cooling tower, no evaporative water loss and no large refrigeration plant—and therefore less electricity use.

This is a major step in the evolution of data-center cooling.

NVIDIA's 2026 architecture is designed for coolant entering at temperatures of up to 45°C and returning at around 55°C, with 100% liquid cooling. Because the loop operates at relatively high temperatures, dry coolers can release heat directly to outdoor air for much of the year, potentially eliminating evaporative cooling. The architectural shift is simple but important: rather than using cold water to cool a hot chip, warm liquid carries the heat to an outdoor heat exchanger.

In favorable climates, NVIDIA says this approach can reduce facility cooling-water consumption from roughly 2.6 million gallons per MW per year with conventional cooling towers to almost zero.

The evolution of Chip Cooling

The transition is visible in NVIDIA's Blackwell platform. The GB200 NVL72 is a 120 kW rack-scale system with 72 Blackwell GPUs and 36 Grace CPUs, using direct-to-chip liquid cooling to capture heat at its source. Liquid cooling itself does not necessarily lead to high water consumption: the coolant circulating through the cold plates stays in a closed loop. The facility's water footprint depends largely on how it ultimately rejects the heat.

NVIDIA estimates that conventional cooling-tower systems can consume roughly 2.6 million gallons (≈9,840 m³) of water per MW of data-center capacity each year. Its newer 45°C liquid-cooling architecture can use dry coolers to release heat directly to the outdoor air, bringing facility cooling-water consumption close to zero in favorable climates. Against the conventional benchmark, a 100 MW AI facility could theoretically avoid almost 1 million m³ of cooling-water consumption annually. The significance of liquid cooling therefore extends beyond keeping GPUs within their operating-temperature limits: it can change the data center's entire energy-and-water architecture.

The current direction of travel is clear: air cooling → direct-to-chip liquid cooling → warm 45°C liquid → dry heat rejection → near-zero cooling-water consumption. More powerful AI does not inevitably have to mean more water use. It could mean more heat, better thermal architecture and, potentially, less water consumed per unit of computation.

NVIDIA is not the only company moving in this direction. Microsoft has redesigned its next generation of AI-optimized data centers around direct-to-chip liquid cooling. In August 2024, it introduced a design that circulates coolant through cold plates attached directly to AI processors in a closed loop, avoiding evaporative water cooling during normal operation. Microsoft estimates that this design can save more than 125,000 m³ of water per data center each year compared with its previous design baseline.

The improvement involves more than replacing air with liquid. Microsoft has redesigned the thermal system around the chip so it can operate at higher temperatures, reducing the need for energy-intensive refrigeration. The company is also investigating microfluidic cooling, which brings coolant even closer to the processor.

Microsoft also says that around 90% of its owned data-center fleet in 2025 used low- or zero-water cooling systems, including direct-air and other water-efficient approaches. Its fleet-wide metrics reflect this progress: global Water Usage Effectiveness (WUE) fell from 0.49 L/kWh in 2021 to 0.30 L/kWh in FY2024 and 0.27 L/kWh in FY2025—about a 45% reduction in water intensity since 2021. Microsoft reports that average WUE has improved by nearly 90% compared with its first-generation data centers, when it was approximately 2.3 L/kWh.

MSF new approach to AI cooling already incorporates dry heat rejection and close loop liquid cooling operation.

Zero-water-evaporation cooling is the next step. Microsoft acknowledges that eliminating evaporative cooling can increase electricity use, but says warmer coolant temperatures and high-efficiency economizing chillers keep the overall energy penalty small across its global fleet.

The broader lesson is not just that AI cooling can use less water. Data-center architecture can be redesigned so that computing capacity and water consumption are no longer directly tied to one another.

10. Why Water Consumption Is Becoming a Geographic Problem

Location matters because a liter of water does not have the same environmental impact everywhere. Using water in the Netherlands is not equivalent to using the same amount in Arizona, Spain or the Middle East.

Freshwater availability, rainfall, groundwater stress and competing agricultural or municipal demand all affect the local impact. Google, for example, uses watershed-health assessments to understand how data-center cooling choices interact with local water conditions.

This means future data-center location decisions increasingly involve three resources:

Electricity: How much power is available?

Water: How much cooling water is available?

Heat-rejection conditions: How easily can the environment accept the waste heat?

The best site for an AI data center may therefore not be the one with the cheapest electricity alone. It may be the location offering the best combination of electricity, climate, water availability, grid capacity and cooling architecture.

11. The Next Generation: Cooling the Chip From Inside

Another promising direction is to move cooling even closer to the heat-generating silicon. Direct-to-chip cold plates already bring liquid very close to the processor, but researchers are exploring the next steps: cold plate → package → silicon → microchannels inside or immediately beside the chip.

Recent research suggests that optimized and embedded microchannel designs can reduce thermal resistance and pumping power by tens of percent, and in some cases improve cooling energy efficiency by roughly an order of magnitude. Experimental devices have also demonstrated heat fluxes above 1,000 W/cm² and, in some cases, above 2,000 W/cm².

These approaches could substantially reduce thermal resistance and allow chips to operate at much higher power densities. Research published in 2026 is already showing the potential: some optimized microchannel designs have reduced thermal resistance by roughly 20–30% compared with conventional cold plates, while a separate 1,000 W-class experiment reduced pumping power by 60%. More advanced embedded-manifold microchannels, fabricated directly into a silicon thermal test chip, have demonstrated heat fluxes above 2,000 W/cm² and around ten times the cooling energy efficiency of the previous state of the art.

The long term objective is therefore not simply to make today's cold plate better. It is to eliminate as much of the thermal path between the heat-generating silicon and the coolant as possible. That starts to make thermal management part of semiconductor architecture. And that brings cooling directly into the same convergence: “Semiconductor / photonics / packaging / thermal engineering / AI” is becoming one integrated engineering problem.

12. The Hidden Engineering Problem: Heat Density

Consider two systems that each generate exactly 100 kW of heat. One spreads the load across 10 m² of liquid-cooled surface; the other concentrates it in a small, high-performance AI package. The total heat is the same, but the heat flux can differ by more than two orders of magnitude. The distributed system might operate at around 10 kW/m², while a 100 kW AI system concentrated over 0.065 m² would reach approximately 1,540 W/cm².

That is why next-generation cooling is moving toward microchannels. The goal is not to reduce the total heat that must be rejected—the 100 kW remains 100 kW. Instead, the aim is to reduce thermal resistance between the silicon and the coolant so heat can be removed from a very small area without allowing the silicon to reach unsafe temperatures. The challenge is not only how much heat an AI system produces, but how much heat a few square centimeters of silicon must release.

For this reason, advanced AI infrastructure is increasingly designed around rack-level thermal engineering. Cooling is no longer an afterthought added once the computer is designed. NVIDIA's Blackwell and Rubin platforms illustrate the shift toward system- and rack-level thermal design.

13. Industry Landscape: Who Is Actually Solving the AI Cooling Problem?

The AI cooling market is becoming increasingly diverse. No single technology suits every application. The competitive landscape now includes companies working on cold plates, coolant distribution, two-phase systems, immersion cooling, heat rejection and complete data-center architectures.

The challenge is not simply to remove heat from a GPU. The full thermal path runs through:

Silicon → package → cold plate → coolant → CDU → facility loop → heat exchanger → atmosphere

Different companies are trying to control different parts of this chain, and liquid cooling has become a substantial market in its own right. One 2026 estimate values the global data-center liquid-cooling market at approximately $4.8 billion in 2025, with a projected CAGR of 18.2% through 2035. Another estimate focused on direct-to-chip cooling puts the market at $3.33 billion in 2026, rising to $17.31 billion by 2032. The figures differ because the studies define the markets differently, but both point to rapid growth.

The competitive landscape can be grouped into four broad categories.

Tier 1 — Infrastructure integrators: These companies aim to manage the overall thermal system, not just supply a cold plate.

Vertiv is a clear example. Its CoolChip platform covers direct-to-chip cooling, rear-door heat exchangers, coolant distribution and heat rejection, treating them as parts of one system rather than isolated products. Its CDU portfolio ranges from approximately 100 kW to 2.3 MW, with both liquid-to-liquid and liquid-to-air architectures.

That range matters. A 100 kW CDU can support a high-density rack, while a 2.3 MW unit can serve a larger cluster or thermal zone. Vertiv is positioning itself as a complete thermal-infrastructure partner rather than simply a hardware supplier. It is also addressing the practical challenge of retrofitting existing facilities. Its CoolPhase CDU uses pumped-refrigerant economization to bring liquid cooling to sites without conventional chilled-water infrastructure, allowing air- and liquid-cooled racks to operate together.

What sets Vertiv apart is system integration and deployment flexibility. Its advantage is less about a single breakthrough cooling mechanism and more about integrating the complete thermal chain into working data centers.

Schneider Electric approaches the issue from a broader infrastructure perspective, spanning:

power → rack → cooling → controls → data-center management

This matters because cooling is becoming inseparable from the power architecture. For a 100–200 kW rack, operators need to know not only whether the GPUs can be cooled, but also:

Companies with an established data-center infrastructure business are well placed as cooling becomes a facility-level engineering challenge.

Specialist companies focus more directly on the thermal bottleneck itself.

CoolIT Systems is one of the most prominent specialist suppliers in direct-to-chip liquid cooling. Its work goes beyond cold plates to include:

cold plates → manifolds → cold-plate loops → CDUs → complete thermal-management systems

The company says it is currently supplying multi-gigawatts of CDU capacity to AI customers. Its technology roadmap is also moving quickly.

In June 2026, CoolIT announced a 15 kW cold plate—nearly 4× the cooling capacity of its previous demonstrated design—to address future ultra-high-density GPUs and AI accelerators.

CoolIT says its systems can cool AI chips above 4 kW and server racks above 500 kW. This illustrates the progression from:

700-W GPU

→ 1–1.5 kW GPU

→ 4-kW+ future device

→ 100-kW+ rack

→ 500-kW+ future rack

What is technically different?

CoolIT's strategy is to extend conventional single-phase direct liquid cooling as far as possible before moving to a fundamentally different architecture. That could be a relatively low-risk route from today's data centers to tomorrow's AI factories.

JetCool takes a different approach: it targets hotspots instead of cooling the entire surface uniformly. Its Microconvective Cooling technology uses arrays of tiny fluid jets to cool specific hot regions rather than simply flowing liquid across a cold plate.

JetCool says its SmartPlate cold plates can support devices above 3 kW TDP and offer approximately 3× lower thermal resistance than leading microchannel cold plates. Its SmartLid approach brings cooling closer to the silicon by directing liquid toward the processing die and removing conventional thermal-interface layers.

JetCool is also developing SmartSilicon, which embeds cooling structures in the semiconductor substrate itself. The company says this approach can support more than 4 kW in a single socket and isolate individual chiplets in heterogeneous 2.5D/3D packages. The progression is:

cold plate → liquid-to-die → embedded cooling

The key difference is localized cooling. Rather than increasing coolant flow everywhere, JetCool aims to direct cooling capacity precisely where heat is generated.

ZutaCore is pursuing a more fundamental change through two-phase cooling. Its HyperCool system uses a dielectric fluid that boils directly at the processor. Instead of simply warming up as it carries heat away, the liquid changes phase into vapor. That phase change absorbs substantial latent heat and can transport large amounts of heat without the same flow rates required by conventional single-phase cooling.

ZutaCore says its system can cool processors at 1,500 W and above, and reports power reductions of 35% or more under its comparison methodology. The approach is notable because it changes the thermodynamic mechanism rather than just improving the pump, cold plate or heat exchanger.

The technical distinction is two-phase pool boiling directly at the chip.

The technology potentially provides:

The trade-off is a more complex system, with the two-phase fluid needing to be managed safely and reliably.

Accelsius follows a similar two-phase approach, with a strong focus on high-density rack deployment. Its NeuCool system uses a dielectric refrigerant that boils at the chip and condenses elsewhere in a closed system. Accelsius claims cooling capacity above 4.5 kW per socket, with thermal resistance of approximately 0.020°C/W at 700 W+ TDP.

Its rack products are already moving toward very high density:

In July 2026, Accelsius published third-party testing claiming that its two-phase design maintained performance comparable to a single-phase system using 45°C facility water, while operating with a 54°C inlet. If confirmed in deployment, this could be useful for warm climates and free-cooling systems. The key technical difference is:

Phase change provides more temperature headroom and reduces reliance on very cold facility water. That is especially attractive when paired with dry heat rejection.

Startups and challengers are particularly worth watching from a technology perspective. Established suppliers tend to refine architectures that can be deployed today; startups are more likely to question whether the architecture itself should change.

Examples:

Companies such as Submer, LiquidStack and GRC are pursuing immersion cooling, in which the server or electronics assembly is immersed in dielectric fluid instead of relying on cold plates. This can be attractive at very high power densities, although direct-to-chip cooling currently has a clearer route into mainstream AI server designs. The 2026 supplier landscape includes both approaches, each with different deployment and operational trade-offs.

Research organizations may provide the next major architectural shift. Commercial suppliers generally focus on systems that can be manufactured and deployed today. Researchers are exploring a more fundamental question: how close can coolant be brought to the heat-generating transistor? This is where microchannel research becomes important. Recent academic work has demonstrated:

>1,000 W/cm² and in one 2026 embedded-manifold architecture: >2,000 W/cm² at the experimental device level.

These results do not yet represent a commercial AI GPU, but they show the potential available if manufacturing and reliability challenges can be solved. Key research directions include:

·        Embedded microchannels

·        Channels directly in or immediately adjacent to silicon.

·        Jet impingement

·        Tiny liquid jets aimed directly at hotspots.

·        Microfluidics

·        Very small channels optimized for extremely high heat flux.

·        Two-phase cooling

·        Using boiling to exploit latent heat directly at the source.

·        Advanced thermal interfaces

·        Reducing the thermal resistance between die and cooling structure.

Who is doing something genuinely different?

Taken together, the 2026 market includes direct-to-chip, two-phase and immersion cooling, along with CDUs and complete infrastructure platforms. The supplier ecosystem is spread across all of these approaches.

The eventual winner will not necessarily be the technology with the highest heat-transfer coefficient. It will need to perform well across a much broader set of requirements:

Performance+Reliability+Manufacturibility+Cost+Servicabiltiy+Scalability

A research microchannel cooler capable of removing 2,000 W/cm² is impressive. But a cold plate that removes 1,000 W/cm², can be manufactured millions of times, does not leak, lasts five years and costs a few hundred dollars may have greater commercial value. The market can therefore be viewed through four competing approaches:

Evolutionary — CoolIT / Vertiv: Extend today's direct-liquid architecture to handle tomorrow's chips.

Architectural — JetCool: Move the coolant closer to the actual heat hotspots.

Thermodynamic — ZutaCore / Accelsius: Change the cooling mechanism through phase change.

Extreme integration — Research organizations: Put cooling structures directly into the semiconductor package or silicon.

The industry is making rapid progress on cold plates, microchannels and immersion cooling, but a harder question remains: how do we reject hundreds of megawatts of heat after collecting it? Evaporative and dry cooling are not the only options, although several alternatives are not yet mature enough for widespread industrial use. The table below compares the main approaches.

Cooling innovation is not limited to the equipment outside a processor. Semiconductor designers are also trying to reduce the heat that needs to be removed in the first place. Gate-all-around transistors can improve performance per watt, while backside power delivery reduces electrical losses and frees the front of the chip for denser signal routing. At system level, chiplets and 3D integration shorten data paths and may reduce one of AI computing's major energy costs: moving data.

However, 3D integration introduces a trade-off. The same architecture that reduces data-movement energy can make cooling harder. A die buried inside a three-dimensional stack has no direct thermal path to the cooler. Compute density, bandwidth, electrical efficiency and access to cooling must therefore be optimized together.

The future AI chip may be defined by more than its transistor count. Its design will increasingly depend on where heat is generated, how efficiently data moves, where power enters the chip and how quickly heat can escape.

Two architectural changes are particularly relevant to improving chip energy efficiency:

1.     FinFET → Gate-All-Around (GAA): Modern transistor design is moving from FinFETs toward gate-all-around structures. Intel's RibbonFET is one example. In a FinFET, the gate controls the channel from three sides; in a GAA design, the gate surrounds the ribbon-shaped semiconductor channel.

Better electrostatic control can allow lower operating voltages and reduce leakage. Intel reports that its 18A process with RibbonFET provides up to 38% lower power at equivalent performance than Intel 3. This is a process-level comparison, not a direct measurement of cooling-system performance.

For cooling, the implication is simple: the same computation can require less electrical power and therefore produce less heat.

2.     Backside power delivery: This is another important development. Traditionally, signal wiring and power-delivery wiring have both been routed through the same side of the chip.

As transistors become smaller and currents become larger, this creates congestion and electrical resistance. Intel's PowerVia moves much of the power delivery to the backside of the silicon.

Separating power and signal routing can reduce resistance and voltage losses. Intel reports that PowerVia can reduce worst-case dynamic voltage droop by as much as 10× and deliver up to 11% block-level area compaction in routed designs. Intel also says its 18A-P technology improves the thermal resistivity of the overall stack by 20–40% compared with the previous platform. This matters because electrical resistance generates heat: reducing resistance can prevent some electrical energy from becoming heat in the first place. Backside power delivery therefore addresses part of the thermal problem before the heat is generated.

Different approaches to make the chip architecture more energy and thermal efficient.

16. The Photonics Intel Verdict

Thermal engineering is becoming one of the fundamental limits on AI scaling. For decades, computing performance was constrained mainly by transistor density, then power, then data movement—and increasingly, by the ability to remove heat.

This is creating a new competitive field focused on:

The main contest in AI cooling may no longer be air versus liquid; at high power densities, that debate is increasingly settled. The next question is which liquid-cooling architecture can scale economically to 1, 5, 10 and eventually 20+ kW per processor. At the same time, the industry must move from 100 kW racks toward racks rated in hundreds of kilowatts, while reducing water consumption and avoiding a proportional rise in cooling electricity.

Four technological paths are emerging:

·        Higher-performance cold plates

·        microconvective / microchannel cooling

·        two-phase cooling

·        embedded / in-silicon cooling

Companies that can turn exceptional laboratory performance into reliable mass production could become major players in AI infrastructure. The strategic lesson is worth noting:

The eventual winner may not be the company with the most impressive cooling technology. It may be the one that makes cooling almost invisible—integrating it so thoroughly into the chip, rack and data center that thermal management becomes part of the computing architecture itself.