The world’s data centers are becoming giant furnaces—and researchers now propose using that heat to make more cooling. A study led by engineers Jinkyun Cho and Joo Hyun Moon describes a “cooling-from-cooling” system that captures heat from liquid-immersed computer hardware and feeds it into an adsorption heat pump. The device then produces chilled water for cooling other equipment, creating a thermal recycling loop designed to reduce dependence on electrically driven cooling machinery. In a prototype-scale design, 50 kilowatts of heat recovered from a 200-kilowatt immersion-cooling system would generate approximately 20 kilowatts of additional cooling. The concept could offer data-center operators a way to expand computing capacity while making use of heat that would otherwise be rejected to the atmosphere.
The proposal arrives as artificial intelligence, cloud computing and connected devices drive an extraordinary expansion in computing infrastructure. Data centers consumed an estimated 240–340 terawatt-hours of electricity in 2022, equivalent to roughly 1–1.3 percent of global electricity use, according to the International Energy Agency. That demand is expected to more than double by 2030. Although electricity powers the processors, memory and networking equipment, nearly all of the energy consumed by information technology ultimately becomes heat. Keeping that heat away from sensitive electronics can account for about 30–40 percent of a facility’s energy use under some operating conditions. Conventional systems move heat into chilled water or exhaust air, then reject it through cooling towers or dry coolers. Much of the thermal energy is therefore lost despite being continuously available.
The challenge is not simply the amount of heat, but its temperature and thermodynamic quality. Traditional air-cooled servers generally release heat at only about 35–45 °C, while condenser-water loops may return at around 37 °C. That is abundant energy, but it is “low grade”: its temperature is too close to the surrounding environment to drive many useful heat engines or industrial processes. Warm-water liquid cooling can raise the outlet temperature to roughly 50–65 °C, and two-phase systems may reach 70–90 °C, but these more advanced technologies introduce engineering concerns involving leakage, vapor distribution, reliability and long-term stability. The researchers identify an intermediate range of approximately 45–60 °C as the most practical for their approach. It is hot enough to regenerate an adsorption cycle, yet compatible with liquid cooling systems that can keep high-power processors within safe operating limits.
An adsorption heat pump does not use a mechanically compressed refrigerant in the way a conventional electric chiller does. Instead, it relies on the ability of a porous solid to attract and release a refrigerant, in this case water vapor, as its temperature changes. The proposed machine contains two beds of adsorbent material, a condenser and an evaporator. While one bed is heated, or regenerated, hot water drives water vapor out of the material. That vapor condenses, and the liquid refrigerant then evaporates at low pressure, absorbing heat from a chilled-water loop. The evaporator thereby produces cooling. At the same time, the second bed adsorbs vapor and releases heat to cooling water. The two beds periodically switch roles, allowing the cycle to continue. Because the regeneration energy comes from data-center waste heat, the system can produce cooling without adding a compressor or requiring an electrically powered heat-up stage.
To test whether the cycle could function with data-center heat, the researchers modeled two adsorbent–water pairs: silica gel–water and MIL-101(Cr)–water. Their calculations used a Linear Driving Force model, a standard approximation for transient adsorption and desorption. In that model, the rate at which refrigerant is taken up or released depends on the difference between the material’s instantaneous uptake and its equilibrium uptake. The model also included the heat released during adsorption, the heat required for desorption, the thermal masses of the bed and adsorbent, and the temperatures and flow rates of the working fluids. The coupled equations were solved with adaptive numerical time stepping because the cycle involves rapidly changing, nonlinear interactions between heat and mass transfer. A cycle time of about 200 seconds produced stable behavior in the simulation.
The modeled system used hot water entering the adsorption beds at 53 °C and cooling water entering at 32 °C. Under these conditions, the silica-gel system delivered a coefficient of performance, or COP, of about 0.77 in the reported calculation. COP is the ratio of cooling delivered to thermal energy supplied; it is not directly comparable to the COP of an electric compressor because the energy input is primarily heat rather than electricity. The model showed that raising the regeneration temperature could actually reduce performance: silica gel’s COP fell to 0.27 at 63 °C and 0.17 at 73 °C. The researchers attribute this decline to greater refrigerant uptake and larger enthalpy flows during short cycles, increasing the thermal input required for each unit of cooling. The MIL-101(Cr)–water pair produced COP values of approximately 0.25–0.40 and showed stronger adsorption potential at lower cold-bed temperatures. These results suggest that choosing the adsorbent and operating temperature together will be crucial.
The proposed engineering configuration combines six immersion tanks, two coolant distribution units, an adsorption heat pump and a rear-door heat exchanger. The immersion system is designed for a 200-kilowatt information-technology load, with each tank handling about 34 kilowatts. One 50-kilowatt section of the cooling loop sends hot coolant to the adsorption heat pump, while a separate 150-kilowatt section continues to the cooling tower. The recovered heat enters the heat pump at approximately 53 °C and leaves at about 48 °C. The resulting 20-kilowatt chilled-water output leaves the evaporator at 20 °C and is sent to a rear-door heat exchanger serving an additional 18-kilowatt air-cooled load. Pumps, fans and heat exchangers were sized with allowances for flow resistance and piping losses. The design therefore does not claim that all data-center heat can be converted into cooling; rather, it diverts a carefully selected fraction into a thermal cycle while the remaining heat follows the conventional rejection path.
The energy analysis produced a more complicated result than a simple claim of universal electricity savings. Compared with immersion cooling alone, the internal-reuse configuration required additional pumps, a heat exchanger and a rear-door fan. At full design load, the baseline system had a mechanical load component of 0.051 and a cooling-focused power usage effectiveness, or PUE, of 1.051. With heat recovery used internally, those values increased to 0.065 and 1.065 because the extra equipment consumed mechanical energy. However, the same configuration enabled the facility to cool an additional 18-kilowatt air-cooled information-technology load, raising the total modeled IT load from 200 to 218 kilowatts without adding compressor power. In a second scenario, where recovered heat was credited against an external cooling demand, the calculated mechanical load component fell to 0.031 and PUEcooling to 1.031. The distinction matters: recycling heat within the facility can increase cooling capacity even when it does not reduce the cooling system’s own electricity use, while exporting the recovered thermal benefit can improve the accounting efficiency more directly.
The study is a theoretical and prototype-scale demonstration rather than proof of commercial performance. Its calculations assume continuous operation at full rated load for 8,760 hours per year, whereas real computing demand fluctuates and cooling towers perform differently across climates and seasons. Adsorption machines also tend to have lower COPs and slower dynamic responses than mechanical chillers, and maintaining a stable 200-second switching cycle under changing server loads may be difficult. The researchers say a pilot system is being developed for experimental validation, including tests of operating stability, thermal matching and real-world energy consumption. If those tests confirm the model, the approach could be especially valuable in high-density computing facilities where cooling capacity—not processor availability—is becoming the main limit on expansion. Rather than treating waste heat as an unwanted by-product, the system turns it into a second chance to remove heat, closing a loop that could help data centers handle the escalating thermal burden of artificial intelligence.
