The rapid expansion of AI is creating a growing challenge for data centre operators. Training and running large AI models forces computer chips to work at maximum capacity continuously, generating massive amounts of heat that require intense cooling systems.
Cooling can consume enormous amounts of water, particularly where facilities rely on evaporative cooling – a technique which involves spraying water onto a heat exchange. This puts data centre expansion into competition with other demands for increasingly constrained water supplies.
The issue is becoming particularly acute in the UK, where trade body Water UK has warned that the country’s water planning does not adequately account for future data centre demand.
Against this backdrop, CO2 could offer an alternative route to cooling and managing the heat generated by data centres.
And it is not limited to the UK. One example is a recently announced project backed by the US Department of Energy (DOE) to develop a supercritical carbon dioxide (sCO2)-based system that could provide both power and cooling for data centres and military installations.
The Strategic Partnership Project will be developed by advanced nuclear company Elemental Nuclear Energy (ENE) and Sandia National Laboratories.
According to the partners, the closed-loop sCO2 Brayton cycle generators will initially target a one-megawatt electric system that can use natural gas and waste heat to generate power and cooling.
ENE then plans to develop a scale-up 10MW unit engineered to operate with a range of heat sources, including the Elemental ISTR nuclear reactor.
Built to train xAI’s Grok chatbot, the Colossus data centre consumes estimated 350 million gallons of blue water per year ©Steve Jones Flight by Southwings for the Southern Environmental Law Center
“We intend to prove that sCO2 power generation can be delivered reliably and at commercial scale, as an available source of on-site power to data centres, industrial operators, and critical installations,” said David Blythe, CEO of ENE.
The larger system will focus on behind-the-metre power applications for data centres, microgrids, and industrial customers.
Named after 19th-century engineer George Brayton, the Brayton Cycle uses heated sCO2 instead of steam to generate electricity.
In a closed-loop Brayton cycle, the sCO2 is heated before the energy is extracted from the CO2 in a turbine. The CO2 then exits the turbine and is cooled before being compressed and returned to the heater to continue the cycle.
SCO2 can operate at high temperatures and pressures, allowing power-generation equipment to be more compact than conventional steam-based systems.
And because the system is closed-loop, the fluid stays sealed inside the pipes and heat exchangers. It is not consumed during operation, so industrial supply shortages of commercial CO2 do not restrict its use.
ENE says the first one-megawatt system is expected to be operational in 2027, with commercial deliveries targeted from 2028.
Data centres’ growing water problem
The DOE project illustrates one potential way in which CO2 could become part of the infrastructure supporting the next generation of data centres. But the underlying challenge is broader than power generation.
Data centres are already facing growing scrutiny over the amount of water required to keep computing equipment cool, with the issue becoming more acute as AI increases the power density of servers.
Water UK has warned that the country’s water forecasts are “fatally flawed” because they “explicitly exclude” data centres.
“There appears to be an assumption that the country will always have enough water for its economic needs,” the trade body said in a written briefing to MPs. “Nothing could be further from the truth.”
The concern comes as data centre development accelerates in areas where water resources are already under pressure.
According to figures reported by The Guardian, UK data centres currently consume around 6.6 million litres of drinking water per day. That could rise to around 19.8 million litres per day if data centre capacity triples by 2030, while some proposed facilities are seeking supplies of up to three million litres per day.
The problem is not simply the amount of water used inside a data centre. Operators are increasingly having to balance water consumption against electricity consumption when deciding how to remove heat from their facilities.
And although many UK data centres use closed-loop cooling, where water is reused, many facilities still rely on secondary cooling towers during warmer periods, where water is sprayed or evaporated to reject heat.
The daily average also surges during intense summer heatwaves, when water usage in data centres more than doubles from baseline averages.
That is driving interest in cooling systems that can reduce or eliminate the need for evaporative water use.
Where CO2 could come in
CO2 is already being explored as a working fluid for data centre cooling, rather than simply as a substance to be captured or emitted.
US-based Infinity Turbine, for example, has explored the use of sCO2 to capture waste heat from data centre GPUs and use the pressure drop generated as the CO2 expands to provide additional cooling.
The company has also proposed a pulsed sCO2 heat pump concept for data centre cooling, which it says could reduce energy consumption by up to 50% compared with conventional air- or water-cooling methods.
US thermal technology startup Karman Industries has developed a CO2-based heat processing unit (HPU) designed to manage the cooling requirements of high-density data centres.
The company says its system can provide cooling without water consumption, while also recovering heat from computing equipment for cooling, electricity generation or other applications.
Infinity Turbine’s pulse supercritical CO2 heat pump. © Infinity Turbine
“We applied an aerospace systems-engineering approach to data centre thermodynamics,” said CJ Kalra, CTO and co-founder. “Our team designed HPUs to process the extreme heat of gigascale racks using a first principles based approach.”
HPUs enable heat reuse and PUE (power usage effectiveness) ratings quickly approaching 1.0 (the theoretical ideal score) without water consumption or PFAS chemicals for AI factories, explained Kalra.
Its modular system is designed around CO2 as the working fluid and is aimed at high-density AI infrastructure, where the amount of heat generated by GPU clusters is becoming a major engineering challenge.
Karman claims its technology can reduce cooling energy consumption by 25% compared with conventional air-cooling systems, although those figures are the company’s own claims and would need to be assessed against independent performance data.
There could also be a regulatory incentive for data centre operators to consider CO2-based cooling. The US American Innovation and Manufacturing Act is driving a phasedown of hydrofluorocarbon refrigerants, which are used in some chillers and other data centre cooling equipment.
As restrictions on high-GWP refrigerants tighten, such as CO2 could become more attractive alternatives.
