Cooling and on-site generation heat
As computing density in artificial intelligence (AI) factories rises to unprecedented levels, thermal management is moving beyond operational support to become a key factor determining performance, reliability and growth.
Hyperscale AI factories generate 20 to 30 times as much heat per unit area as ordinary commercial buildings, and the electricity required for computing and cooling increases as AI workloads and chip power densities rise. Cooling alone uses up to 40% of a data center’s total electricity. Against this background, McKinsey and the International Energy Agency (IEA) expect global data center energy demand in 2030 to reach about three times its 2024 level. However, demand on electricity grids and lengthy related procedures are making it increasingly difficult to secure new power capacity.
On-site generation, which is becoming widespread at data centers, provides electricity quickly and improves operational control and resilience, but further complicates thermal management. Only some of the fuel energy supplied to on-site generation equipment, such as gas engines, gas turbines and fuel cells, is converted into electricity. About 57% emerges as thermal energy and may be released into the atmosphere without being used.
Because this waste heat represents energy that has already been paid for, data centers need to manage heat from on-site generation equipment as well as information technology (IT) equipment.
Effectively managing both heat sources requires more than improving the efficiency of individual cooling units. The entire heat flow needs to be designed as an integrated system, covering the transfer, rejection and reuse of heat generated by chips, as well as the use of waste heat from on-site generation for cooling. This can reduce energy use and carbon emissions, while allowing electricity saved on cooling to be used for additional AI computing.
Integrated designs and absorption chillers
Johnson Controls has spent decades designing, integrating and servicing thermal management systems for mission-critical environments. Drawing on this experience, it has developed standardized reference design guides integrating generation equipment, cooling infrastructure and energy optimization. These help operators use energy more productively and improve sustainability.
Absorption chillers are one option for managing both heat sources in an integrated way. Unlike conventional electric chillers, they produce chilled water using thermal energy such as hot water or steam, allowing waste heat from on-site generation to replace some of the electricity needed for cooling. Johnson Controls’ “1-gigawatt (GW) absorption chiller reference design guide” modeled an integrated system combining absorption chillers, combined heat and power units, dry coolers and related thermal management infrastructure. This model can reduce cooling electricity demand by up to 44%, allowing up to an additional 97 megawatts (MW) of electricity to be allocated to AI computing without increasing on-site generation capacity.
Cooling system design needs to consider not only power usage effectiveness (PUE), but also carbon usage effectiveness (CUE) and water usage effectiveness (WUE). Trade-offs can arise, as with evaporative heat rejection, which lowers electricity demand but increases water use.
The Johnson Controls model can lower PUE to as little as 1.23 and provide cooling without an on-site water supply through dry-cooler-based heat rejection. Reducing energy use by lowering cooling demand can also cut the cooling system’s carbon emissions by up to 43%.
Designing for technological change
Integrated thermal management design also needs to accommodate long-term technological change. GPU generations, rack densities and liquid-cooling technologies change rapidly, while cooling infrastructure remains in use for long periods. A modular design enables flexible responses to technological change and allows a data center to expand from a 100 MW campus to a GW-scale AI factory without a fundamental redesign.
As electricity becomes a key constraint on AI infrastructure development, competitiveness will depend not only on how much energy can be secured, but also on how productively it can be used through thermal management.
Original Korean article: Maeil Business
English brief by Korea Wind Intelligence, koreawindintel.com
