What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
AI data centers need different power and cooling designs because AI accelerators can concentrate much more computing load—and therefore heat—into each rack. Facilities must deliver that electricity reliably and remove the resulting heat without disrupting equipment. The right design depends on the servers, rack configuration, and site; there is no single rack-power threshold at which every facility must switch to liquid cooling.
How AI changes the demands on a data center
AI workloads often run on servers packed with accelerators. The International Energy Agency (IEA) says the growth of AI is accelerating deployment of high-performance accelerated servers and increasing data-center power density. In practical terms, more electrical load and heat can be concentrated in a smaller part of the facility, especially within a rack.
That concentration matters even when a facility has enough total electricity on paper. Power must reach the equipment through the facility’s electrical systems, and cooling must capture and carry away heat where it is produced. A room-level design suited to less-dense equipment may not be sufficient for a rack whose servers place a much larger load in the same footprint.
Power and cooling are coupled: nearly all electricity consumed by IT equipment ultimately becomes heat that the facility must remove. Raising the IT load therefore increases both the electricity needed at the servers and the heat-removal burden. The exact requirements depend on the system and site, so general data-center averages cannot substitute for a design based on the intended equipment.
What the electricity figures do—and do not—tell you
Data-center electricity demand is growing, but the figures below measure different things. Annual energy consumption, demand growth, and the cooling share of a facility’s electricity use should not be treated as interchangeable measures of rack power.
| Measure | Figure | Scope and qualification |
|---|---|---|
| Data-center electricity use | About 415 TWh in 2024 | IEA estimate for data centers worldwide; about 1.5% of global electricity consumption that year. IEA, Energy and AI (2025). |
| Projected data-center electricity use | About 945 TWh in 2030 | IEA global Base Case projection, not a measured outcome. IEA, Energy and AI (2025). |
| Growth in data-center electricity demand | 17% in 2025 | Growth reported by the IEA in its 2026 summary; this is a one-year growth figure, not the 2030 projection above. |
| Cooling share of facility electricity | About 7% to over 30% | IEA estimates spanning efficient hyperscale data centers to less-efficient enterprise facilities. The share varies by facility; it is not a universal cooling allowance. IEA, Energy and AI (2025). |
The IEA also describes how the global increase in data-center electricity use from 2024 to 2030 is divided in its Base Case: nearly half of the net increase is attributed to accelerated servers, about one fifth to conventional servers, around one tenth to other IT equipment, and around one fifth to cooling and other infrastructure. These are scenario attributions, not universal measured shares for individual facilities.
Servers average around 60% of electricity use in modern data centers, according to the IEA, but the proportion varies. Storage, networking, cooling, UPS equipment, backup generation, and other infrastructure also contribute to facility demand. That broader system context is why planning only around a server’s nameplate load can miss important facility requirements.
Why the electrical design must adapt
A higher rack load means more electricity has to be delivered to the IT equipment and supported by the facility infrastructure around it. The design must account for the power path as a whole, including UPS systems and backup generation, rather than treating the grid connection as the only relevant component. The precise configuration and capacity depend on the planned servers and facility; the available figures do not establish a universal rack-power specification.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #3
- Plan for the intended IT load. Accelerator-heavy racks change the amount and concentration of power required at the equipment. Facility plans need to reflect the systems that will actually be installed.
- Include supporting infrastructure. Servers are only part of a data center’s electricity use. UPS, backup generation, storage, networking, and cooling belong in the facility-level picture.
- Coordinate capacity with heat removal. A power plan that supports a denser rack also needs a cooling plan capable of handling its heat output. Neither should be sized in isolation.
How cooling approaches differ
Cooling designs differ in where they capture heat. Conventional room air cooling moves heat from equipment into the room and then out through the facility cooling system. Rear-door heat exchangers capture heat at the rack’s rear. Direct-to-chip liquid systems use cold plates to collect heat closer to processors and other components; a coolant distribution unit (CDU) manages coolant between the equipment and the facility system. Immersion cooling places equipment in a liquid environment. These are different ways to move heat, not interchangeable solutions with one universally best option.
| Approach | Where heat is captured | Design considerations |
|---|---|---|
| Room air cooling | Air moving through the equipment and room | Room and airflow design must suit the equipment’s heat output. Dense racks can make assumptions based on lower-density equipment inadequate. |
| Rear-door heat exchange | At the rear of the rack | Heat is captured nearer the rack than with room-level cooling; compatibility depends on the rack and facility design. |
| Direct-to-chip liquid | At cold plates on heat-producing components | Requires compatible servers and coolant routing, plus a facility-side path to reject heat. NVIDIA describes cold plates and CDUs in its vendor-authored material. |
| Immersion cooling | In the liquid surrounding immersed equipment | Requires an equipment and facility design built for immersion; the sources available here do not establish an apples-to-apples cost or efficiency comparison with other approaches. |
Liquid cooling is a response to concentrated heat, not an automatic requirement for every AI server. NVIDIA’s April 2025 article presents liquid cooling as a way to reduce dependence on chillers and improve heat rejection, but that is a vendor statement, not an independent benchmark applicable to all systems. Its reference-design material also describes examples such as redundant CDU groups and rack-level isolation. Those illustrate possible design features; they are not universal requirements.
Choosing among approaches requires more than comparing a cooling technology’s label. Operators need to consider equipment compatibility, where heat is collected, how the secondary facility loop rejects it, operational access, redundancy, isolation, and leak monitoring. Site climate and water availability can also affect heat-rejection choices, but there is no location-independent water or cost outcome established by these sources.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the demand outlook means for facility planning
Global and U.S. projections provide context for the scale of planning, but they describe different geographies and horizons. They should not be combined into a single forecast.
| Projection | Figure | Scope and qualification |
|---|---|---|
| Data-center electricity use growth | Could double or triple by 2028 | U.S. projection summarized by the Department of Energy in a December 2024 announcement about an LBNL study. |
| Share of U.S. electricity use by decade’s end | 11.8%, with a 9.5%–15.3% range | LBNL 2025 update estimate as summarized in the Department of Energy’s 2026 resource hub. This is a U.S. share estimate, with a stated range, and uses a different horizon and framing from the 2024 projection. |
For an individual project, these national and global outlooks do not determine the required electrical or cooling capacity. They indicate why capacity planning has become more consequential; the facility’s actual design still has to be matched to its equipment, operating needs, and location.
What is established—and what is not
The evidence supports the central design case: AI is driving deployment of accelerated servers, increasing power density, and making heat removal a more demanding facility problem. It does not establish one universal point at which air cooling stops being adequate, a single best cooling method, or independent comparative costs, water use, and efficiency for air, direct-to-chip liquid, and immersion systems. Those outcomes need to be evaluated for a specific server and site rather than inferred from a vendor example or a sector-wide statistic.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
