Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsSome links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
AI is turning data-center thermal design into a first-order construction and capacity issue. The main change is not simply that AI consumes more electricity. AI accelerators concentrate substantially more power and heat in individual racks, rows, and pods, while training and inference workloads can produce rapid load changes. Every watt used by IT equipment ultimately becomes heat: a 100 kW AI rack requires approximately 100 kW of continuous heat-removal capacity at full IT load.
Air cooling remains suitable for lower-density and mixed environments. However, high-density accelerator deployments increasingly require direct-to-chip liquid cooling, rear-door heat exchangers, immersion, or a hybrid architecture. The correct choice depends on rack power, workload variability, climate, water availability, electrical capacity, retrofit constraints, reliability requirements, and the facility’s ability to operate and maintain liquid systems.
Why AI changes data-center heating and cooling
AI workloads typically combine large numbers of GPUs or other accelerators with high-speed networking, CPUs, memory, and interconnect equipment. This creates much greater heat concentration than a conventional enterprise server room. The International Energy Agency reports that AI-server power density increased elevenfold between 2020 and 2025 and could increase another fourfold by 2027. That trend is global analysis, not a universal specification for every AI rack, but it illustrates the direction of travel.
Free tools Windows power users keep installed
One-click scans. No signup required.
The construction implication is that total site megawatts are only part of the design problem. Engineers must also determine how much power and heat are concentrated in each rack, row, pod, cooling loop, electrical busway, and structural zone.
Current rack-scale systems show the density challenge. NVIDIA’s GB200 NVL72 design connects 72 Blackwell GPUs and 36 Grace CPUs in a rack-scale configuration and uses liquid cooling for its most power-intensive components. The exact power and thermal requirements still depend on the installed configuration and operating limits.
AI workloads also vary considerably. Training, inference, model evaluation, batching, and test-time reasoning can have different power profiles. Rapid changes in demand can stress pumps, fans, chillers, controls, and heat-rejection equipment even when the annual average load appears manageable.
Cooling can account for approximately 7% of consumption in efficient hyperscale facilities and more than 30% in less-efficient enterprise data centers, according to the IEA. Those percentages vary with climate, design, utilization, and measurement boundaries.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The basic heat calculation
For practical facility planning:
- 1 kW of IT power produces approximately 1 kW of heat.
- A 100 kW rack produces roughly 100 kW of IT heat at full load.
- A 1 MW AI cluster produces approximately 1 MW of IT heat before cooling, power-conversion, networking, and other facility overheads.
These figures should not be confused with total facility demand. IT load covers processors, memory, storage, networking, and server components. Cooling load includes fans, pumps, chillers, cooling towers, dry coolers, coolant-distribution units, and controls. Facility load includes IT and all mechanical, electrical, lighting, and support systems.
Designers must size for both average and peak conditions. Accelerator power depends on the model, batch size, precision, utilization, communication intensity, memory traffic, power caps, and workload scheduling. Nameplate power alone is not a reliable prediction of sustained heat output.
When air cooling reaches its practical limits
Air has much lower heat capacity and thermal conductivity than liquid. As rack power rises, an air-cooled facility needs more airflow, larger or faster fans, tighter containment, shorter air paths, and more carefully controlled supply and return conditions.
High-density air cooling can be constrained by:
- hot spots at GPUs, CPUs, networking ASICs, or power supplies;
- fan energy and pressure drop;
- air recirculation and bypass airflow;
- perimeter cooling capacity;
- raised-floor or overhead-distribution limits;
- return-air capacity and heat-rejection capability;
- the amount of floor space required for cooling equipment.
There is no universal rack-power threshold at which air cooling suddenly fails. Air cooling can remain appropriate for conventional servers, storage and networking equipment, lower-density AI deployments, and accelerator systems operating within their manufacturer’s air-cooled specification. The more accurate conclusion is that high-density AI deployments increasingly need liquid or hybrid cooling.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11ASHRAE’s 2026 AI Data Center Energy Performance Framework discusses direct-to-chip liquid cooling, rear-door heat exchangers, and thermally segmented zones for racks in the 50–100+ kW range. This is design guidance, not a universal liquid-cooling cutoff.
Cooling architecture options
Traditional air cooling
CRAC or CRAH units condition room air, while cold aisles, hot aisles, containment, and raised-floor or overhead distribution guide airflow.
Best suited to: low- and medium-density racks, mixed workloads, incremental deployments, and existing facilities with adequate CRAH/CRAC capacity.
Advantages: mature maintenance practices, broad technician familiarity, no liquid near server electronics, and relatively simple server replacement.
Limitations: high airflow and fan requirements, difficult hot-spot control, limited rack-density headroom, and potentially extensive room or floor modifications.
Rear-door heat exchangers
A rear-door coil removes heat from server exhaust before it enters the data hall. This can be an effective retrofit when the rack is too dense for room-air cooling but does not require liquid connections to every chip.
Advantages: preserves much of the existing server architecture, reduces hot-air recirculation, and supports a staged move toward higher density.
Limitations: adds rack weight and service complexity, requires facility-water distribution, does not remove all room-air cooling, and may not capture heat from the highest-power components as directly as cold plates.
ASHRAE identifies rear-door heat exchangers as an important option alongside direct-to-chip systems for high-density AI zones. Selection should be based on the complete rack and facility design.
Rank #3
Direct-to-chip liquid cooling
Cold plates mounted directly to GPUs, CPUs, or other high-power components transfer heat into a technology cooling loop. A coolant-distribution unit, or CDU, uses heat exchangers to connect that loop to the facility loop.
Advantages: captures heat at the source, reduces dependence on room airflow, supports higher rack power, can reduce fan and mechanical-cooling energy, and may permit warmer facility-water temperatures that increase free-cooling hours. The U.S. Department of Energy describes direct liquid cooling as transferring heat into a recirculating liquid loop rather than first transferring it to room air.
Construction and operational requirements: CDUs, manifolds, piping, valves, quick disconnects, sensors, water-quality management, leak detection, service clearances, compatible cold plates, and procedures for isolating or draining equipment. NVIDIA’s GB200 documentation includes liquid-cooling manifolds and leak detection, demonstrating that liquid cooling adds operational systems rather than eliminating thermal-management work. OEM requirements and warranties must be checked before procurement.
Recommended Free Tools
Single-phase immersion cooling
Servers are submerged in a nonconductive dielectric fluid. Heat transfers into the fluid and then through a heat exchanger to the facility loop.
Immersion can provide high heat-transfer performance, reduce server-fan dependence, and support dense purpose-built deployments. It also changes the building and operations model: tanks may require greater structural capacity, lifting equipment, filtration, fluid handling, different service procedures, and hardware specifically validated for immersion.
It is therefore not automatically superior to direct-to-chip cooling. It may be a poor fit for mixed infrastructure, conventional rack-by-rack servicing, or hardware with incompatible warranties.
Hybrid cooling
Hybrid systems use liquid cooling for GPUs and CPUs while retaining air cooling for memory, drives, power supplies, networking, optical equipment, and other components. This is often the most practical approach because it matches the cooling method to component heat density.
NVIDIA’s GB200 documentation and Vertiv’s reference design both use hybrid arrangements. A liquid-cooled GPU does not make the entire rack liquid cooled.
| Architecture | Strongest fit | Main construction concern |
|---|---|---|
| Air | Low or moderate density and mixed workloads | Airflow, hot spots, and room cooling headroom |
| Rear-door heat exchanger | High-density retrofit | Water distribution, rack weight, and residual air cooling |
| Direct-to-chip | New high-density GPU facilities | CDUs, manifolds, leak detection, water quality, and redundancy |
| Immersion | Purpose-built extreme-density deployments | Tank structure, fluid management, service, and compatibility |
| Hybrid | Mixed generations and liquid-cooled accelerator racks | Coordinating two thermal systems and their controls |
What changes in facility construction
Cooling plant and heat rejection
AI facilities may require larger heat exchangers, higher-temperature liquid loops, additional dry-cooler or cooling-tower capacity, and chillers selected for the target supply and return temperatures. Warm-water designs can increase free-cooling hours in suitable climates. ASHRAE identifies high-temperature chillers and dry coolers as retrofit and modernization strategies.
Liquid cooling moves heat; it does not make heat disappear. The facility must still reject heat outdoors or reuse it. A dry cooler generally minimizes on-site evaporative water use but may require more fan energy or equipment in hot weather. Cooling towers can be energy-efficient but consume water through evaporation and blowdown. Adiabatic systems use water during hot or peak conditions to improve heat rejection.
Electrical and mechanical coordination
Pumps, CDUs, chillers, fans, controls, and heat-rejection systems consume electricity and need reliable power. An undersized electrical system can disable cooling, while an undersized cooling loop can strand available compute capacity.
Cooling redundancy must be assessed from the server cold plate or tank through the rack manifold, CDU, facility loop, pumps, controls, chiller or dry-cooler plant, electrical supply, leak detection, emergency shutdown, and outdoor heat rejection. Redundant chillers do not eliminate a single point of failure in a CDU or rack manifold.
Structure, space, and access
AI racks can impose greater floor loading and power-distribution demands. Liquid systems add CDUs, manifolds, piping, valves, sensors, and service clearances. Immersion tanks add structural and handling requirements. Construction documents should reserve routes for overhead piping, provide isolation valves and drainage where appropriate, and preserve access for maintenance and replacement.
Controls and transient response
A plant designed only around average load may not respond safely to rapid AI power changes. Commissioning should evaluate ramp rates, thermal inertia, pump response, control-loop stability, temperature excursions, workload coordination, and behavior during loss of a pump, CDU, chiller, or electrical source.
Energy, water, and carbon consequences
Liquid cooling can reduce cooling electricity, particularly by reducing fan energy and enabling warmer water temperatures. It does not guarantee a lower total facility energy use. Pumps, CDUs, heat exchangers, dry coolers, chillers, and controls all add load.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Water use depends largely on how the facility rejects heat:
- Dry coolers: usually minimize on-site evaporative water consumption, with possible energy and equipment trade-offs.
- Cooling towers: reject heat efficiently but consume water through evaporation and blowdown.
- Adiabatic systems: use water during hot or peak periods.
- Warm-water liquid cooling: may enable more dry cooling or free cooling where climate and allowable temperatures permit.
ASHRAE describes warm-water liquid cooling and dry coolers as pathways toward very low or near-zero operational cooling-water use in suitable designs, not as a universal outcome. The site climate, heat-rejection system, coolant temperatures, and operating profile determine the result.
Reports should distinguish on-site water withdrawal, on-site water consumption, water embedded in electricity generation, manufacturing water, local watershed stress, and annual average versus peak-season demand. A “zero-water” claim may refer only to operational site cooling under specified conditions.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.New construction versus retrofit
Greenfield AI facility
Use the target accelerator generation and rack density as early architectural inputs, not as late equipment substitutions. Reserve space for redundant CDUs, manifolds, heat exchangers, pumps, controls, service clearances, and expansion. Specify supply and return temperatures, water quality, leak response, heat-rejection mode, structural loading, electrical redundancy, and a path for future accelerator generations.
Enterprise facility adding a few AI racks
First verify whether the existing air system, electrical busway, floor loading, containment, and heat-rejection plant can support the proposed rack’s sustained and peak power. If room cooling is the bottleneck, rear-door heat exchangers or a limited hybrid zone may be more practical than converting the entire building to direct-to-chip cooling.
High-density retrofit
Do not assume that spare floor area or utility power means the building is AI-ready. Check structural loading, ceiling height, overhead piping routes, white-space clearances, CDU locations, supply and return temperatures, pump and heat-exchanger redundancy, water treatment, leak containment, fire protection, controls integration, maintenance access, warranties, and commissioning capability. ASHRAE specifically identifies density and power volatility as major modernization issues.
Metrics that should guide decisions
Power Usage Effectiveness (PUE) is:
PUE = Total facility energy / IT equipment energy
PUE shows facility overhead but does not isolate cooling, water, carbon, or useful AI output. Pair it with:
- WUE: annual site water usage divided by IT equipment energy;
- WUI: a location-sensitive measure of water-use impact;
- CUE: carbon emissions related to facility operation, influenced by grid mix and accounting method;
- DCRE and IT-work-capacity measures: indicators of how effectively infrastructure supports useful compute.
ASHRAE’s framework recommends tracking PUE, WUE, WUI, CUE, DCRE, and server utilization or IT-work-capacity measures. The most meaningful comparison is useful AI work per unit of electricity, water, carbon, and occupied capacity.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →How to evaluate vendor claims
Require every efficiency or water claim to state:
- the baseline system and technology generation;
- whether the boundary is chip, rack, cooling plant, or whole facility;
- climate and design outdoor conditions;
- workload, utilization, power caps, and measurement period;
- coolant supply and return temperatures;
- heat-rejection technology;
- redundancy assumptions;
- whether results are measured, modeled, or vendor estimates;
- whether water means withdrawal, consumption, or only on-site operational use.
For example, NVIDIA reports substantial water and energy advantages for liquid-cooled Blackwell systems, but these are vendor-reported comparisons based on specified conditions. They should not be generalized to an entire facility without examining the baseline, climate, workload, heat-rejection method, and system boundary. Vertiv’s 1.2 MW reference design includes eight 132 kW racks using a 76% direct-to-chip and 24% air-cooling topology, but it is a reference design rather than an industry-wide average. Site validation remains essential.
Pre-construction and procurement checklist
Hardware
- Accelerator model and maximum board power
- CPU, memory, storage, networking, and power-supply configuration
- Rack nameplate and expected sustained power
- OEM-approved cooling method and warranty requirements
- Residual air-cooled components
Workload
- Training, inference, evaluation, or mixed use
- Expected utilization and batch profile
- Peak and average demand
- Workload ramp rate and thermal transients
- Redundancy, failover, and thermal-throttling tolerance
Facility
- Available electrical capacity and redundancy
- Cooling capacity at design outdoor temperature
- Rack and floor loading
- CDU, manifold, piping, and service locations
- Supply and return temperatures
- Pump, heat-exchanger, CDU, and heat-rejection redundancy
- Leak detection, water treatment, isolation, and emergency procedures
- Maintenance access, controls integration, and commissioning capability
Sustainability
- PUE, WUE, WUI, and CUE
- Annual and peak water demand
- Local water stress and source water
- Water associated with electricity generation
- Heat-reuse potential
- Useful AI work per unit of energy and water
Commercial planning
- Interoperability between servers, CDUs, manifolds, and controls
- Service-level agreements and replacement-part availability
- Telemetry for power, flow, temperature, and leaks
- Commissioning and operator training
- Expansion capacity for the next accelerator generation
The practical buying path is staged: measure current electrical and thermal headroom, establish a target rack density, validate the cooling topology, then obtain quotations for CDUs, heat rejection, controls, monitoring, and commissioning. Public pricing is generally quote-based and site-specific.
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.
Leave a Reply