Data Center Cooling in 2026: Technology Options, Site Constraints, and the AI Advantage
As AI compute drives rack densities beyond what air can handle, cooling has moved from an operational detail to a development-level constraint.
For most of data center history, cooling was an operational concern, not a development one. Sites were selected for power and connectivity; cooling was engineered in afterward. That assumption broke around 2023, and it has not returned.
The cause is GPU density. A rack of Nvidia H100s draws 10 to 14 kW per server, meaning a standard 42U rack can hit 60 to 80 kW of heat load. The newer Blackwell B200 generation pushes higher still. Traditional air cooling maxes out at roughly 20 to 25 kW per rack under the best conditions. The gap between what AI compute demands and what air cooling delivers is now the defining constraint in data center engineering.
The downstream effects reach the developer before a single server is installed. Cooling technology choice drives site requirements for water, land area, and utility infrastructure. Getting it wrong at the design stage means either stranded capacity or a building that cannot serve the tenant mix it was built for.
The Cooling Technology Stack
Air Cooling: Still the Default, But Narrowing Fast
Conventional air cooling, computer room air conditioning (CRAC) and computer room air handlers (CRAH) with cold aisle/hot aisle containment, remains the baseline for general-purpose and enterprise workloads.
Where it works:
Rack densities below 20 kW
Enterprise and cloud storage workloads where GPU density is not the constraint
Markets where water costs or regulations make evaporative cooling expensive
Where it breaks down:
Any AI training or inference workload above 30 kW per rack
High-density deployments where floor area for airflow management is limited
Markets with high ambient temperatures where free cooling hours are limited
PUE for well-designed air-cooled facilities typically runs 1.25 to 1.45, depending on climate and operating efficiency.
Rear-Door Heat Exchangers
Rear-door heat exchangers (RDHx) are a transitional technology: water-cooled panels mounted to the back of standard racks that capture heat at the source before it enters the room airspace.
Capabilities:
Handles 20 to 40 kW per rack reliably
Can be retrofitted into existing raised-floor environments
Lower capital cost than full liquid cooling infrastructure
Limitations:
Not sufficient for current-generation GPU clusters at full density
Requires chilled water infrastructure in the room
Maintenance complexity increases compared to air-only environments
Direct Liquid Cooling
Direct liquid cooling (DLC) routes chilled water or other coolants directly to server components via manifolds and cold plates mounted to CPUs and GPUs. Heat is transferred to the coolant at the chip level and carried out of the rack entirely.
Capabilities:
Handles 40 to 100+ kW per rack
PUE as low as 1.10 to 1.15 with high-efficiency cooling plants
Required by most hyperscalers for AI training clusters as of 2025
Site implications:
Requires cooling distribution units (CDUs) at the row or room level
Needs higher-capacity chilled water plant infrastructure
Increases mechanical complexity and specialized maintenance requirements
Most new hyperscale builds announced in 2025 and 2026 specify DLC-ready infrastructure as a base requirement. Developers who are not designing for this are building legacy product.
Immersion Cooling
Immersion cooling submerges servers directly in dielectric fluid, either single-phase (liquid that remains liquid) or two-phase (liquid that boils and recondenses). It is the highest-density cooling approach currently in production deployment.
Capabilities:
Handles 100 kW per rack and above
PUE of 1.03 to 1.05 in optimized deployments
Significant reduction in water use compared to evaporative cooling
Limitations:
High capital cost: immersion tank infrastructure adds $1 to $2 million per MW compared to DLC
Limited vendor ecosystem for maintenance and fluid management
Not all server hardware is validated for immersion deployment (this is improving but still a constraint)
Primarily viable for dedicated AI training clusters with long operational stability
Evaporative and Hybrid Cooling
Many large facilities use evaporative cooling, wet cooling towers, as a heat rejection mechanism, either in combination with mechanical refrigeration or in climate-appropriate direct or indirect evaporative systems.
The water constraint:
A 100 MW data center using evaporative cooling can consume 200 to 400 million gallons of water annually, roughly the annual residential water use of a city of 2,000 to 4,000 homes. This is now a live regulatory issue.
Phoenix, in Maricopa County, has imposed water use restrictions on new data center developments. Singapore's national moratorium on new data centers from 2019 to 2022 was partly driven by water consumption concerns. Several California water districts have flagged data center expansion as a competing demand on limited groundwater.
Sites in markets with water stress, much of the Southwest, parts of Texas, and increasingly the Southeast, need a credible cooling water plan before they clear the site screening stage.
Site Constraints by Cooling Type
| Cooling Approach | Water Demand | Land Premium | CapEx Premium | Viable Rack Density |
|---|---|---|---|---|
| Air cooling | Minimal | None | Baseline | Up to 20 kW |
| Rear-door heat exchangers | Low | None | Low | Up to 40 kW |
| Direct liquid cooling | Moderate | Low | Moderate | Up to 100 kW |
| Immersion cooling | Low | Moderate | High | 100 kW+ |
| Evaporative (wet cooling) | High | Moderate | Low-moderate | Variable |
The right cooling approach depends on the tenant mix, the market's water position, and the target power density. There is no universal answer in 2026, which is exactly why modeling it early matters.
Where AI Fits in Cooling Analysis
Cooling decisions made at the feasibility stage are difficult to reverse. A building designed for 20 kW per rack average density cannot easily be converted to serve 80 kW AI compute clusters without major mechanical plant replacement.
AI-assisted feasibility modeling helps developers make these decisions with better data:
Thermal load modeling: AI can simulate heat loads across different rack density assumptions and cooling configurations, identifying the crossover points where technology choices change the economics.
Water budget analysis: Given site water availability data and planned cooling technology, AI can model annual water consumption across operational scenarios and flag sites where water constraints would bind.
Life cycle cost comparison: Upfront capital cost is not the only variable. Energy efficiency (PUE), maintenance cost, water cost, and operational complexity all feed into a 10-year NPV model. AI can run these comparisons across cooling configurations at the early design stage, before engineers are engaged.
Climate sensitivity analysis: Free cooling hours vary by market. A facility in Seattle captures free cooling potential for 6,000 or more hours per year. The same facility in Phoenix may capture fewer than 1,000. AI can adjust PUE assumptions and energy cost models by location, improving the accuracy of feasibility underwriting.
The Developer Takeaway
Cooling technology selection is no longer an engineering afterthought. It is a development decision that shapes site requirements, capital structure, and tenant eligibility before the first design document is produced.
Developers building for the AI compute market in 2026 need to answer three questions before they select a cooling approach: What rack density does the target tenant base require? What water constraints does the site impose? What is the all-in cost difference between cooling configurations over a 10-year hold?
AI makes those questions answerable at feasibility speed. The alternative is designing to assumptions that may not match the market by the time the building delivers.