Data Center Cooling Trends and Innovations in 2026

Liquid Cooling Is Now a Data Center Design Requirement

In a January survey from IDC, about three-quarters of data center service providers reported having racks dedicated to AI or high-performance computing in production — a share IDC expects to reach 93% by 2027.

“Demand is strong and accelerating, but it’s still largely supplier-led rather than customer-led,” says Luis Fernandes, IDC senior research manager for European infrastructure strategies.

He explains that data center consultants and providers are making thermal management a core design requirement to handle new AI-driven rack densities.

Fernandes says data center cooling providers are joining design discussions earlier, reducing the need for later retrofits. Less intensive inference workloads may still use air cooling, but new and modernized facilities should evaluate whether they will eventually support racks at 100 kilowatts or higher.

WATCH: A renewed interest in on-premises data centers is driving escalated power consumption.

“Demand is surging as AI workloads push rack densities beyond the practical limits of traditional air cooling,” says Derek Chung, global application engineer at Honeywell Technologies. “Data center operators are seeking cooling capacity along with energy efficiency, water reduction and operational reliability.”

Liquid transfers heat more effectively than air, supporting greater rack density while reducing fan energy and stabilizing temperatures.

“Air cooling has run out of physical headroom,” Stengel says. “Current-generation accelerators draw 700 to 1,200 watts each, and air cooling stops being economical above roughly 40kW to 50kW per rack.”

IDC forecasts that 90% of large-scale, performance-intensive computing deployments will use direct liquid cooling by 2027.

Comparing Direct-to-Chip vs. Immersion Cooling for AI Deployments

Direct-to-chip cooling has become the leading option for new AI deployments. Coolant passes through cold plates attached to high-heat components, carries heat to a rack manifold and reaches a coolant distribution unit, which transfers it to a separate facility loop.

“Direct-to-chip has emerged as the default for OEM server vendors shipping AI-ready infrastructure,” Fernandes says. “It arrives largely pre-integrated from the factory, which matters when hyperscalers are deploying at speed.”

Pre-integrated systems can reduce implementation risk, but operators still need plumbing, leak detection and trained personnel. Active rear-door heat exchangers offer existing facilities a less disruptive retrofit that requires no server or floor layout redesign.

Immersion cooling submerges servers in a nonconductive dielectric fluid. It cools the entire system, eliminates fans and can reduce thermal stress, making it useful at extreme densities or edge sites with dust, humidity or inadequate ventilation.

Adoption is harder because facilities must accommodate tanks, technicians need new procedures and servers require immersion-compatible components. Direct-to-chip fits more readily into conventional rack operations.

How AI Systems Adapt Facility Cooling to Local Conditions

Cooling efficiency depends on far more than the selected hardware. Climate, water quality, power stability, facility design and local regulations change the performance of the same system from one site to another.

Fernandes says no two data centers have the same combination of size, terrain, climate, air and water quality, available power, and local regulation.

Digital twins allow operators to model those variables and test configurations before applying them. One site may prioritize power usage effectiveness, which compares total facility energy with energy delivered to IT equipment. Another may emphasize water usage effectiveness or redundancy.

AI can analyze real-time temperature, pressure and flow data to predict hotspots, adjust pumps and fans, forecast demand, and identify equipment degradation.

“AI-driven cooling optimization is IDC’s highest-conviction near-term software lever,” Stengel says.

He points to Google DeepMind’s reinforcement-learning cooling controller, which he says cut cooling energy by roughly 40% versus rule-based baselines.

“We are seeing greater use of sensor-driven, adaptive control strategies that adjust cooling based on real-time conditions rather than relying solely on fixed setpoints,” Chung adds.

Those controls depend on accurate measurement. Sensors operating around liquid cooling systems must withstand coolant exposure, moisture and condensation while detecting abnormal pressure, flow or temperature before it affects performance or uptime.

How Water and Heat Reuse Affect Data Center Site Planning

Closed-loop cooling can reduce direct water consumption because coolant circulates instead of evaporating. It does not automatically eliminate a facility’s wider water impact: Captured heat must still be rejected through chillers, dry coolers, cooling towers or another system.

Dry cooling can increase energy demand, while reclaimed water is not available everywhere. Fernandes says operators must assess water use across the entire facility, including the effects of transferring heat into natural water sources.

Some facilities can treat captured heat as an output rather than a disposal problem. Liquid-cooled data centers can return water at temperatures between 45 and 60 degrees Celsius (113-140 degrees Fahrenheit), making it usable directly or through heat pumps for district heating.

“It’s not really waste heat; it’s high-grade heat with a high potential temperature difference,” Fernandes says. “If captured and distributed correctly, that remaining energy can be put to use for community or industrial reuse.”

READ MORE: How are data centers adapting for artificial intelligence?

However, the economics remain highly local. Stengel says a heat buyer generally needs to be located within about three kilometers of the data center, while contracts can extend beyond 20 years. District heating infrastructure has made reuse more viable in Stockholm, Helsinki, Dublin and Frankfurt than in most U.S. markets.

IT leaders comparing cooling systems should match the approach to current rack density and future expansion, then evaluate installation cost, energy and water use, coolant compatibility, monitoring requirements, and available skills.

They should also determine whether the system can adjust to changing workloads without compromising reliability.

“As data centers transition from air cooling to intelligent liquid cooling, sensing becomes even more mission-critical,” Chung says. “With more adaptive, connected and demand-driven systems comes the need for trusted sensor data to help optimize operation dynamically.”