Liquid Cooling and the Changing Reality of AI Infrastructure

From Air to Liquid: An Operator’s Blueprint for AI-Ready Data Centers

ARTICLE SOURCE: Jack Elsea, Vice President – Design Integration

Atlanta, GA – July 1, 2026 — The rapid adoption of AI and advanced computing workloads is fundamentally changing data center design. GPU (Graphic Processing Units) applications now routinely drive rack densities beyond 100kW per rack, pushing well beyond the practical limits of traditional air-cooling systems, which typically support approximately 30kW per rack. As AI deployments continue to scale, operators must increasingly adopt cooling strategies specifically designed to handle these unprecedented thermal loads.

The End of Air-Cooling as We Know It

While it remains possible to deploy GPU infrastructure at lower densities or distribute high-density deployments across a larger air-cooled footprint, doing so introduces significant inefficiencies. Spreading workloads across more white space increases both construction costs and ongoing operational expenses by requiring larger facilities than would otherwise be necessary. Additionally, greater physical separation between GPU racks results in longer network cabling distances, which can negatively impact the performance of interconnected GPUs where low latency is critical to achieving optimal computational efficiency.

To address these density challenges, the industry has largely converged around two primary liquid-cooling methodologies: immersion cooling and direct-to-chip liquid cooling as essential components of modern AI-ready infrastructure. Both technologies provide significantly greater heat removal capabilities than air-cooling alone but the choice between them carries implications for maintenance access, long-term flexibility and operator familiarity.

In the U.S. market, direct-to-chip liquid cooling has emerged as the preferred solution. Most deployments still require a hybrid approach because many GPU rack components remain air-cooled, while network racks often continue to operate entirely on air-cooled systems and are typically located alongside GPU production environments. As a result, today’s facilities commonly operate within cooling ratios ranging from approximately 30% air and 70% liquid cooling to as much as 10% air and 90% liquid cooling.

When Legacy Infrastructure Meets AI Demand

Retrofitting existing facilities for liquid cooling presents substantial challenges. Existing chilled water plants may have temperature ranges, cooling capacities, and pipe sizing that do not align with the requirements of coolant distribution units (CDUs) or the desired air-to-liquid cooling ratios needed for peak performance. Successfully integrating these systems often requires significant modifications to both mechanical and electrical infrastructure. For operators, this means often shifting how capital planning is approached, requiring earlier coordination between development, operations, and tenant requirements.

Organizations with experience modifying sophisticated systems in live operating environments are uniquely positioned to navigate these retrofit challenges. Performing major upgrades while maintaining uninterrupted service to active data halls requires detailed planning, rigorous safeguards, and extensive expertise. The process is often comparable to conducting open-heart surgery on critical infrastructure while keeping the patient fully operational, all in support of next-generation computing technologies. As AI-driven infrastructure demands continue to accelerate, organizations with proven retrofit capabilities will play a critical role in enabling operators to modernize existing facilities quickly, cost-effectively, and with minimal operational risk.

These challenges are not only technical, but they are also reshaping how data center infrastructure is owned, operated, and maintained. As liquid cooling systems extend deeper into white space environments, traditional boundaries between stakeholders are beginning to shift.

The rise of direct-to-chip cooling is redefining traditional boundaries between landlords and tenants. In air-cooled environments, the mechanical demarcation point typically keeps landlords outside the white space, limiting their operational responsibilities. With liquid cooling, landlords may now own, operate, and maintain portions of the technical water loop that extend into customer environments.

This introduces new operational and contractual complexity. Systems located above or near high-value customer equipment raise concerns around leaks, maintenance activities, and risk allocation, requiring updated approaches to accountability, procedures, and risk management.

Service level agreements are also evolving. Questions arise around how performance guarantees are defined when components such as CDUs may be purchased by the tenant but installed or maintained by the landlord. Determining what constitutes an SLA violation requires closer coordination among all stakeholders.

As the industry matures, landlords, tenants, operators, and manufacturers must collaborate to establish clear, practical frameworks that define responsibilities and resolution paths when issues occur—balancing operational reality with contractual clarity.

Building for Hardware That Doesn’t Exist Yet

Complicating matters further is the pace of GPU innovation itself. Technology roadmaps are advancing faster than data centers can be designed and constructed. Developers are often designing facilities around future hardware platforms whose specifications may not yet be finalized, or may not even exist, at the time design decisions are being made. This creates significant uncertainty when planning infrastructure that may not become operational for 18 months or longer.

GPU manufacturers are also pursuing different approaches to technical water loop operating temperatures and cooling architectures adding another layer of complexity. These variations make it increasingly difficult to create standardized designs capable of supporting multiple generations of hardware over the life of a facility.

Where Cross-Industry Expertise Creates an Advantage

While the industry has extensive experience designing and operating large, chilled-water plants, technical cooling systems introduce new requirements and intricacies.

Unlike industries that have long used process water systems, data centers still lack universally accepted standards for technical water loop purity, fluid specifications, monitoring, commissioning, and operations. As liquid cooling adoption accelerates, owners, operators, designers, and builders have an opportunity to establish the best practices that will define next-generation AI infrastructure.

Many of these challenges have already been solved elsewhere. Industries such as life sciences, healthcare, laboratories, food and beverage, and advanced manufacturing have decades of experience designing and operating high-purity fluid systems under strict performance and regulatory standards. Their established testing methodologies and quality control practices offer valuable guidance for the data center sector.

Experience Matters When the Stakes Are High

As AI infrastructure evolves, success depends on executing complex transformations without disrupting operations. Upgrading live facilities, integrating liquid cooling systems, and preparing for next-generation hardware require technical expertise, operational discipline, and precise execution.

Evans brings deep experience delivering complex mechanical and electrical infrastructure in mission-critical environments, with a proven track record of modifying sophisticated systems in live facilities while maintaining full operational continuity in active data halls during construction and upgrades.

Whether supporting liquid cooling integration, expanding capacity, or preparing for future AI deployments, Evans understands that infrastructure modernization often requires detailed sequencing, contingency planning, and a deep understanding of how mechanical and electrical systems interact under real-world operating conditions without interrupting service.

The transition from air-cooled to liquid-cooled data centers is reshaping how facilities are designed, built, and operated. Over the next several years, liquid cooling will not be a niche solution. It will become a baseline requirement for high-performance computing environments where success will depend on aligning innovative and advanced cooling strategies with proven execution to lead the industry forward.