ADVANTAGES OF LIQUID COOLING IN AI DATA CENTERS Many liquid cooling deployments are driven by compute-intensive workloads that require higher performance and thermal design power (TDP) than conventional applications. Generative AI (GenAI), for example, relies on graphics processing units (GPUs) that deliver up to 12 times the processing performance of general purpose CPUs. Beyond performance, data center operators increasingly prioritize energy efficiency, lower operational costs, and improved hardware reliability. Liquid cooling helps maintain optimal operating temperatures, reducing thermal-induced failures and extending equipment lifespan. It also supports heat reuse strategies by capturing and redirecting waste heat to nearby systems or buildings—advancing both sustainability and energy efficiency goals. Liquid cooling further reduces facility-wide water consumption and decreases reliance on high speed fans. Immersion cooling, in particular, can optimize space by eliminating traditional hot and cold aisle layouts. Some configurations also support concurrent maintainability—allowing operators to service or replace rack-level components without taking the system offline. Although overall data center energy demands may increase over time, liquid cooling can improve total cost of ownership (TCO) by increasing thermal efficiency and reducing cooling-related energy consumption. LIQUID COOLING ARCHITECTURES: DLC AND IILC EXPLAINED Liquid cooling systems are generally categorized as either DLC or IILC. DLC uses a liquid coolant— typically water or dielectric fluid—to transfer heat directly from electronic components. It provides significantly higher efficiency than traditional air cooling, especially for infrastructure with high thermal design power (TDP) ratings. There are two primary types of DLC configurations. In direct-to-chip (DTC) cooling, liquid is pumped
through cold plates attached to heat-generating components, with air cooling handling any remaining thermal load. In immersion cooling, entire servers are submerged in a dielectric liquid that absorbs and dissipates heat. Both DTC and immersion systems can operate in single-phase or two-phase modes. Single-phase systems circulate liquid continuously through a heat exchanger without phase change, while two-phase systems use engineered fluids that evaporate during heat absorption and condense back into liquid form. Some hybrid systems combine DTC and immersion within a single deployment. Unlike DLC, IILC removes heat from exhaust air at the rack or server level using liquid-based components such as rear door heat exchangers (RDHx). While IILC is less efficient than DLC, it significantly outperforms traditional room- or row-level air cooling and is easier to integrate into existing deployments. RDHx units can operate as standalone solutions or integrate with DTC systems for optimized thermal performance. OPPORTUNITIES AND BARRIERS Despite its benefits, liquid cooling deployments remain relatively limited compared to traditional air cooling, which offers proven reliability, lower initial costs, and broad industry familiarity. In contrast, liquid cooling typically requires higher upfront capital investment in specialized infrastructure and equipment, particularly in retrofit scenarios. Adoption on a broader scale has been slow due to challenges, including a lack of standardized metrics, design best practices, and performance benchmarks for liquid cooling systems. While issues like skilled labor shortages and evolving supply chains affect the entire data center industry, they can complicate liquid cooling deployments in particular. Additionally, restrictions on refrigerants with high global warming potential (GWP) and dielectric fluids containing per- and polyfluoroalkyl substances (PFAS) chemicals introduce environmental and compliance complexities. Concerns about long term reliability and limited rack level redundancy further contributes to operators' hesitation to adopt this promising capability.
Scaling for the Future: Liquid Cooling’s Role in AI Data Centers
By Mike Connaughton and Jacques Fluet
The TIA-942 standard has been a resource for global data center design and implementation for many years. The “c” revision (ANSI/TIA-942-C) was released in May 2024, and it covers a variety of aspects such as physical infrastructure, cabling, power, cooling, redundancy, and security. Data centers must support increasingly higher rack power densities as compute-intensive artificial intelligence (AI) and machine learning (ML) workloads scale. At the same time, operators are aiming to reduce energy consumption, extend equipment life, lower costs, and comply with stringent regulatory requirements. To meet these demands, many AI data center operators are considering liquid cooling solutions, which use water or dielectric fluids to transfer heat up to 20 times more efficiently than air. Despite these advantages, liquid cooling introduces a range of new design, operational, and infrastructure challenges. This article explores why data centers deploy direct liquid cooling (DLC) and in-rack indirect liquid cooling (IILC) systems, outlines key opportunities and barriers, and addresses infrastructure considerations for cabling and coolant routing. It also covers space, structure, and scalability requirements, highlights essential monitoring parameters, and details safety and operational needs— including redundancy and resilience.
To meet these demands, many AI data center operators are considering liquid cooling solutions, which use water or dielectric fluids to transfer heat up to 20 times more efficiently than air.
I
I
14
ICT TODAY
July/August/September 2025
15
Made with FlippingBook - Online catalogs