Liquid Cooling the AI Data Center

Data centers have always been thermal management problems. The compute efficiency gains from transistor scaling have historically been accompanied by proportional increases in heat density, and the data center industry has spent decades building increasingly sophisticated air handling systems to manage the consequences. Through the server generations of the 2000s and 2010s, air cooling kept pace. Hot-aisle/cold-aisle containment, precision computer room air handlers (CRAHs), and economizer cooling allowed power densities of 5–15kW per rack to be managed without fundamental infrastructure rethinking.

The AI accelerator generation has broken that model decisively. A single NVIDIA H100 DGX system — a full 8-GPU configuration — draws approximately 10–12kW. A fully loaded H100 rack runs 60–80kW. The GB200 NVL72, NVIDIA's flagship configuration for the Blackwell generation, draws approximately 120kW per rack. The physics of convective air cooling cap reliable heat removal at 20–30kW per rack — a ceiling set by the heat capacity of air, fan speeds that remain within acoustic and mechanical limits, and the practical limits of airflow velocity in a server chassis. At 120kW rack density, air cooling does not merely underperform; it fails entirely. Coolant temperatures would exceed safe operating thresholds within minutes without supplemental liquid cooling.

This is not a future problem that the industry can defer until it becomes urgent. It became urgent in 2024, when major hyperscalers began qualifying H100 clusters at densities that already required rear-door heat exchangers or cold plate cooling to maintain uptime. The GB200 ramp makes liquid cooling the default specification for any new AI compute deployment, and the installed base of air-cooled facilities faces a 2–3 year retrofit cycle that will absorb significant capex regardless of the macroeconomic backdrop.

The Power Density Problem

The fundamental constraint is the heat capacity of air. Air has a volumetric heat capacity of approximately 1.2 kJ/m³·K — meaning that moving a cubic meter of air through a server chassis and allowing it to warm by 10°C removes roughly 12 kJ of heat. To remove 120kW (120,000 Joules per second) via air cooling, you would need approximately 10,000 cubic meters of airflow per second through a single rack — a physically impossible volumetric flow rate in any data center environment. The practical ceiling for high-end air cooling, with state-of-the-art containment and airflow management, is approximately 20–30kW per rack.

Water has a volumetric heat capacity roughly 3,400 times that of air. A small flow of coolant — 5–10 liters per minute — can absorb the full thermal output of a 120kW rack with a modest temperature rise, provided the cooling loop is connected directly to the heat-generating components via cold plates or immersion. This is the fundamental physics advantage of direct liquid cooling, and it is why the transition is not a question of if but when — and the answer is: now, for every new GB200-class deployment.

The industry threshold was crossed definitively in 2024. Prior to H100 density deployments, the majority of enterprise and hyperscaler compute ran at rack densities compatible with advanced air cooling. The H100 DGX pushed the edge of air cooling viability; the GB200 NVL72 is well beyond it. The PUE (Power Usage Effectiveness) implications are also significant: air-cooled facilities typically run PUE of 1.4–1.6, meaning that for every watt of IT load, 0.4–0.6 additional watts go to cooling infrastructure. DLC systems achieve PUE of 1.1–1.2, and immersion cooling can reach 1.02–1.05. At scale, this difference in PUE represents tens of millions of dollars annually in electricity cost for a large hyperscaler cluster.

Cooling Technologies Compared

The data center cooling market is not monolithic. Four principal technology categories address different points on the power density spectrum, with distinct cost, complexity, and PUE profiles. Understanding where each technology sits is essential for mapping which deployments drive demand for which supply chain layers.

The practical near-term market is dominated by direct liquid cooling (DLC) with cold plates, rather than immersion. Immersion cooling offers the best PUE and the highest power density ceiling, but requires purpose-built facility infrastructure — floor loading for immersion tanks, specialized fluid handling, and server hardware designed for fluid contact. For the vast majority of existing data center facilities deploying H100 and GB200 clusters, cold plate DLC is the technically feasible and cost-optimal solution within the 2–3 year retrofit window.

Rear-door heat exchangers represent the lowest-disruption upgrade path for facilities in the 30–50kW density range. They can be retrofitted on existing racks without modifying server hardware, making them attractive for operators who need an incremental thermal solution while planning a more comprehensive DLC deployment. Several major colocation providers have adopted rear-door units as a bridge technology while constructing DLC-ready white space.

Supply Chain & Key Players

The DLC supply chain operates in four layers. Each layer has a distinct competitive structure, customer base, and margin profile. Investors need to map their exposure to the specific layers where demand and pricing power are most concentrated.

Layer 1: Chilled Water Plant & Building Infrastructure

The foundation of any DLC installation is the chilled water plant that provides sub-15°C water to the cooling distribution units. This layer is dominated by large HVAC/industrial players — Johnson Controls, Trane Technologies, Carrier Global — who build the chillers, cooling towers, and distribution infrastructure at the facility level. Margins in this layer are governed by large project contract economics; growth is tied to new data center construction rather than retrofit cycles.

Layer 2: Coolant Distribution Units (CDUs)

CDUs are the critical interface between the facility chilled water loop and the rack-level cooling manifolds. They take facility water at 7–15°C, manage flow rate and pressure regulation, and deliver coolant to individual rack manifolds. CDUs are engineered products with meaningful IP and qualification cycles — hyperscalers qualify specific CDU models for their standard rack designs, creating switching costs similar to those in the memory market. Key players include Vertiv Holdings (NYSE: VRT), Schneider Electric, and specialist vendors Motivair Corporation and CoolIT Systems.

Layer 3: Cold Plates & Manifolding

Cold plates attach directly to CPUs, GPUs, and memory controllers, providing the heat-exchange surface where the thermal transfer actually occurs. Cold plate design is a precision engineering problem: the micro-channel structure, fluid velocity profile, and thermal interface material selection all affect thermal resistance. Key players include Asetek (NASDAQ Copenhagen: ASTK), Boyd Corporation, Aavid (Modine), and Alfa Laval (Stockholm: ALFA). NVIDIA's GB200 NVL72 ships with a specified cold plate design that third-party vendors must qualify against — a dynamic that concentrates volume in NVIDIA-validated suppliers.

Layer 4: Rack-Level Integration & Management

The final layer integrates CDUs, cold plates, manifolding, and leak-detection systems into a managed rack solution with DCIM (data center infrastructure management) software. This layer captures the highest integration margin and creates the stickiest customer relationships. nVent Electric (NYSE: NVT), Vertiv, and emerging pure-plays like ZutaCore and Iceotope compete at this level. The software management layer — monitoring coolant temperature, flow rates, and thermal envelope compliance — is increasingly differentiating as hyperscalers demand tighter integration with their cluster management systems.

Hyperscaler Adoption Curves

The five major hyperscalers — Microsoft, Google, Amazon AWS, Meta, and Oracle — are in different stages of DLC adoption, reflecting their different AI cluster architectures, facility vintage, and capex cycles. Their adoption trajectories collectively define the demand curve for DLC supply chain vendors over the 2025–2028 period.

Microsoft's Azure AI infrastructure, built around the H100 and now GB200 platform, has been among the most aggressive in deploying cold plate DLC at scale. Their partnership with NVIDIA on GB200 NVL72 configurations has essentially mandated DLC-ready white space in every new AI cluster build. Azure's capex guidance for 2025–2026 implies continued accelerated data center construction spend, with a higher DLC content per square foot than any prior generation.

Google's TPU-based infrastructure presents a partially different picture. Google custom-designs its AI accelerators and the thermal management systems around them, which means they are less constrained by NVIDIA's GB200 form factor. However, Google's GPU fleet for third-party AI workloads (Google Cloud A3/A3+ instances) is NVIDIA-based and faces the same thermal density challenges as Azure. Google has publicly discussed DLC adoption for its next-generation AI campuses.

Meta's AI infrastructure build — heavily focused on NVIDIA H100 and now GB200 clusters for Llama model training — has disclosed facility-level capex of $35–40B for 2025 alone. A meaningful fraction of that represents DLC-ready facility upgrades and new builds. Meta's disclosed plans for a dedicated 1GW+ AI campus represent a greenfield opportunity for immersion and cold plate DLC at a scale that would anchor supplier revenue for multiple years.

Amazon AWS and Oracle's OCI have followed similar trajectories, with OCI notably aggressive in building out NVIDIA-optimized cluster infrastructure as a differentiator. The 2–3 year lead time for DLC-ready facility construction creates a pipeline of demand that is visible to supply chain vendors 18–24 months in advance — a more predictable revenue profile than typical enterprise IT equipment cycles.

Investment Implications

The investment case for DLC infrastructure divides into two sub-theses: diversified exposure to the AI buildout through thermal management leaders, and concentrated optionality through pure-play DLC specialists. The risk/reward profiles are meaningfully different.

Diversified Thermal Management Leaders

Vertiv Holdings (NYSE: VRT) is the most accessible large-cap proxy for DLC adoption. Vertiv's product portfolio spans power distribution, thermal management, and IT infrastructure management — all of which benefit from the shift to higher-density AI compute. Their CDU and integrated cooling product lines are qualified across multiple hyperscalers. The valuation has re-rated significantly as the market recognized the AI tailwind; forward multiples reflect high expectations, but the visibility of hyperscaler capex commitment provides unusual earnings predictability for a capital equipment company.

nVent Electric (NYSE: NVT) offers broader industrial diversification with a growing data center segment. Their SCHROFF and Hoffman enclosure brands have a long history in high-density compute, and their growing liquid cooling portfolio addresses the rack-level integration layer. Lower AI-specific revenue concentration than Vertiv, but also lower multiple risk if hyperscaler capex were to slow.

Eaton Corporation (NYSE: ETN) benefits indirectly through its power distribution and UPS businesses: every DLC installation requires upgraded power management infrastructure, and the GB200's 120kW rack density demands precision power delivery systems that Eaton's data center power segment is positioned to supply. Eaton is a lower-beta, higher-margin AI infrastructure play — not a DLC pure-play, but a structural beneficiary.

Pure-Play Specialists

The private players — Motivair, CoolIT, ZutaCore, Iceotope — represent pre-IPO optionality that is accessible only through venture or crossover funds currently. However, their valuation benchmarks will be set by how Vertiv and nVent trade as the sector matures. Any IPO activity in 2026–2027 should be evaluated against the established public company comps.

Asetek provides the most accessible listed pure-play, but at a small-cap valuation with material execution risk. Their design wins at hyperscale customers are commercially meaningful, but revenue concentration and capital requirements for scaling create binary outcomes that larger-cap investors may find unattractive. Position sizing accordingly.

Risk Factors

The retrofit timeline risk deserves particular attention. Data center retrofits involve not just installing new cooling hardware but also upgrading electrical systems (higher kW per rack requires upgraded PDUs and power distribution), reinforcing floor loading for immersion tanks where applicable, and integrating leak detection systems that hyperscalers require before going live on liquid cooling. Each of these workstreams has independent permitting and contractor scheduling requirements. In tight construction labor markets, these timelines can extend well beyond initial plans. Investors in pure-play DLC vendors should monitor order-to-revenue conversion cycles closely in 2026 earnings reports as a leading indicator of timeline execution.

The water usage concern is a real but frequently overstated ESG risk for DLC. Closed-loop DLC systems are not consumptive of water in the way evaporative cooling towers are — the coolant circulates in a sealed loop without evaporative loss. The primary water usage in a DLC facility is in the chilled water plant's cooling towers, which consume roughly 3–5 liters of water per kWh of heat rejected. At 120kW per rack and a 1.15 PUE, a 1,000-rack AI cluster would consume on the order of 400,000–500,000 gallons per day — material, but comparable to a medium-sized industrial facility. The ESG framing is manageable; the permitting risk in water-scarce jurisdictions (parts of the US Southwest, Northern Africa, and the Middle East) is real and should inform site selection analysis for greenfield AI campuses.