In 2025, Vertiv launched its CoolPhase Flex two-phase direct liquid cooling system specifically designed for GPU rack densities exceeding 50 kW
The System Pressure
Cooling has historically been managed below the energy strategy level. That separation is breaking down. As AI GPU clusters push rack densities into the 30–60 kW range and above, air cooling systems — which still represent the majority of installed data center thermal infrastructure — are approaching their physical capacity limits. The consequence is not a hardware swap. It is a structural shift in how energy is consumed, distributed, and planned within your facility, at the same moment your procurement strategy is already under pressure from AI demand growth.
Market research from SNS Insider estimates the global data center cooling sector at roughly USD 20.73 billion in 2025, with projections reaching USD 83.54 billion by 2035 at a compound annual rate of approximately 15%. Forecast figures at this horizon carry inherent uncertainty, but the directional signal is clear: cooling infrastructure is moving from a commodity line item to a capital-intensive investment category, with direct implications for how energy spend is budgeted and justified.
Air conditioning and CRAC systems reportedly held around 52% of the cooling market in 2025. Liquid cooling, however, is growing at an estimated 22.5% CAGR through 2035 — more than double the rate of the broader market. That installed-base lag and growth-rate divergence define the transition pressure operators are now navigating.
The Drivers, Dependencies, and Constraints
Three forces are converging. First, GPU rack density has moved beyond what air systems can handle at acceptable energy cost. AI training clusters routinely operate above 50 kW per rack, and the cooling overhead required to manage that heat via air erodes PUE to levels increasingly difficult to defend under sustainability reporting obligations or board-level carbon commitments.
Second, infrastructure supply chains are maturing around liquid cooling at commercial scale. In 2025, Vertiv launched its CoolPhase Flex two-phase direct liquid cooling system specifically designed for GPU rack densities exceeding 50 kW. In 2024, Schneider Electric partnered with NVIDIA to develop optimized power and cooling reference architectures for GPU deployments. These are commercial system launches from major infrastructure suppliers, signaling that procurement pathways exist today.
Third, AI-driven DCIM platforms are being positioned as an energy management mechanism — not just a facilities tool — with market-cited claims of 15–25% reductions in cooling energy consumption through dynamic thermal load matching. Those figures come from market research and should be treated as planning scenarios rather than independently audited benchmarks until confirmed at operational scale in environments comparable to yours.
The binding constraint is transition sequencing. Liquid cooling requires dedicated coolant distribution units, leak-detection systems, and often structural modifications that air-cooled designs do not. Retrofitting an existing facility involves capital, downtime risk, and a procurement timeline that compresses poorly. For facilities currently mid-construction or in interconnection queues, the decision window is narrowing faster than planning cycles anticipated.
Open Dependencies
Several assumptions embedded in this market picture carry material uncertainty that energy strategy planning cannot quietly absorb.
The DCIM energy reduction range is vendor-adjacent, not independently verified. Actual outcomes will vary with workload variability, facility vintage, and baseline PUE. Treat 15–25% as a scenario bracket, not a contracted deliverable, when building the business case for cooling-layer software investment.
Geographic concentration matters operationally. North America accounts for an estimated 43.5% of global cooling revenues, with the U.S. holding the large majority of that share. That concentration reflects where AI infrastructure build-out is densest, but it also means localized pressure on transformer supply chains, water availability for cooling support systems, and utility grid infrastructure — constraints that a better cooling technology choice does not resolve.
The installed-base crossover timeline — when liquid cooling surpasses air cooling in new deployments — remains unconfirmed. No established data point anchors when retrofit economics improve sufficiently to accelerate replacement cycles. Multi-year CapEx plans should hold this transition timing as explicitly open rather than assuming a specific inflection year.
The Operating Exposure for Global Heads of Data Center Energy
Liquid cooling changes your facility’s load profile in ways that flow directly into your tariff structure, demand charge exposure, and behind-the-meter storage calculus. Liquid-cooled AI facilities tend to sustain higher density loads with less diurnal variation than air-cooled general-purpose infrastructure. That profile shift affects how utilities model grid impact and structure demand charges — a variable worth quantifying before the next interconnection application or tariff renegotiation.
More immediately: if your AI compute build-out is sized on the assumption that air cooling remains viable at current and near-term rack density projections, the risk is stranded or underperforming capacity. Facilities designed around air cooling limits may not absorb the next GPU generation without re-engineering that carries energy infrastructure dependencies — substation headroom, coolant distribution power loads, backup generation sizing — that sit inside your mandate.
The DCIM software boundary is also shifting. As AI-driven platforms claim influence over cooling energy consumption, the operational divide between energy procurement and facilities management is blurring. Clarifying budget ownership and performance accountability for that optimization layer before systems are procured avoids both attribution disputes and missed efficiency capture.
Signals the System Is Shifting
Three indicators would confirm this transition is outrunning current planning assumptions. First, if major colocation providers begin listing liquid cooling availability as a standard specification rather than a premium option, the installed-base gap is closing materially — and your facility RFP criteria need to reflect that shift before you sign the next long-term lease.
Second, watch utility interconnection requests for data center projects specifying dramatically different load profiles from prior generations. Sustained high-density loads with flatter consumption curves represent a new demand signature, and how ISOs and utilities respond will shape interconnection timelines and tariff design in key markets.
Third, monitor coolant distribution equipment and thermal management component lead times. If these approach the 2–3 year range already documented for large power transformers, the constraint shifts from technology readiness to supply chain — and the sequencing decision needs to move earlier in your capital planning cycle than current processes assume.
Sources
- Snsinsider — Top Data Centre Cooling Companies for AI Infrastructure (Link)
