Deployments that have fully optimized for liquid cooling are reported to reach PUE levels of 1.1, a material improvement over typical industry performance

Decision Lens

The AI compute build-out is typically framed as a procurement challenge — more megawatts, faster interconnection, longer PPAs. But a significant portion of the constraint is internal. Liquid cooling technologies have demonstrated the ability to reduce cooling overhead by up to 40%, with best-in-class deployments reaching PUE levels of 1.1. The strategic question is whether infrastructure planning cycles are moving fast enough to capture that gain before the next capacity commitment is locked in.

90-Second Brief

Today, aI workloads are driving unprecedented increases in power density across data center portfolios, but a significant share of procured energy is consumed by cooling infrastructure before it reaches compute hardware. The cooling-to-compute energy ratio reported in the source, approximately 43% directed to cooling in U.S. Facilities, lacks a named primary study, so the figure should be treated as indicative rather than settled. Liquid cooling can close much of that gap, with some deployments already achieving PUE of 1.1.

What’s Actually Happening

The core mechanism is straightforward but underweighted in most energy procurement models: as AI workload density increases, thermal load per rack rises sharply, and conventional air-based cooling systems require disproportionately more energy to manage it. The result is that a larger fraction of total procured power is consumed by the cooling layer, not the compute layer.

Liquid cooling directly attacks this inefficiency. By removing heat at the component level, direct liquid-to-chip and immersion systems dissipate thermal load more effectively than air circulation, with reported reductions in cooling-related power consumption of up to 40%. Deployments that have fully optimized for liquid cooling are reported to reach PUE levels of 1.1, a material improvement over typical industry performance.

A parallel constraint is water. Large hyperscale facilities can consume up to 2.5 billion liters of water annually. Older evaporative cooling systems are the primary driver of that volume. Closed-loop systems using liquid-to-air heat exchangers offer a technically viable path to reducing water intensity, though the transition requires capital planning that intersects directly with energy infrastructure decisions. Workload virtualization adds a third dimension: eliminating idle compute reduces both direct energy draw and the thermal load that cooling systems must manage.

Why It Matters for Global Heads of Data Center Energy?

The energy head’s primary mandate is securing power at cost and scale. But if a substantial share of procured energy is diverted to cooling overhead — the source indicates this may be close to 43% in some U.S. facilities, though that figure lacks a named primary study — the effective cost of compute power is materially higher than the tariff or PPA rate suggests. For AI-optimized builds, where rack densities are already pushing conventional thermal limits, that overhead will worsen without deliberate infrastructure intervention.

This creates a direct lever for the energy function beyond supply procurement: influencing cooling technology selection. A facility that transitions to liquid cooling and reaches a PUE of 1.1 effectively unlocks stranded compute capacity from its existing interconnection position — without requiring additional grid capacity. That is a meaningful offset against the 3-to-7-year interconnection timelines constraining expansion in most markets.

Water consumption introduces a secondary constraint with site-level consequences. In jurisdictions where permits are required or where usage is subject to regulatory scrutiny, a hyperscale facility consuming billions of liters annually faces exposure that can affect permitting timelines and community agreements — both of which fall within scope for energy and infrastructure leadership. Transitioning to closed-loop cooling is therefore not only an efficiency move; it is a site risk management decision.

The Forward View

As AI inference workloads proliferate and rack densities continue to rise, facilities still running conventional air cooling will face a widening efficiency gap relative to liquid-optimized peers. The capital case for transitioning strengthens with each increment of AI density, and the competitive disadvantage of delay compounds across a multi-site portfolio.

Water constraints are likely to become a harder regulatory factor in the next planning cycle, particularly in markets where hyperscale concentration has already drawn public and regulatory attention. Energy heads should expect water usage to gain weight as a site-selection variable alongside grid interconnection timelines in certain geographies.

The circular economy framing — hardware designed for longer lifecycles, closed-loop cooling, as-a-service infrastructure procurement — also shifts how capital is deployed. Rather than large periodic refresh cycles, the emerging model implies continuous optimization with embedded sustainability metrics. This intersects directly with how Scope 2 and Scope 3 obligations are measured, and may reshape procurement frameworks and vendor agreements in the near term.

What We’re Uncertain About?

  • Liquid cooling adoption pace at portfolio scale: The efficiency gains are technically established, but the rate at which existing fleet transitions can occur — given facility lifetimes, capital constraints, and equipment lead times — is not quantified in available evidence. What would resolve this: operator-disclosed upgrade roadmaps or utility-reported load factor changes at transitioning facilities.

  • Primary attribution of the cooling energy share: The approximately 43% figure appears in the source without a named primary study or measurement year. The actual ratio varies significantly by facility vintage, climate zone, and workload mix. What would resolve this: facility-level PUE and energy use disclosures segmented by cooling technology type.

  • Regulatory trajectory on water permits: Whether water consumption triggers binding regulatory constraints — versus reputational or community risk only — depends on local frameworks that differ sharply by jurisdiction. What would resolve this: specific permitting decisions or legislative action in key data center markets such as Northern Virginia, the Netherlands, or Singapore.

  • Cost parity timeline for liquid cooling conversion: Up-front capital requirements for liquid cooling infrastructure are not addressed in the source. Whether energy savings justify accelerated conversion versus phased adoption requires facility-specific financial modeling that is not yet visible in available evidence.

One Question to Bring to Your Team

For each facility in our portfolio currently operating above a PUE of 1.4, what is the estimated energy and cost differential against a liquid-cooled equivalent at PUE 1.1 — and does that calculation change our interconnection strategy or next site commitment in that market?

Sources

  • Msn — Three ways data centers can operate more sustainably (Link)

Tags