In some climates, the closed-loop system requires evaporative or chiller-assisted cooling for only roughly one percent of the year
Decision Focus
Nvidia announced the Rubin generation of AI infrastructure on June 22, 2026, describing it as the world’s first platform to achieve 100% liquid cooling, with every chip and networking component running in a closed loop that contains no fans. The operational signal is direct: any data center operator deploying Rubin-based compute is no longer choosing between air and liquid cooling architectures. The choice has been made upstream, and the infrastructure requirements that follow — dry coolers, closed-loop plumbing, revised mechanical plant sizing — land on energy and facilities teams now, not at the end of a design cycle.
90-Second Brief
In recent days, nvidia’s Rubin platform circulates coolant at up to 45°C through cold plates directly on processors, eliminating evaporative cooling towers in all but the most extreme conditions. According to Nvidia, this cuts facility cooling water consumption from roughly 2.6 million gallons per megawatt per year under conventional cooling-tower systems to near zero in favorable climates. In some climates, the closed-loop system requires evaporative or chiller-assisted cooling for only roughly one percent of the year. Operators currently scoping AI infrastructure buildouts must determine whether their mechanical plant, water rights, and utility interconnections align with this new baseline, or require redesign before construction begins.
What Is Really Happening?
The underlying pressure is chip power density. As AI accelerators have grown denser, the thermal loads they generate crossed a point where air movement cannot carry heat away fast enough without consuming significant electrical power in the process. Liquid cooling has been available for years, but partial implementations left networking components and auxiliary systems air-cooled, creating hybrid thermal management problems that delivered only partial efficiency gains.
Rubin’s distinction is the completeness of the transition. The 45°C coolant ceiling is the mechanism that makes dry-cooler operation viable across most of the year. Higher coolant temperatures mean heat exchange between the closed loop and outdoor air can occur without mechanical refrigeration, since the delta between process fluid and ambient air remains large enough for passive or fan-assisted rejection rather than active chilling. That is not a marginal improvement over conventional chiller plant design — it is a structural removal of the largest single driver of cooling-related energy consumption in a conventional facility.
The waste heat dimension opens a second operating angle. When coolant exits the loop at temperatures high enough to be useful, residual heat can be transferred to district heating systems, commercial building HVAC, or low-temperature industrial process loads. That transforms a cost line that previously appeared only as energy consumption into a potential offset or contractual revenue stream in jurisdictions with heat-network infrastructure.
Why It Matters for Global Heads of Data Center Energy
The first consequence is mechanical plant sizing. Chiller plants represent a major fraction of the electrical capacity a data center draws from the grid and are the primary driver of water utility requirements at the point of interconnection. Removing them from standard operating conditions changes both the utility load calculation and the water supply agreements that municipalities require before issuing site approvals. At 50 MW scale, figures cited in Nvidia’s announcement put annual cooling-related energy and water cost savings at more than $4 million — a number that compounds across a multi-site portfolio and accrues to energy budget lines, not just sustainability reporting. That figure is an estimate embedded in the source announcement, not an audited result from a commissioned deployment.
The second consequence is procurement sequencing. Closed-loop liquid cooling requires different infrastructure from day one: dry coolers instead of cooling towers, glycol-mix distribution systems, and in-rack plumbing rather than perimeter air handlers. Operators who finalize mechanical specifications assuming a conventional chiller plant can be retrofitted later risk capital write-downs or commissioning delays. The evaluation window is early design, not after civil work has begun.
Third, the floor layout implications affect power density per cabinet and total facility footprint. Without hot-aisle and cold-aisle containment requirements, the spatial logic of the computer room floor changes. That shifts how operators calculate power delivery per square meter and how substation infrastructure is sized relative to IT load — a sequencing issue with direct bearing on interconnection queue timing.
Forward View
Three fronts are worth tracking. First, whether major hyperscalers confirm Rubin-based deployments in near-term earnings or construction announcements will determine how quickly 100% liquid-cooled facilities move from architecture specification to large-scale procurement. That signal matters because it will compress dry-cooler and closed-loop component lead times in the same way transformer constraints emerged from the prior wave of expansion.
Second, water rights and permitting processes in constrained markets — Northern Virginia, the Phoenix metro, Singapore — may shift if operators can credibly demonstrate near-zero process water consumption. That creates a speed advantage for early movers and may trigger revision to existing permit conditions negotiated under conventional cooling assumptions.
Third, waste heat offtake agreements represent an emerging contractual space. In European markets with established district heating infrastructure, operators running high-temperature closed-loop systems are likely to find receptive municipal counterparties faster than those still running conventional cooling towers.
What Is Still Uncertain
The performance figures originate from Nvidia’s own announcement and have not been independently verified through third-party testing at commercial scale. The near-zero water use claim applies to unspecified favorable climates; operators in Phoenix, Singapore, or the UAE should model local wet-bulb conditions against the 45°C coolant ceiling before using headline numbers in capital plans or regulatory submissions.
Whether Rubin deployments can be retrofitted into existing facilities or require ground-up mechanical redesign has not been addressed in the published material. The waste heat recovery opportunity depends on proximity to offtake infrastructure and local regulatory frameworks — conditions that are not uniform across a global portfolio.
One Question for Your Team
Which sites currently under design or early construction have mechanical plant specifications built around a conventional chiller plant — and what would it cost to revise those specifications before the Rubin procurement cycle forces the decision on an accelerated timeline?
Sources
- Ndtvprofit — ‘100% Liquid Cooling’: Nvidia Unveils Tech That Could Nearly Eliminate Data Center Water Use (Link)
