From Air to Liquid: How to Evaluate Cooling ROI in the Age of AI

Data centre teams are being asked to support denser AI infrastructure while keeping power, cooling and operating costs under control. That balance is becoming harder to maintain as newer CPUs, GPUs and accelerators increase the heat generated within each rack. The question is no longer whether liquid cooling can remove more heat than air cooling. It is when the economics justify the move, and which cooling model fits the facility, workload and growth plan. 

Transitioning from traditional air-based systems to liquid cooling introduces complex financial questions. While the engineering benefits of liquid cooling are clear in high-density environments, justifying the capital expenditure (CapEx) requires a rigorous return-on-investment (ROI) framework. 

AI workloads are changing the cooling equation 

For decades, standard data centre designs accommodated an average power density of 5-15 kW per rack. Traditional raised-floor air systems and hot/cold aisle containment handled these thermal loads efficiently. The explosion of AI and machine learning workloads has fundamentally altered this baseline, primarily due to the increased power requirements of GPUs. 

According to IDC survey data collected from a diverse group of 300 European CIOs, rack power densities are moving from historical 15 kW baselines to 100 kW or more in 2026. They are projected to reach up to 500 kW per rack by 2030. 

 This rapid escalation creates a significant operational gap: 

• The financial burden: 28% of data centre operators explicitly cite “expensive cooling overhead” as a primary operational bottleneck. 

• The ROI issue: Despite these rising overhead costs, a significant portion of operators do not actively track their granular cooling costs. 

• The hesitation gap: 40% of operators state that direct-to-chip liquid cooling is “not ready” for widespread deployment due to a perceived “lack of clear ROI”. 

• The sceptics: Another 21% maintain that direct-to-chip liquid cooling is not necessary for their current roadmaps. 

This hesitation is understandable. Many teams are not rejecting liquid cooling outright. They are struggling to justify the move because they do not yet have the cooling cost data needed to compare the cost of staying with air cooling against the cost of moving to liquid cooling. That makes measurement the starting point for any credible ROI discussion.  

The choice to use liquid cooling can be either required or a financial decision. Understanding the CapEx and operational expenditure (OpEx) associated with liquid cooling requires assessing both costs and whether the data centre is new or needs retrofitting. 

AI is bringing air to breaking point 

Air cooling becomes a challenge when a single rack’s density exceeds the 30-35 kW threshold, based on customer feedback we have received at Supermicro and extensive testing under constant high-performance workloads that keep the systems at the highest percentage of their TDP. 

Air has a low volumetric heat capacity. Moving enough air to cool a 100 kW rack requires massive, loud, energy-hungry fans that can consume a significant amount of power, if even possible. 

Liquid cooling starts to become the more economical choice around 40 kW per rack and becomes required at 100 kW+ to prevent severe hardware thermal throttling, which reduces the performance purchased. 

The threshold will vary by facility design, workload profile, utilisation and cooling infrastructure. The important point is that as rack densities rise, the question shifts from whether air cooling can technically remove enough heat to whether it can do so economically, predictably and without constraining the performance of the infrastructure. 

Know your numbers first 

You cannot build a credible business case for an infrastructure migration without establishing a precise baseline. Operating without accurate cost tracking makes it virtually impossible to justify the capital switch from air to liquid. To calculate the cost trajectory as rack density increases, data centre operators must track two core metrics: 

1. Power Usage Effectiveness (PUE)

PUE is the standard industry metric for data centre energy efficiency, calculated using the following formula: 

PUE = Total Facility Energy Delivered / Total Energy Used by IT Equipment 

While useful for the overall data centre metric, standard PUE can obscure the true cost of local air-cooling inefficiencies at the rack level. 

2. Partial PUE (pPUE)

To isolate cooling performance, operators should measure partial PUE, or pPUE. This looks strictly at the energy used by the cooling subsystem relative to the IT load.  

Because data centres are rarely homogeneous, teams should identify the specific zones or rack clusters where high-density AI systems will be deployed. The ROI case should then focus on a small set of practical inputs: current rack power density, expected future rack density, current cooling overhead, local energy cost, existing heat rejection capacity, available power, water availability, and whether the facility would otherwise require a cooling upgrade. This keeps the discussion focused on the business case: what does it cost to keep scaling air cooling, and when does RDHX, hybrid liquid cooling or direct-to-chip liquid cooling become the more practical option? 

Tracking methods and cost trajectories 

Operators can capture these metrics using Data Centre Infrastructure Management (DCIM) tools paired with smart Power Distribution Units (PDUs) and environmental sensor networks. 

As rack density scales up, the cost trajectory of air cooling is non-linear. At low densities (under 15 kW), doubling the density results in a predictable, linear increase in fan and chiller power. As racks move toward 50kW, 100kW and beyond, the cost curve becomes less predictable because facilities may need higher fan speeds, additional cooling capacity, revised airflow management, new heat rejection capacity or a move to liquid cooling. That is why the ROI case should compare the cost of staying with air cooling against the cost of moving to RDHX, hybrid liquid cooling or direct-to-chip liquid cooling. 

There is no one-size-fits-all approach to liquid cooling. Data centre managers must choose a deployment strategy that matches their physical architecture: either a retrofit path for legacy space or a new-build design with liquid-cooling infrastructure. 

There are several options available when liquid cooling is deemed critical for a data centre. 

Traditional air-cooling works best when the heat output per rack does not top 20kW to 30kW. Computer room air conditioning (CRAC) is already in most data centres. For many lower-density enterprise environments, this makes air cooling the practical default. It is familiar, widely deployed and easier to maintain. The challenge is that the economics change as racks become denser and the cost of moving enough air increases. 

Understanding the different implementation pathways 

For existing data centres, ripping out concrete floors and installing complex liquid piping loops can be cost-prohibitive. One retrofit option is a rear-door heat exchanger, or RDHX, which can support higher-density racks without requiring a broader facility redesign. It can be used in environments where the rack produces up to 50 kW of heat, depending on the facility design and cooling infrastructure. 

• How it works: Standard server racks remain air-cooled internally, but a liquid-chilled radiator door is attached to the back of the rack. The door captures up to 100% of the hot exhaust air before it enters the room, neutralising the heat source. 

• Economics: RDHX offers a much lower initial CapEx than full liquid infrastructure. It allows legacy facilities to safely support isolated high-density AI pockets (up to 40 kW-50 kW) without altering the broader room’s cooling profile. 

For new-build greenfield data centres constructed for 100 kW+ AI clusters, Direct-to-Chip (Cold Plate) cooling is the most widely adopted method for long-term reductions in Total Cost of Ownership (TCO). 

• How it works: Liquid coolant loops circulate directly through micro-channel cold plates mounted on top of the highest-heat components, such as CPUs and GPUs. This system captures up to 94% of the heat directly at the CPU or GPU level (and can also cool memory or other electronics). 

• Economics: While direct-to-chip engineering demands a higher upfront capital investment for piping, Coolant Distribution Units (CDUs), and fluid management, it delivers measurable OpEx reductions. 

Matching the cooling strategy with business realities 

An effective ROI framework for liquid cooling must extend beyond basic energy efficiency to account for broader business realities. For this, there are three main considerations: 

1. Cost reduction (OpEx): Transitioning from high-density air to liquid cooling can lower overall cooling energy consumption by up to 70%, based on rack-scale simulations and individual server testing. Liquid systems eliminate the high energy draw of server chassis fans and allow facilities to use warmer water loops, reducing or eliminating expensive chiller workloads. 
2. Sustainability and compliance: Liquid cooling dramatically improves a facility’s carbon footprint. It enables data centres to hit aggressive sustainability targets and comply with tightening environmental regulations around corporate energy consumption. 
3. Future-proofing: AI hardware lifecycles are reasonably brief. Investing in a robust, liquid-ready infrastructure ensures that the physical plant can accommodate next-generation architectures without requiring another costly facility overhaul every 3-to-5-years. 

Real-world payback scenarios 

Financial outcomes vary significantly depending on the deployment strategy: 

• The retrofit scenario: An enterprise data centre installs RDHX doors onto 20 existing racks to accommodate a new AI inference project. By capturing heat at the rack door, they avoid upgrading their entire room’s chiller system. With lower upfront capital, the energy savings and avoided costs yield a rapid payback period of 12 to 24 months, based on simulations for a typical deployment in a moderate climate zone. 

• The new-build scenario: A hyperscale colocation provider builds a dedicated AI facility using direct-to-chip liquid cooling for 1,000 high-density servers. The upfront investment for the liquid infrastructure is high. However, by reducing the facility’s PUE from an estimated 1.4 to 1.1, the operation saves hundreds of thousands of dollars annually in utility bills, a savings that can be simulated in advance. This comprehensive TCO model delivers a reliable payback period of 3-to-5-years while maximising computing density within the available real estate. 

For data centre operators, the first step is not choosing a cooling technology, but rather understanding what cooling currently costs, how those costs change as rack density rises, and which facility constraints will shape the next AI deployment. Once those numbers are clear, the decision becomes more practical: continue with air cooling where density allows, use RDHX or hybrid approaches where retrofits make sense, and consider direct-to-chip liquid cooling where higher-density AI infrastructure changes the economics.