1.1. Research Background

Recent advances in generative artificial intelligence (AI) have brought significant changes to the data center industry. Large language models (LLMs), such as ChatGPT, Gemini, and Claude, require thousands of GPU servers for training and inference, rapidly increasing power consumption and cooling demand in AI data centers. Conventional data centers were designed for CPUbased workloads and typically operate at a power density of 5–15kW per rack, whereas AI data centers require 50–120kW per rack, and some stateof-the-art GPU racks have been reported to consume more than 150kW. Maintaining appropriate thermal conditions for IT equipment becomes more critical as rack power density and heat generation increase^[1], and conventional air cooling alone is insufficient under such conditions. Ensuring the reliability and efficiency of the power infrastructure has consequently become an equally critical challenge. A substantial portion of a data center's operational cost is associated with power supply and cooling, and decisions made at the planning stage directly affect long-term operating cost and sustainability. Efficient operation of AI data centers therefore requires an integrated Mechanical, Electrical, and Plumbing (MEP) design approach that simultaneously considers cooling, electrical, and piping systems.

1.2. Research Objectives

This study investigates the design characteristics of MEP systems applied to AI data centers and evaluates their impacts on energy efficiency, operational reliability, Total Cost of Ownership (TCO), and sustainability. It comparatively evaluates air, liquid, and hybrid cooling and UPSbased power supply architectures (N, N+1, and 2N). Energy efficiency is quantified through Power Usage Effectiveness (PUE); reliability through UPS redundancy architectures; economic feasibility through TCO considering both capital expenditure (CAPEX) and operational expenditure (OPEX); and environmental impact through Water Usage Effectiveness (WUE) and Carbon Usage Effectiveness (CUE). Based on these analyses, the study develops and evaluates a quantitatively grounded MEP design framework that links rack power density, cooling technology, UPS redundancy, and PUE/WUE/CUE targets, and provides practical guidelines for improving energy efficiency, reliability, economic performance, and sustainability in future high-density computing environments.

2.1. Characteristics of AI Data Centers

AI data centers are built around large-scale Graphics Processing Unit (GPU) servers and exhibit substantially higher power density and heat generation than CPU-based facilities. Modern GPUs typically require more than 700 W per chip, multiGPU servers may consume several kilowatts, and rack-level density has risen from approximately 5–15kW to 50–120kW, with some ultra-high-density deployments exceeding 150kW per rack. AI data centers are also more susceptible to localized overheating (hot spots), which can cause hardware degradation, reduced computational performance, and system failures. Because AI services typically operate continuously on a 24/7 basis, highly reliable thermal management and power supply are of paramount importance, which calls for an integrated MEP design approach.

2.2. Components of MEP Systems

MEP systems consist of mechanical, electrical, and plumbing systems whose design quality directly influences data center performance and energy efficiency; their importance is greater in AI data centers because of higher power densities and thermal loads.

(1) Mechanical Systems

Remove heat generated by IT equipment and maintain appropriate environmental conditions. Representative components are chillers, Computer Room Air Handlers (CRAHs), cooling towers, and Coolant Distribution Units (CDUs). As liquid cooling spreads, CDUs have gained importance: a CDU regulates coolant circulation and heat exchange between servers and the cooling system and is a key component of Direct-to-Chip (D2C) cooling.

(2) Electrical Systems

Provide reliable and continuous power through transformers, Uninterruptible Power Supply (UPS) systems, generators, and Power Distribution Units (PDUs). Because AI data centers consume far more power than conventional facilities, N+1 or 2N redundancy architectures are commonly adopted to enhance reliability.

(3) Plumbing Systems

Produce, circulate, and distribute cooling water; major components are the cooling water system, secondary loop, and distribution piping. D2C cooling, which delivers coolant directly to CPUs and GPUs, demands more sophisticated piping than conventional facilities to ensure stable delivery and prevent leakage.

2.3. Trends in Cooling Technologies

Cooling technologies have evolved with the increasing heat generation of IT equipment, and the growth of AI data centers has exposed the limits of conventional air cooling, driving the adoption of liquid cooling.

(1) Air Cooling

The most widely used method, supplies cooled air through CRAHs or Computer Room Air Conditioners (CRACs). It offers low installation cost and simple maintenance, but its efficiency decreases as rack density increases; above 50kW/rack it often cannot provide adequate capacity and increases the likelihood of hot spots. Cho et al. classified data center HVAC and cooling systems and proposed corresponding performance test methods, providing a domestic reference framework for evaluating air-cooling capacity limits in high-density environments^[6].

(2) Liquid Cooling

Exploits the high specific heat capacity of water or other fluids and is well suited for high-density computing. Representative technologies are D2C cooling, Rear Door Heat Exchangers (RDHx), and immersion cooling. D2C has emerged as an important solution for modern AI data centers because it removes heat directly from high-power CPU and GPU components; RDHx uses heat exchangers on rack rear doors, and immersion cooling submerges servers in dielectric fluids. Cho examined rack-based cooling for high-density hyperscale data centers and reported that D2C-type rack cooling maintains stable chip-level temperatures even beyond 50 kW/rack, consistent with the density thresholds adopted in this study^[7]. Song and Hwang reviewed immersion cooling and noted its advantages in eliminating hot spots above 100 kW/rack, but also identified higher fluid and maintenance costs as a barrier to widespread adoption^[10].

(3) Hybrid Cooling

Applies conventional air cooling to standard server areas and liquid cooling to high-density GPU zones, combining the costeffectiveness of air cooling with the thermal performance of liquid cooling. It can be implemented relatively easily when upgrading existing data centers, and it is consistent with Cho's discussion of zone-based hybrid cooling as a transitional strategy for facilities migrating from CPU-based to GPU-based workloads^[8].

2.4. Sustainability Performance Metrics

Sustainability evaluation has expanded beyond energy efficiency to broader environmental impacts, and PUE, WUE, and CUE are the most widely used indicators. Their definitions, units, and typical ranges are summarized in Table 1 and applied quantitatively in Section 4.4. PUE, the ratio of total facility power consumption to IT equipment power consumption, is the most widely recognized efficiency metric; values closer to 1.0 indicate higher efficiency. WUE measures the water consumed during operation and has become more important as liquid cooling spreads. Byeon et al. compared electric-chiller and absorption-chiller cooling sources for a data center within a mandatory district energy supply zone in Korea, concluding that electric chillers were more cost-effective and reliable and noting that the absorption option was incompatible with high-temperature liquid cooling^[9]. CUE evaluates the carbon emissions associated with operation, calculated from electricity consumption and the carbon emission factor of the energy source, and has become increasingly important with ESG initiatives and carbon-neutrality policies.

Note: Typical ranges are compiled from the literature reviewed in this study (Section 3.1.1) as representative reference values and should not be interpreted as measured performance from a specific facility.

3.1. Research Framework

This study is based on a literature review and comparative scenario analysis. Domestic and international publications, technical reports, industry standards, and design guidelines were reviewed to identify the characteristics of AI data centers and the components of MEP systems, and the design characteristics of cooling, electrical, and plumbing systems were then comparatively analyzed. The research procedure, developed by the authors on the basis of Section 2, consists of six sequential stages: (1) literature review and selection, (2) MEP system analysis, (3) energy efficiency evaluation, (4) reliability assessment, (5) TCO analysis, and (6) sustainability evaluation (Fig. 1). Three cooling configurations (air, liquid, hybrid) and three UPS redundancy architectures (N, N+1, 2N) are evaluated using PUE (energy efficiency), availability and expected downtime (reliability), TCO incorporating CAPEX and OPEX (economic performance), and WUE and CUE (environmental performance). The quantitative analyses in Sections 4.1–4.4 and 4.6 are based on a common 10 MW reference facility and the scenario assumptions in Section 3.2; the numerical results are therefore comparative scenario estimates rather than measured performance of an actual AI data center.

Fig. 1. Research framework

3.2. Research Conditions

A hypothetical 10MW AI data center was considered as the reference case: a high-density facility composed primarily of GPU-based AI servers, with 100 racks at an average of 100kW per rack, operating continuously (24 hours per day, 365 days per year; 8,760h/year). The 10MW IT load represents a mid-to-large-scale AI data center building block, a scale commonly referenced in capacity-planning discussions such as the IEA Energy and AI report^[3] and Schneider Electric's AI data center guidance^[4], rather than hyperscale campuses (hundreds of MW) or small edge deployments; it also keeps the CAPEX/OPEX and reliability calculations of Sections 4.2–4.3 transparent and reproducible. The 100kW/rack density is a representative “mainstream” value within the 50–120kW/rack range of Section 2.1, corresponding to the >50–100kW/rack tier of Table 11, and was deliberately chosen away from legacyadjacent (below 30kW/rack) and ultra-high-density (above 150kW/rack) extremes; rack densities above 150kW/rack are outside the scope of this study. The IT load was held constant while only the cooling and power configurations were varied, so that performance differences are attributable solely to MEP design strategies.

The all-air-cooled scenario is a comparative baseline rather than a universal recommendation for 100kW/rack: its thermal feasibility depends on server configuration, airflow management, containment, supply-air conditions, and supplemental heat removal, and no rack-level thermal simulation was performed. It should be read as a scenario benchmark for facility-level energy assumptions, not as a validated design recommendation

Note: The zonal composition is an author-defined reference scenario and a modeling assumption, not a measured rack distribution from an operating facility. Zone A is air- cooled; Zone B serves 0.5MW of its 1.5MW IT load with D2C liquid cooling and Zone C 2.5MW of its 4.0MW; Zones D and E are fully D2C-cooled. The liquid-cooled IT load is 7.0MW, i.e., 70% of the total IT load ($f = 0.70$); the remaining 3.0MW is air-cooled.

3.3. Target Systems

(1) Mechanical System

Three cooling technologies were compared: CRAH-based air cooling, D2C liquid cooling, and hybrid cooling (air cooling for standard server areas and D2C for highdensity GPU zones). These architectures represent the dominant approaches documented in Section 2.3. CRAH-based air cooling is the most widely deployed baseline owing to its low installation cost and mature maintenance practices^[6]. D2C was selected over RDHx and immersion cooling because it is reported to maintain stable chip-level temperatures beyond 50kW/rack^[7] and has emerged as the leading industry solution for GPUintensive workloads in the density range considered^[4]. RDHx is treated as a sub-option in the aircooling-adjacent tier of Table 11, while immersion cooling, although reported to offer additional hotspot elimination above 100 kW/rack, was excluded from the primary comparison because of its higher fluid and maintenance costs and lower current adoption^[10]. The hybrid configuration represents the zone-based transitional strategy for facilities migrating from CPU-based to GPU-based workloads^[8].

(2) Electrical System

The study focuses on the transformer, UPS, and PDU; because the UPS architecture directly affects reliability, the N, N+1, and 2N configurations are comparatively evaluated.

(3) Plumbing System

The study focuses on the Cooling Water System and the Secondary Loop, the major components of modern liquid-cooling infrastructure.

3.4. Performance Evaluation Metrics

Four evaluation criteria were adopted: energy efficiency (PUE), reliability (availability, considering the UPS redundancy level and the ability to maintain continuous service during equipment failures), economic feasibility (TCO, the sum of CAPEX and OPEX over the system lifecycle), and sustainability (WUE and CUE). Their definitions are given in Section 2.4 and Table 1.

3.5. Scope and Limitations

The study is based on a literature review and practical considerations from the design and construction of AI data centers, while the quantitative analysis relies on literature-informed scenario assumptions and predefined representative operating conditions. Owing to security and confidentiality constraints, operational and equipment performance data from specific projects could not be disclosed or included. The study focuses on cooling and power systems; regional electricity prices, climatic conditions, manufacturer differences, and project-specific operational characteristics were not fully considered. In the Korean context, an industrial electricity price of USD 0.10/kWh was adopted as an illustrative assumption, and the 2023 Korean consumption-side electricity emission factor of 0.4173tCO₂eq/MWh (0.4173kgCO₂eq/kWh) reported by the Korea Energy Agency's EG-TIPS was applied consistently to the total facility electricity consumption when calculating CUE and annual carbon emissions. Because seasonal temperature and humidity influence free-cooling availability and heat-rejection efficiency, and local water availability affects the feasibility of evaporative cooling, these effects were considered qualitatively, as no site-specific hourly weather or water-resource data were used. None of the quantitative results reported in Sections 4.1–4.4 and 4.6 were validated against metered or measured operating data from an actual operating AI data center. All PUE, availability, TCO, WUE, and CUE figures are scenario-based engineering estimates derived from the conditions in Section 3.2 and the literature ranges identified in Section 3.1.1; their absolute magnitudes, including monetary and carbon-emission values, should not be interpreted as project-specific forecasts but as an internally consistent basis for comparing the relative performance of the evaluated configurations. Future research should validate the framework with operational data from actual AI data centers and digital-twin-based simulations.

4.1. Energy Efficiency Analysis

The cooling system is among the most critical factors influencing the power consumption of AI data centers. In the air-cooling scenario, non-IT infrastructure energy accounts for 31.0% of total facility energy consumption; this value combines cooling systems, UPS systems, transformers, power distribution equipment, and other auxiliary infrastructure and should not be interpreted as the proportion attributable to cooling equipment alone. In high-density AI data centers, air cooling requires additional air-handling equipment and fan power, which increases energy consumption, whereas liquid cooling, in particular D2C, exploits the high specific heat capacity of water to remove heat more efficiently and reduce cooling power, improving PUE. Hybrid cooling applies air cooling to standard areas and liquid cooling to GPU-intensive zones, combining the economic advantage of the former with the thermal performance of the latter.

Note: The PUE values are scenario-based engineering assumptions for comparative analysis, not measured values from a specific operating facility. The hybrid value is derived from the IT-loadweighted average of the air- and liquid-cooling values using a liquidcooling fraction of 70%.

The air-cooling PUE (1.45) is a scenario assumption for a purpose-built CRAH-based cooling architecture, set below the 1.56 global average PUE reported in the Uptime Institute Global Data Center Survey 2024 (Fig. 2, p. 8)^[2]. The liquid-cooling PUE (1.20) is a scenario assumption for a highefficiency D2C configuration; although the literature supports D2C liquid cooling for high-density AI workloads and its potential to reduce cooling-related power consumption^{[4, 7]}, 1.20 is not a directly measured or universally reported D2C PUE and was selected as the baseline for the comparative analysis. For the hybrid scenario, the liquid- cooled fraction of the IT load was calculated from the zonal composition in Table 3 rather than back-calculated from a target PUE: 7.0MW of the 10.0MW IT load is D2C-cooled ($f = 0.70$) and 3.0MW is air-cooled. The 100 kW/rack value in Table 3 is the facilityaverage density, whereas the hybrid scenario uses the mixed-density hall in which the higher-density racks are predominantly liquid- cooled. The hybrid PUE was calculated as the IT-load- weighted average $\text{PUE}_{\text{hybrid}} = \text{PUE}_{\text{air}} \times (1 - f) + \text{PUE}_{\text{liquid}} \times f = 1.45 \times 0.30 + 1.20 \times 0.70 = 1.275$, which assumes that shared cooling plant, UPS, transformer, and distribution losses are embedded in the scenario PUE values and scale with IT load, with no separate shared-infrastructure term. The same weighting gives $\text{WUE}_{\text{hybrid}} = 1.8 \times 0.30 + 0.5 \times 0.70 = 0.89\text{L/kWh(IT)}$. Because this linear weighting does not account for shared infrastructure, auxiliary cooling loads, or residual heat released from D2C cold plates to room air, it is a simplified comparative model, and the 70% fraction is an author-defined assumption rather than a measured or literaturereported industry ratio; its robustness is tested for 50–90% in Section 4.6. As shown in Table 4, liquid cooling has the lowest assumed PUE, and hybrid cooling is approximately 12.1% lower than air cooling. For the 10MW / 8,760h reference facility, $\text{Total Annual Energy} = \text{IT Load} \times \text{PUE} \times \text{Operating Hours}$ and $\text{Non-IT Overhead Energy} = \text{Total Annual Energy} - \text{IT Annual Energy}$, where IT Annual Energy is 87,600MWh/year in all cases (Table 5). The non-IT overhead includes cooling-related energy as well as UPS, transformer, power-distribution, and other auxiliary losses and is not interpreted as cooling-only energy.

Liquid cooling reduces annual non-IT overhead energy by approximately 55.6% and hybrid cooling by approximately 38.9% relative to air cooling, providing a quantitative basis for the PUE comparison in Table 4. The PUE values of 1.45 and 1.20 remain author-defined baseline scenario inputs informed by the literature, not directly transferable measured values.

4.2. Reliability Analysis

AI data centers support mission-critical services such as financial systems, cloud computing, and generative AI, so the reliability of the power supply is a primary design consideration. The N configuration uses only the minimum required equipment (lowest capital cost), and a UPS failure may interrupt the entire service. N+1 adds one standby UPS that immediately assumes the load upon a failure. 2N completely duplicates the power path, giving the highest reliability and availability at significantly higher capital and maintenance cost. The qualitative differences are compared in Table 6.

To quantify these ratings, the availability values in Table 7 (99.671%, 99.982%, and 99.995%) were set as scenario assumptions that preserve the generally recognized ordering N < N+1 < 2N ^{[2, 5]}. They do not represent manufacturer-specific or fieldmeasured values, because actual availability depends on component failure rates, maintenance procedures, switching performance, repair logistics, common-cause failures, and system topology. The MTTR was fixed at 4 hours as a comparative repairtime assumption. The equivalent system-level MTBF was derived from $A = \text{MTBF} / (\text{MTBF} + \text{MTTR})$, i.e., $\text{MTBF} = A \times \text{MTTR} / (1 - A)$, and annual expected downtime was calculated as $8,760 \times (1 - A)$. For the N configuration, $\text{downtime} = 8,760 \times (1 - 0.99671) \approx 28.8\text{ h/year}$ and $\text{MTBF} = 0.99671 \times 4 / (1 - 0.99671) \approx 1,212\text{ h}$; the same procedure gives the N+1 and 2N values.

Note: The values are derived from the scenario assumptions described in this study. The N, N+1, and 2N configurations represent UPS redundancy architectures and should not be interpreted as being directly equivalent to Uptime Institute Tier classifications.

The analysis is an illustrative comparison based on assumed system-level availability and a fixed MTTR. A component-level reliability model (UPS module ratings and quantities, failure and repair rates, static bypass and transfer behavior, battery failures, common-cause failures, and independence of distribution paths) was not developed because the required manufacturer-specific and operational data were unavailable; the derived MTBF values are therefore equivalent system-level values and do not independently validate the reliability improvement of each architecture. The N+1 configuration reduces expected annual downtime by approximately 94.4% relative to N while requiring substantially less redundant infrastructure than 2N. Under the assumed conditions, N+1 thus provides a balanced relationship between reliability and redundancy cost and may serve as a practical baseline architecture for medium-and large-scale AI data centers, whereas 2N, which provides a further reduction in expected downtime, may be considered for highly critical zones. The final selection should depend on the required availability, load criticality, additional UPS and distribution CAPEX, efficiency-related losses, maintenance requirements, allowable interruption time, the economic loss caused by service interruption, and project-specific cost constraints; this comparison is a reliability-oriented reference rather than a lifecycle cost optimization.

4.3. Total Cost of Ownership (TCO) Analysis

TCO, the sum of CAPEX and OPEX over the lifecycle, is a widely used indicator of the economic feasibility of design alternatives. Air cooling requires the lowest initial investment but incurs higher operating costs through higher energy consumption. Liquid cooling requires additional components (CDUs, piping networks, heat exchangers) and a higher initial investment, which is offset by lower electricity consumption. Hybrid cooling offers a moderate initial investment together with reduced operating costs. A scenario-based cost model was constructed for the 10 MW reference facility using representative engineering cost assumptions established by the authors: CAPEX of USD 6.0 / 7.2 / 6.6 million per MW for air, liquid, and hybrid cooling (liquid 20% and hybrid 10% above the air-cooling baseline); an industrial electricity rate of USD 0.10/kWh; annual maintenance of 2% of CAPEX; a water cost of USD 2.0/m³; and a mid-life replacement cost of 15% of initial CAPEX. The TCO for an evaluation period of $n$ years is $\text{TCO}_n = \text{CAPEX} + (\text{Annual OPEX} \times n) + \text{Replacement Cost}$, $\text{Replacement Cost} = 0.15 \times \text{CAPEX}$, applied identically to the 10-year and 15-year periods. The cost boundary covers cooling-related facility infrastructure and associated operating costs; IT equipment and building-shell construction are excluded. The CAPEX values are comparative scenario inputs, not vendor quotations or projectspecific estimates, since actual CAPEX varies substantially with project scope, location, cooling architecture, and construction conditions. Annual OPEX includes electricity, maintenance, and water costs. The replacement allowance is a simplified mid-life provision included once in each analysis (nominally at about Year 5 in the 10-year and Year 8 in the 15-year evaluation); because the base case applies no discount rate, inflation, financing cost, tax, or residual value, its timing does not affect the cumulative nominal TCO. The maintenance and replacement percentages are simplified lifecycle parameters, not industry standards. The hybrid CAPEX of USD 6.6 million/MW is below the USD 6.84 million/MW obtained by linear interpolation at a 70% liquid-cooling fraction; this simplification is examined in Section 4.6.

Note: CAPEX, electricity price, maintenance cost, water cost, and replacement cost are scenario-based engineering assumptions established by the authors for comparative analysis and are not vendor quotations or project-specific costs.

Despite its higher CAPEX, liquid cooling yields the lowest 15-year TCO, approximately 6.7% below air cooling, owing to its lower energy and water costs. With the base-case hybrid CAPEX, hybrid cooling achieves a slightly lower 10-year TCO than liquid cooling (USD 202.3 million versus USD 203.2 million) while requiring approximately 8% less initial CAPEX; this ranking depends on the assumed hybrid CAPEX (Section 4.6). The economically preferable configuration therefore varies with the evaluation period and investment constraints: liquid cooling is lowest over 15 years, whereas hybrid cooling provides a lower initial investment with comparable 10-year TCO, which is attractive where up-front budget constraints or phased deployment are important.

4.4. Sustainability Analysis

ESG management and carbon-neutrality initiatives have made sustainability a fundamental consideration in the design of AI data centers. PUE, WUE, and CUE were used to evaluate energy efficiency, water consumption efficiency, and carbon emissions, respectively. Liquid cooling does not inherently result in higher water consumption, which depends strongly on the heat-rejection architecture and the use of evaporative cooling. In the D2C configuration assumed here, the lower WUE is attributed to the assumed heatrejection configuration and reduced dependence on evaporative cooling. The WUE values of the air, liquid, and hybrid configurations are assumed to be 1.8, 0.5, and 0.89L/kWh(IT) (Table 10), giving annual water consumption of 157,680, 43,800, and 77,964m³ from $\text{Annual Water Use (m³)} = \text{WUE} \times \text{IT Annual Energy (kWh)} / 1,000$ (air cooling: $1.8 \times 87,600,000 / 1,000 = 157,680\text{ m³/year}$). CUE was calculated as $\text{Grid Emission Factor} \times \text{PUE}$ using the 0.4173kgCO₂eq/kWh factor of Section 3.5 (air cooling: $0.4173 \times 1.45 = 0.605\text{ kgCO₂eq/kWh(IT)}$), and annual emissions as $\text{CUE} \times \text{IT Annual Energy (kWh)} / 1,000$ (air cooling: $\approx 53,005\text{ tCO₂eq/year}$). The same procedures were applied to the liquid- and hybrid-cooling scenarios.

Note: The WUE values are representative scenario assumptions, not measured values from specific facilities. The air-cooling value (1.8 L/kWh(IT)) represents a CRAH/chiller plant with conventional cooling-tower evaporative heat rejection. The liquid-cooling value (0.5 L/kWh(IT)) represents a D2C scenario with warm-water supply, dry-cooler-based heat rejection, and limited adiabatic assistance, in which evaporative water consumption is substantially reduced; it is a scenario parameter rather than an inherent WUE value of D2C cooling. The hybrid value is the IT-load-weighted average ($f = 0.70$) of the two.

Actual WUE depends strongly on the heatrejection architecture, climate, operating conditions, water-reuse strategy, and cooling-water management, so the liquid- and hybrid-cooling values should be interpreted as scenario parameters rather than measured facility performance; their sensitivity is examined in Section 4.6.

4.5. Proposed AI Data Center MEP Design Framework

Fig. 2 illustrates the integrated architecture of the proposed MEP framework. The electrical path consists of the utility supply, transformer, main switchgear, UPS redundancy, and PDU, while the thermal path connects the cooling source and heatrejection system to the CRAHs and CDUs serving the IT racks. The BMS/EMS and monitoring layer collects operating information from both and provides the basis for coordinated operational control; the predictive-control layer is a conceptual extension for future optimization, not a validated control algorithm.

Fig. 2. Integrated power, cooling, and predictive control architecture for AI data centers

The domestic studies reviewed in Section 2 provide the technical context: studies on data-center HVAC and cooling-system classification support CRAH-based air cooling as a reference architecture, while rack-based cooling studies support D2C liquid cooling in high-density GPU zones ^{[6, 7]}. The present study extends these findings by evaluating the selected configurations under a common 10 MW reference condition and linking them to PUE, TCO, WUE, CUE, and UPS reliability indicators; because the boundary conditions and evaluation methods differ from those of the cited studies, this is a conceptual and methodological extension rather than a direct comparison of measured performance.

Building on the results of Sections 4.1–4.4, this study proposes an explicit, rack-density-based MEP design framework of four integrated layers: (1) a Zoned Cooling Layer, (2) a Redundant Power Layer, (3) a Plumbing/CDU Layer, and (4) an Intelligent Predictive Control Layer. Rather than applying a single uniform strategy to the entire facility, the model partitions the data hall into four rack-density tiers and assigns a matched cooling technology, UPS redundancy level, and PUE/WUE target to each tier (Table 11).

Note: The PUE and WUE values are proposed design targets derived from the comparative scenario analysis and literature-informed boundary conditions, not measured performance values, and should be adjusted for project-specific climate, load profile, heat-rejection architecture, and equipment characteristics. The UPS redundancy column shows the default candidate for each tier; the final selection follows the criticality-based criteria of Layer 2 rather than rack density alone.

Note: The values represent reference scenario results rather than measured facility performance; the N+1 values follow the scenariobased reliability assumptions in Table 7. The TCO values exclude IT equipment and building-shell construction. CUE uses the Korean consumption-side emission factor of 0.4173kgCO₂eq/kWh.

The four layers of the proposed model are defined as follows.

Layer 1 - Zoned Cooling Layer partitions the data hall into the four density tiers, keeping air cooling in lower-density zones to minimize CAPEX and concentrating D2C liquid cooling in GPUintensive zones, consistent with the hybrid findings of Sections 4.1 and 4.3.

Layer 2 - Redundant Power Layer uses N+1 UPS redundancy as the candidate baseline for general facility loads and considers 2N for missioncritical loads that require greater power-path independence or have particularly low allowable interruption times. Redundancy should not be selected solely on rack power density; service criticality, allowable downtime, fault-isolation requirements, maintenance strategy, and the economic consequences of interruption must also be considered through a project-specific reliability and cost evaluation.

Layer 3 - Plumbing/CDU Layer sizes the Secondary Loop and CDU capacity per zone according to the liquid-cooling fraction of each tier, avoiding facility-wide over-provisioning that would inflate CAPEX in lower-density zones. CDU capacity is not prescribed as a fixed value because it depends on the zonal liquid-cooled IT load, coolant supply and return temperatures, design temperature difference, flow rate, heat-exchanger performance, and redundancy requirements.

Layer 4-Intelligent Predictive Control Layer is intended as a future operational optimization framework: by integrating real-time GPU workload information with chiller, CDU, pump, and fan operating data, it could predict changes in thermal and electrical demand and support coordinated control of cooling and power infrastructure. Since no AI prediction algorithm or operational dataset was implemented, it is a conceptual extension rather than a validated AI control system. Rack densities above 150kW/rack are regarded as project-specific ultra-high-density applications requiring separate assessment, so no universal design target is prescribed. Applied to the 10MW reference facility (Table 3), the framework lets operators concentrate capital where higher cooling capacity and redundancy are required (7.0MW liquid-cooled and 3.0 MW air-cooled IT load, with N+1 as the candidate baseline and 2N for missioncritical zones) instead of applying a uniform and, for lower-density zones, unnecessarily costly specification. The main advantage of the framework is therefore not a single new technology but the coordinated allocation of existing MEP technologies according to load density, service criticality, and infrastructure requirements, explicitly linking rack density, cooling technology, UPS redundancy, and plumbing capacity.

4.6. Sensitivity Analysis

Because several key inputs are scenario assumptions, one-at-a-time sensitivity analyses were performed. The base-case 15-year TCO was USD 282.3, 263.4, and 265.6 million for air, liquid, and hybrid cooling (10-year: 211.2, 203.2, and 202.3 million). The tested ranges were CAPEX ±20%; liquid-cooling CAPEX premium 0–40% (hybrid at half of that premium); maintenance 1–3% and replacement 10–20% of CAPEX; electricity price 0.05–0.15USD/kWh; IT utilization 50–100% of the reference IT load; and discount rate 0–8% (Table 13).

Cost parameters and IT utilization. Under all tested CAPEX, maintenance, replacement, and electricity-price assumptions, the 15-year TCO of liquid and hybrid cooling remained below that of air cooling, but the ranking between them depended on the assumptions. When the hybrid CAPEX was raised to the linearly interpolated USD 68.4 million, the hybrid TCO rose to USD 269.1 million and remained above that of liquid cooling, showing that the liquid-versus-hybrid ranking is sensitive to the assumed hybrid CAPEX. In the IT-utilization analysis, electricity and water consumption varied proportionally while CAPEX, maintenance, replacement, and PUE were held constant; the TCO gap between air and liquid/hybrid cooling widened as utilization increased because energy-related costs grew with IT load. Hybrid cooling had the lowest 15-year TCO at 50% and 75% utilization (USD 180.7 and 223.2 million versus 183.9 and 223.7 million for liquid cooling), whereas liquid cooling was lowest at 100% (USD 263.4 versus 265.6 million). Because PUE was held constant, the low-utilization results likely understate the energy penalty at partial load.

Hybrid liquid-cooling fraction. Varying the liquid-cooling fraction from 50% to 90% gave hybrid PUE of 1.325, 1.300, 1.275, 1.250, and 1.225; WUE of 1.15, 1.02, 0.89, 0.76, and 0.63 L/kWh(IT); CAPEX (varied with the fraction) of USD 63.6, 64.8, 66.0, 67.2, and 68.4 million; and 15-year TCO of USD 269.3, 267.5, 265.6, 263.7, and 261.8 million, respectively. Increasing the fraction generally reduces PUE, WUE, and 15-year TCO, so the 70% base case is a representative scenario condition rather than a universally applicable industry ratio.

PUE. Varying the liquid-cooling PUE from 1.10 to 1.30 (70% hybrid fraction) gave hybrid PUE values of 1.205, 1.240, 1.275, 1.310, and 1.345; the 15-year TCO of liquid cooling was USD 250.3, 256.8, 263.4, 270.0, and 276.5 million, and that of hybrid cooling USD 256.4, 261.0, 265.6, 270.2, and 274.8 million. When the air-cooling PUE was varied from 1.35 to 1.56, the hybrid PUE was 1.245, 1.275, and 1.308, the air-cooling TCO USD 269.1, 282.3, and 296.7 million, and the hybrid TCO USD 261.6, 265.6, and 269.9 million (liquid cooling remained at USD 263.4 million). Liquid and hybrid cooling stayed below air cooling over the tested PUE ranges, whereas the ranking between liquid and hybrid cooling changed as the liquid-cooling PUE increased or the air-cooling PUE decreased.

Discount rate. For the discounted analysis, each annual OPEX (end-of-year payment) and the replacement allowance (year-8 payment) were converted to present values. Liquid cooling had the lowest TCO at 3% (USD 224.3 versus 224.7 million), whereas hybrid cooling was slightly lower at 5% and 8% (USD 203.9 versus 204.3 million; USD 179.5 versus 180.9 million), because discounting lowers the weight of later-year energy savings relative to the initial CAPEX. The base-case TCO should therefore be read as an undiscounted cumulative cost, with the discount-rate results as an additional perspective on lifecycle performance.

Reliability (MTTR). MTTR was varied from 2 to 8 hours while the derived MTBF values were held constant. For MTTR of 2, 4, and 8 hours, the annual expected downtime was 14.4, 28.8, and 57.5 hours for N; 0.79, 1.58, and 3.15 hours for N+1; and 0.22, 0.44, and 0.88 hours for 2N; the corresponding availability was 99.835%, 99.671%, and 99.344% for N; 99.991%, 99.982%, and 99.964% for N+1; and 99.9975%, 99.995%, and 99.990% for 2N. The ordering N < N+1 < 2N was unchanged: absolute downtime is sensitive to repair time, whereas the relative ordering is not.

WUE. WUE was varied independently (air 1.2– 2.4, liquid 0.2–0.8, hybrid 0.6–1.6 L/kWh(IT)), giving annual water consumption of 105,120–210,240 m³ for air cooling, 17,520–70,080 m³ for liquid cooling, and 52,560–140,160m³ for hybrid cooling. Because the water cost (USD 2.0/m³) is a small component of lifecycle cost, WUE had a limited effect on 15-year TCO; it influences the magnitude of water consumption more than the relative TCO. Because the liquid and hybrid ranges overlap, the water-consumption ranking is not guaranteed under all combinations.

Overall, the advantage of liquid and hybrid cooling over air cooling is maintained across the principal cost, PUE, and hybrid-fraction scenarios, whereas the liquid-versus-hybrid ranking is sensitive to the evaluation period, hybrid CAPEX, liquidcooling fraction, electricity price, cooling efficiency, IT utilization, and discount rate. The N < N+1 < 2N availability relationship is unchanged when MTTR varies. The quantitative results should therefore be interpreted within the defined scenario boundaries rather than as universally applicable performance values.