Cost-Optimal Decarbonization Pathways for Data Centers in Japan: A Bottom-Up Model Integrating Location, Energy Systems, and Carbon Pricing

Abstract

This study develops a bottom-up cost optimization model (DC-DECOM) to evaluate decarbonization pathways for Japan’s data center industry, targeting carbon neutrality of the information and communications technology (ICT) sector by 2040. The model represents Power Usage Effectiveness (PUE) as a dynamic function of ambient temperature and cooling technology, and integrates technology selection, regional energy supply, and carbon pricing within a single cost-minimization framework. Three scenarios are compared: a reference case (REF), a centralized carbon-neutral scenario (C-CN) that restricts new capacity to metropolitan areas, and a regional decentralization scenario (R-CN) that allows for nationwide siting. Input parameters are calibrated against data from the International Energy Agency (IEA), the Uptime Institute, Japan’s Ministry of Internal Affairs and Communications (MIC) White Papers, and the Japan Science and Technology Agency (JST). The R-CN scenario achieves the 2040 net-zero target at 18–23% lower total system cost than C-CN. The cost gap decomposes into four channels (cooling-energy reduction ∼35%, lower regional renewable procurement cost ∼30%, lower carbon cost ∼25%, and lower siting-related cost ∼10%). Sensitivity analysis identifies the carbon-price trajectory and the hardware-efficiency improvement rate as the most influential parameters; the R-CN advantage remains positive across all $\pm 1\sigma$ parameter variations and across two combined-scenario stress tests.

1. Introduction

The International Energy Agency (IEA) estimated global data center electricity consumption at approximately 415 TWh in 2024, about 1.5% of total global electricity demand [1]. The IEA’s Energy and AI report (April 2025) projects this will more than double to approximately 945 TWh by 2030 in the base case, a level that exceeds Japan’s entire FY2023 electricity consumption of approximately 940 TWh [1]. In the United States, Lawrence Berkeley National Laboratory (LBNL) reported that data center electricity use rose from 58 TWh in 2014 to 176 TWh in 2023, representing 4.4% of total U.S. electricity consumption, with projections of 325–580 TWh by 2028 [2].

Japan faces a particularly acute form of this challenge. Wood Mackenzie’s 2025 analysis estimated Japan’s data center electricity consumption at 19 TWh in 2024, projecting it to triple to 57–66 TWh by 2034. This is equivalent to the annual electricity use of 15–18 million households, and the incremental demand from data centers alone accounts for roughly 60% of the total growth in Japan’s national electricity demand expected over the same period [3]. Japan’s grid monitor OCCTO projects that total national electricity demand will reach 846 TWh in FY2035, up 5.3% from 803 TWh in FY2025, with data centers and semiconductor factories identified as the primary growth drivers [4].

The Japan Science and Technology Agency (JST) has published a series of five reports (2019–2023) systematically quantifying the impact of information society advancement on energy consumption [5,6,7,8,9]. These reports collectively provide the most detailed bottom-up analysis of Japan’s information and communications technology (ICT) energy trajectory currently available. JST Vol. 1 (2019) estimated total IT-related electricity consumption at 41 TWh in 2016, and projected that without efficiency measures this would rise to 1480 TWh by 2030 and to 176,200 TWh by 2050 [5]. The latter figure is more than 187 times Japan’s FY2023 total national electricity consumption—an unrealistically large magnitude that serves to illustrate the urgency of implementing efficiency and decarbonization measures [5]. JST Vol.2 (2021) focused specifically on data center power consumption, establishing the baseline estimates used in subsequent analyses [6]. JST Vol.3 (2021) extended the analysis to network-related energy consumption, estimating total network power consumption at 23 TWh in Japan and 490 TWh worldwide in 2018, and projecting that with 27% annual traffic growth and technology levels fixed at 2018 standards, network consumption alone would reach 93 TWh by 2030 and 9000 TWh by 2050 in Japan [7]. For data centers specifically, the Vol.3 “Modest” scenario—assuming continued efficiency improvements at historical rates (CPU performance doubling every 2 years, GPU every 1.5 years, memory and network switches halving power per unit every 2 years, storage improving 10–30× per decade)—projected domestic data center consumption of 24 TWh by 2030 and 500 TWh by 2050 [7]. JST Vol.4 (2022) and Vol.5 (2023) updated these projections to account for the emergence of generative artificial intelligence (AI) [8,9].

The Ministry of Internal Affairs and Communications (MIC) has tracked the growth of Japan’s digital infrastructure across successive editions of its White Paper on Information and Communications. The 2022 edition addressed the geographic concentration risks of data centers and the growing importance of digital infrastructure resilience, noting that about 80% of Japan’s data center capacity is concentrated in the Tokyo and Kansai metropolitan areas [10]. The 2023 edition highlighted the structural transformation of ICT infrastructure driven by cloud computing and AI adoption, reporting that 70.4% of Japanese enterprises already use cloud services but only about 10% use them for system development [11]. The 2024 edition reported the rapid expansion of Japan’s data center market, projected to reach ¥3.3 trillion by 2030 [12]. The 2025 edition specifically examined the penetration of digital technologies as “social infrastructure,” including the energy implications of AI-driven data center growth [13].

However, this demand trajectory stands in direct conflict with Japan’s climate commitments. Japan’s newly updated Nationally Determined Contribution (NDC), released in 2025, sets greenhouse gas reduction targets of 46% by 2030, 60% by 2035, and 73% by 2040, all relative to 2013 levels [14]. The Green Growth Strategy further designates the semiconductor and ICT sector for carbon neutrality by 2040—a decade ahead of the national 2050 target—with specific milestones including a 30% energy efficiency improvement for all newly built data centers by 2030 [15]. On top of this, the GX Emissions Trading System (GX-ETS) became mandatory in April 2026, with an initial price floor of ¥1700/t-CO₂ and a price cap of ¥4300/t-CO₂ [16,17]. As a result, any delay in decarbonization efforts translates directly into an explicit economic pressure on the industry in the form of carbon costs in the ¥1700–4300/t-CO₂ range, leaving the data center sector caught between rising demand and tightening emission reduction requirements.

The Uptime Institute’s Global Data Center Survey 2024 reported a global average Power Usage Effectiveness (PUE) of 1.58, essentially unchanged since 2020 when it was 1.59 [18]. Data reported by the JDCC, as summarized in the Uptime Institute survey [18], indicate that in Japan the domestic average PUE is 1.7 for conventional facilities (range 1.2–2.6) and approximately 1.2 for cloud providers, with the most recent new-build facilities achieving below 1.5. In Japan, the Agency for Natural Resources and Energy, through its Data Center Benchmark System, has set a target of improving the national average PUE to 1.4 by 2030 [19,20]. Internationally, Germany’s Energy Efficiency Act (EnEfG, 2023) goes substantially further, mandating a PUE of 1.2 for new data centers starting in July 2026 and 1.3 for existing facilities by 2030 [21], illustrating the direction of regulatory tightening on data center energy efficiency.

Advanced cooling technologies offer a path to substantially lower PUE values. Mitsubishi Heavy Industries (MHI) demonstrated a PUE of 1.05 using immersion cooling in a 2023 pilot, representing a 94% reduction in cooling energy compared to conventional air-cooled systems [22]. Haghshenas et al. [23] conducted a peer-reviewed evaluation of immersion cooling in Energy Informatics, reporting about 50% reduction in total energy consumption and two-thirds reduction in occupied space relative to air cooling. Liu and Yu [24] evaluated two-phase liquid-immersion cooling optimization in Energies, and Kheirabadi and Groulx [25] provided a comprehensive design review of server cooling technologies in Applied Thermal Engineering. ASHRAE’s Liquid Cooling Guidelines for Datacom Equipment Centers [26] provide the engineering standards framework. Indeed, NVIDIA’s Blackwell architecture—with the B200 at approximately 1000 W TDP and the B300 at approximately 1200 W per chip—requires liquid cooling at rack densities of 120 kW for the GB200 NVL72 system [27], making air cooling alone physically insufficient for next-generation AI infrastructure at design rack densities. The transition from air to liquid cooling is thus not merely an efficiency improvement option but a physical necessity dictated by the trajectory of GPU power density.

The industry has already begun pursuing two complementary approaches to address data center energy challenges: leveraging geographic advantages and proactive renewable energy procurement. In Hokkaido, the White Data Center in Bibai has operated since 2014 using snow-based cooling, reducing cooling costs by 55% [28]. The Hokkaido Government’s Data Center Storage Promotion Project documents that Hokkaido is 7–10 °C cooler than mainland Japan year-round, enabling free cooling for 8–10 months annually [28]. SoftBank and IDC Frontier broke ground in May 2025 on a 300 MW AI data center in Tomakomai, Hokkaido, with a ¥65 billion investment, explicitly citing the region’s cold climate and renewable energy access as location factors [29]. Google signed its first power purchase agreements (PPAs) in Japan totaling 60 MW of solar capacity with Clean Energy Connect and Shizen Energy to power its Inzai, Chiba data center [30].

Against this backdrop of data center energy challenges and industry decarbonization trends, academic research has also increasingly addressed this domain. Masanet et al. [31] provided a seminal assessment in Science, demonstrating that efficiency gains had historically offset demand growth but warning that this decoupling was unlikely to persist. Shehabi et al. [32] developed bottom-up energy models for U.S. data centers at LBNL. Koronen et al. [33] analyzed data centers in future European energy systems in Energy Efficiency. On geographic optimization, Hao et al. [34] examined joint optimization of operational cost and carbon emissions in geo-distributed data centers in Frontiers in Energy Research. On energy system modeling for Japan, Imagawa et al. [35] developed a technology selection model incorporating carbon capture and utilization in the Journal of Japan Society of Energy and Resources, Nishikura et al. [36] analyzed distributed energy system resilience in the same journal, and Yi et al. [37] demonstrated dynamic optimal power expansion planning in IEEJ Transactions. Akimoto [38] conducted model analysis of Japan’s 2050 carbon-neutrality scenarios in the Journal of the Institute of Electrical Engineers of Japan.

Regarding Japan’s ICT sector specifically, prior scenario analyses have evaluated discrete combinations of cloud migration rates and conventional efficiency improvements, finding that only the most aggressive assumptions marginally exceeded Japan’s 2030 emission reduction target of 46% below 2013 levels. The 451 Research analysis for the Asia–Pacific region estimated that cloud migration with 100% renewable energy could achieve up to 93% carbon reduction [39], a finding validated against Deloitte Tohmatsu’s Japan-specific study [40]. Ozaki and Iwatsuki [41] examined the convergence of power and ICT network infrastructure in the IEICE Journal, Nozaki et al. [42] reviewed the direction of telecommunications energy technology to meet decarbonization needs in IEICE Transactions, and Hirose [43] surveyed ICT system energy efficiency trends in the Journal of the Japan Society of Mechanical Engineers.

However, existing studies exhibit several limitations. First, most energy models treat PUE as a fixed parameter rather than a dynamic variable dependent on ambient temperature, cooling technology, and server density [44,45]. Second, few models integrate technology selection with geographic optimization in the context of Japan’s pronounced regional variations in climate, renewable energy potential, and grid carbon intensity. Third, the interaction between the GX-ETS carbon pricing mechanism [16,17] and data center investment decisions has not been quantitatively modeled. Fourth, while the JST reports [5,6,7,8,9] provide detailed energy consumption projections, they do not incorporate economic optimization or policy feedback mechanisms. Fifth, the policy paradox between AI infrastructure subsidies and decarbonization targets, described in the MIC White Papers [10,11,12,13], has not been subjected to integrated cost optimization analysis.

The foregoing literature review reveals that no prior study has simultaneously addressed all five of these limitations within a single integrated framework.

Table 1. Methodological comparison of DC-DECOM with key prior studies. Abbreviations used in this table: DC-DECOM, Data Center Decarbonization Cost Optimization Model; PUE, Power Usage Effectiveness.

Feature	Masanet [31]	Shehabi [32]	Koronen [33]	Hao [34]	DC-DECOM
Dynamic PUE	No	No	No	No	Yes
Geographic optimization	No	No	No	Yes	Yes
Carbon pricing integration	No	No	No	No	Yes
Multi-year investment	No	No	No	No	Yes
Japan-specific	No	No	No	No	Yes

summarizes the methodological positioning of DC-DECOM relative to key prior studies.

Beyond the binary feature comparison in Table 1, the present work differs from these four prior studies in three substantive respects. First, Masanet et al. [31] and Shehabi et al. [32] provide global and U.S. accounting estimates of data center energy use; they do not formulate a normative cost optimization. DC-DECOM, in contrast, is normative: it asks what the cost-optimal configuration is under explicit policy constraints (NDC, GX-ETS), and the answer is a decision variable, not a tabulated estimate. Second, Koronen et al. [33] analyze European data centers under demand-response and waste-heat-recovery scenarios but treat PUE and grid emission factor as static national averages; DC-DECOM endogenizes both as regional, time-varying functions, which is what enables the geographic-arbitrage finding (RQ2). Third, Hao et al. [34] optimize operational dispatch across multiple data center microgrids on hourly timescales but do not include investment decisions or multi-year horizons. DC-DECOM incorporates a 17-year capacity-investment dimension nested over hourly dispatch, allowing CAPEX/OPEX trade-offs to drive the technology trajectory rather than only the within-day allocation. The methodological novelty of DC-DECOM lies in the simultaneous combination of all five features in Table 1, applied to the Japan-specific regulatory context (GX-ETS + 2040 NDC + Green Growth Strategy), which no prior study has addressed.

This study addresses these gaps by proposing DC-DECOM, a multi-regional bottom-up cost optimization model. The model incorporates dynamic PUE modeling calibrated against empirical data, geographic optimization across five regions, endogenous carbon pricing, and demand projections from the JST report series [5,6,7,8,9]. The academic contributions of this study are elaborated in the following subsection.

Research Objectives and Academic Contributions

To address the gaps identified above, this study formulates four explicit research questions that the DC-DECOM model is designed to answer:

RQ1.

Under Japan’s NDC trajectory and the GX-ETS carbon price schedule, what is the cost-optimal mix of cooling technologies, regional siting, and renewable energy procurement that achieves the 2040 ICT carbon-neutrality target?

RQ2.

How large is the system-cost advantage of geographic decentralization (R-CN) over a metropolitan-concentrated decarbonization strategy (C-CN) under identical carbon constraints, and through which channels (cooling efficiency, renewable energy access, grid carbon intensity) does this advantage arise?

RQ3.

What carbon-price threshold is required to make the cost-optimal decarbonization investments economically self-justifying for private-sector decision-makers, relative to the current GX-ETS price band?

RQ4.

How robust are the answers to RQ1–RQ3 to plausible variations in AI demand growth, hardware efficiency trajectory, PPA prices, and discount rate?

Whereas prior data center studies have evaluated predefined combinations of measures—such as cloud migration rates or incremental efficiency improvements—under fixed PUE assumptions and without geographic optimization, this study endogenizes these choices within a cost-minimization framework, deriving the optimal combination as a mathematical solution rather than a prior assumption. The principal contributions of DC-DECOM are threefold:

• Endogenous dynamic PUE: PUE is explicitly modeled as a function of ambient temperature and cooling technology, calibrated against published empirical data. This allows for the geographic cooling advantage of cold-climate regions to be quantified in cost terms.
• Integrated cost optimization: Technology investment (cooling, BESS, cogeneration), energy procurement (grid, PPA, on-site PV), and carbon cost (GX-ETS) are integrated within a single cost-minimization framework, enabling endogenous co-optimization of these interdependent decisions.
• Quantitative comparison of siting strategies: Centralized (C-CN) and decentralized (R-CN) siting strategies are compared under identical carbon constraints on both cost and emissions dimensions, providing the first quantitative assessment of geographic arbitrage in Japan’s data center context.

The remainder of this paper is organized as follows. Section 2 reviews the policy and technology context. Section 3 presents the model formulation, including the objective function, demand module, thermodynamic cooling module, and constraint structure. Section 4 reports the simulation results across three scenarios. Section 5 discusses policy implications, model validation, and limitations. Section 6 concludes with a summary of findings.

2. Policy and Technology Context

2.1. Japan’s Data Center Policy Framework

Japan’s data center policy operates along two axes that are in structural tension, which can be conceptualized as an “accelerator” (demand side) and a “brake” (supply and efficiency side).

On the demand side (accelerator), the Cloud Supply Security Plan under the Economic Security Promotion Act allocated subsidies exceeding ¥90 billion for GPU server deployment (¥50.1 billion to Sakura Internet, ¥42.1 billion to SoftBank) [46], directly increasing electricity demand. Microsoft announced a ¥1.6 trillion ($10 billion) investment in Japan’s AI and cloud infrastructure through 2029 [47]. Oracle, Google, and Microsoft were selected as official government cloud providers, triggering a combined $28 billion (¥4 trillion) investment wave [3].

On the supply and efficiency side (brake), the Green Growth Strategy [15] and the Digital Infrastructure Resilience Program (¥50 billion fund) [48] aim to reduce emissions and promote geographic decentralization. The MIC White Papers [10,11,12,13] have documented this tension across successive editions. A fundamental question thus arises: can the brake be applied forcefully enough to counterbalance the accelerator while maintaining the pace of digital transformation?

The geographic concentration of Japan’s data center infrastructure creates both resilience vulnerabilities and sustainability constraints. Approximately 80% of capacity is located in the Tokyo and Kansai metropolitan areas [48]. The government has designated Hokkaido and Kyushu as strategic regional hubs, with the Expert Group Meetings on the Development of Digital Infrastructures recommending five core and up to ten regional data center hubs [48]. The Agency for Natural Resources and Energy established a data center benchmark system recommending PUE improvement to 1.4 by 2030 [20]. Hokkaido’s electricity demand is expected to grow the fastest nationally, driven by data center and semiconductor investments [4]. The MIC’s Communications Usage Trend Survey found that while 70.4% of enterprises use cloud services, only about 10% use them for system development, and sustainability was not listed among the reasons for cloud adoption—suggesting insufficient awareness of the environmental benefits of cloud migration [49].

2.2. Carbon Pricing and the GX-ETS

As noted in Section 1, the GX-ETS became mandatory in April 2026. The system covers approximately 300–400 companies responsible for about 60% of Japan’s CO₂ emissions [16]. For FY2026, the price floor is set at ¥1700/t-CO₂ and the price cap at ¥4300/t-CO₂ [17]. A carbon levy (GX-Surcharge) on fossil fuels will be introduced from FY2028, and auctioning of allowances will commence in FY2033, with combined revenue targets of at least ¥20 trillion [17]. The Mitsubishi Research Institute projects carbon prices rising to ¥7000–15,000/t-CO₂ by the mid-2030s [50]. Ding [51] analyzed the impacts of carbon pricing on Japan’s regional electricity markets in Humanities and Social Sciences Communications, finding heterogeneous pass-through rates across utilities—with Hokkaido showing a negative relationship between electricity price and carbon cost, suggesting that carbon pricing may actually improve the relative competitiveness of Hokkaido-based data centers. Green [52] found in Environmental Research Letters that carbon pricing systems globally have had limited success in triggering zero-carbon investments at low price levels, indicating that prices substantially above the current GX-ETS range would be required to drive technology transitions.

2.3. Cooling Technology Landscape

The stagnation of global average PUE at 1.55–1.58 since 2020 [18] reflects the physical limits of air-cooling technology in the face of rising server power densities. Brady et al. [44] provided a critical assessment of PUE calculation methodologies in Energy Conversion and Management, and Avgerinou et al. [45] analyzed PUE trends under the European Code of Conduct in Energies. Ni and Bai [53] reviewed air conditioning energy performance in data centers in Renewable and Sustainable Energy Reviews, identifying the nonlinear relationship between ambient temperature and cooling energy that motivates the dynamic PUE model in this study. These findings collectively indicate that air cooling has reached diminishing returns as a pathway for further PUE improvement.

The transition from air to liquid cooling is now being driven by physical necessity. NVIDIA’s H100 GPU consumes approximately 700 W per chip. The Blackwell B200 delivers 2.5× the inference performance of H100 but at a substantially higher TDP of approximately 1000 W per chip (up to 1200 W in the highest-bin SKUs), while the GB200 NVL72 rack-scale system is rated at 120 kW and requires liquid cooling [27]. The B300 (Blackwell Ultra), shipped in January 2026, achieves 14 petaFLOPS of FP4 compute per chip at approximately 1200 W TDP [54]. At these power densities, air cooling is physically inadequate. Direct liquid cooling (DLC) achieves a PUE of 1.10–1.20 [55], while immersion cooling reaches a PUE of below 1.05 [22,23]. Henderson et al. [56] addressed the systematic reporting of energy and carbon footprints of machine learning in the Journal of Machine Learning Research, providing methodological foundations for quantifying AI workload energy impacts. Desislavov et al. [57] analyzed long-term trends in AI inference energy consumption in Sustainable Computing: Informatics and Systems, confirming sustained hardware efficiency gains of approximately 2× per GPU generation.

2.4. Regional Energy Characteristics

Japan’s regional climate variation creates significant differences in data center cooling requirements. Tokyo’s annual average temperature is approximately 16.5 °C, with summer peaks exceeding 35 °C, while Sapporo (Hokkaido) averages 9.2 °C, with summer peaks rarely exceeding 30 °C—a year-round difference of 7–10 °C [58]. NTT Facilities reported that cooling-energy consumption increases nonlinearly above 15 °C ambient temperature, with mechanical chiller activation required above this threshold [58]. In Hokkaido, free cooling is available for 8–10 months annually [28], and the White Data Center in Bibai has demonstrated that snow-based cooling can reduce cooling costs by 55% [28].

Japan’s national average grid emission factor was 0.43 kg-CO₂/kWh in FY2022 [59], but regional variation is significant and will widen as renewable deployment accelerates unevenly. The Ministry of the Environment’s REPOS database shows Hokkaido’s land-based solar potential at 337,471 MW versus Kanto’s 98,833 MW [60]. Japan’s renewable energy share reached 22.9% of electricity generation in FY2023, with solar contributing approximately 10% [61]. The 7th Strategic Energy Plan targets 36–38% by 2030 and 40–50% by 2040, with nuclear power targeted at approximately 20% [62]. Hokkaido’s abundant wind and solar resources, combined with the New Hokkaido-Honshu HVDC Link strengthening grid connectivity to the mainland [28], position the region for rapid grid decarbonization. Google’s first PPAs in Japan—60 MW of solar capacity with Clean Energy Connect and Shizen Energy—demonstrate the growing corporate PPA market [30], as reported by the Renewable Energy Institute [63].

The seven principal Japanese policy instruments most relevant to data center decarbonization—spanning ministerial responsibility, target year, and quantitative thresholds—are summarized in

Table 2. Summary of key government policies related to data center decarbonization.

Policy	Ministry	Objective	Key Target
Green Growth Strategy [15]	METI	ICT carbon neutrality	2040; 30% efficiency gain for new DCs by 2030
Digital Infrastructure Resilience [48]	MIC	Geographic decentralization	5 core + 10 regional hubs
GX-ETS mandatory phase [16]	Cabinet/METI	Carbon pricing	¥1700–4300/t-CO2 (FY2026); auction from FY2033
7th Strategic Energy Plan [62]	METI	Energy mix	RE 36–38% by 2030, 40–50% by 2040
Cloud Supply Security Plan [46]	METI	AI infrastructure	¥90B+ for GPU capacity
NDC (updated 2025) [14]	Cabinet/MOE	GHG reduction	73% by 2040 (vs. 2013)
DC Benchmark System [20]	ANRE	Energy efficiency	National avg. PUE 1.4 by 2030 (target)

. These instruments collectively define the regulatory and incentive boundary conditions that the DC-DECOM model embeds, and are referenced individually throughout Section 3 and Section 4.

3. Materials and Methods

3.1. Information Base

DC-DECOM is data-intensive and integrates seven categories of public input data, summarized in

Table 3. Information base of DC-DECOM: data inputs, sources, granularity, and model components.

Input Category	Primary Source	Granularity	Used in
IT demand projections	JST Vol. 2–5 [6,7,8,9]; Wood Mackenzie [3]	National annual, 2024–2040	Section 3.4, Equation (7)
Hourly temperature	Japan Meteorological Agency (5 cities, 2020–2023 mean) [65]	96 representative hours × five regions	Section 3.5, Equations (8) and (9)
Renewable potential	MOE REPOS [60]	Regional MW, by RE type	Section 3.8, Equation (15)
Technology cost & performance	METI Cost Verification WG [66]; NEDO [67]; REI [63]; Vertiv [55]	National annual	Section 3.6
Grid emission factor	MOE Emission Factors by Utility [59]	Regional, declining 2024–2040	Section 3.7, Equation (10)
Carbon price trajectory	GX-ETS [16,17]; MRI forecast [50]	Annual, 2024–2040	Section 3.7, Equation (10)
Grid electricity tariff	ANRE Survey of Electric Power Statistics [68]	Regional, special-high-voltage rate	Section 3.6 and Section 3.7

. The data-intensive nature is characteristic of the optimization-based class of energy systems models reviewed by Pfenninger et al. [64], in which model fidelity depends critically on the transparency of input data. To aid reproducibility, the table identifies, for each input category, the primary source, its temporal/spatial granularity, and the model component that consumes it. Subsequent subsections describe how each input is incorporated into the corresponding model equation.

All datasets are publicly accessible. The complete numerical parameter sheet, hourly temperature time series, and regional aggregation logic are deposited in the Zenodo software archive (DOI 10.5281/zenodo.20312805), as detailed in the Data Availability Statement.

3.2. Model Structure and Rationale

Existing energy system models for data centers typically treat PUE as a fixed exogenous parameter, do not incorporate geographic optimization, and lack endogenous carbon pricing mechanisms (see Table 1). To overcome these limitations, the present study develops DC-DECOM, a multi-regional bottom-up linear programming model that minimizes total discounted system cost over the period 2024–2040. The modeling approach draws on the TIMES/MARKAL family of energy system models [69,70], which have been widely applied to national energy planning. Imagawa et al. [35] demonstrated the applicability of this approach to Japan’s energy system with detailed technology selection, and Yi et al. [37] extended it to power expansion planning with hydrogen integration. Recent peer-reviewed extensions of this methodology—e.g., Heuberger et al. [71] for power-sector capacity expansion with endogenous learning—demonstrate the flexibility of the linear-programming framework for sector-specific decarbonization analysis. DC-DECOM adapts this methodology to the specific characteristics of data center energy systems, where cooling technology choice, geographic siting, and carbon pricing interact in ways not captured by general energy models. The remainder of this section describes the model’s specific structure: objective function (Section 3.3), demand module (Section 3.4), thermodynamic cooling module (Section 3.5), technology portfolio (Section 3.6), carbon flow module (Section 3.7), constraints (Section 3.8), scenario design (Section 3.9), and implementation details (Section 3.10).

Five geographic regions are modeled: Hokkaido, Tohoku, Kanto, Kansai, and Kyushu. These were selected to represent the range of climatic conditions (annual mean temperatures from 9.2 °C to 17.3 °C), renewable energy endowments (solar potential ranging from 98,833 MW in Kanto to 337,471 MW in Hokkaido [60]), grid carbon intensities, and existing data center infrastructure concentrations reported in the MIC White Papers [10,11,12,13]. Annual time steps govern investment decisions; representative hourly profiles—comprising 24 h for each of four seasons (96 representative hours per year), derived from JMA meteorological data for 2020–2023—are used for operational optimization of energy dispatch and BESS charging/discharging.

3.3. Objective Function

The model minimizes the net present value of total system cost over the planning horizon:

$$min{\displaystyle \sum _{t=2024}^{2040}}{\displaystyle \frac{\mathrm{CAPEX}\left(t\right)+\mathrm{OPEX}\left(t\right)+C_{\mathrm{energy}}\left(t\right)+C_{\mathrm{carbon}}\left(t\right)+C_{\mathrm{curtail}}\left(t\right)}{{(1+d)}^{t-2024}}}$$

(1)

where $\mathrm{CAPEX}\left(t\right)$ is capital expenditure on cooling systems (air, DLC, immersion), battery energy storage systems (BESSs), gas cogeneration systems (CGSs), and on-site photovoltaic (PV) systems; $\mathrm{OPEX}\left(t\right)$ is annual operation and maintenance cost; $C_{\mathrm{energy}}\left(t\right)$ is energy procurement cost from grid electricity and PPA contracts; $C_{\mathrm{carbon}}\left(t\right)$ is carbon cost under the GX-ETS; $C_{\mathrm{curtail}}\left(t\right)$ is demand curtailment penalty representing economic losses from supply interruptions; and d is the social discount rate. Note that the discount rate is denoted d to avoid notational conflict with the regional index r. The specific values of d and the curtailment penalty are documented in Section 3.10. This formulation enables evaluation of the temporal trade-off between technologies with high initial investment but low operating costs (e.g., immersion cooling with on-site renewables) and those with lower upfront costs but higher long-term energy and carbon expenses (e.g., conventional air cooling with grid dependence).

The model uses a two-level temporal structure. Investment decisions (capacity additions) are taken at annual time steps $t\in \{2024,\dots ,2040\}$ (17 years). Operational dispatch and BESS state-of-charge are resolved at representative hourly steps $h\in \{1,\dots ,96\}$ corresponding to 24 h × four seasons within each year. Annual cost terms in Equation (1) aggregate the corresponding hourly outcomes within each year; in particular, $C_{\mathrm{energy}}\left(t\right)$ and $C_{\mathrm{carbon}}\left(t\right)$ are sums over h. The hourly index h thus connects the annual investment problem [Equation (1)] to the hourly dispatch problem in Equations (6)–(12).

Each cost term in Equation (1) is computed as follows:

$$\mathrm{CAPEX}\left(t\right)=\sum _r\sum _{\tau }{I}_{\tau }\left(t\right)\Delta {\mathrm{Cap}}_{\tau }(r,t)$$

(2)

$$\mathrm{OPEX}\left(t\right)=\sum _r\sum _{\tau }{o}_{\tau }{\mathrm{Cap}}_{\tau }(r,t)$$

(3)

$$C_{\mathrm{energy}}\left(t\right)=\sum _r\sum _h{w}_h\left[p_{\mathrm{grid}}(r,t)E_{\mathrm{grid}}(r,h,t)+p_{\mathrm{PPA}}(r,t)E_{\mathrm{PPA}}(r,h,t)\right]$$

(4)

$$C_{\mathrm{curtail}}\left(t\right)={\pi }_{\mathrm{curt}}\sum _r\sum _h{w}_h\mathrm{Curt}(r,h,t)$$

(5)

where $\tau$ indexes technology classes (air cooling, DLC, immersion, BESS, CGS, PV), $I_{\tau }\left(t\right)$ is the unit capital cost in year t, $\Delta {\mathrm{Cap}}_{\tau }(r,t)$ is new capacity addition, $o_{\tau }$ is the per-MW-yr O&M cost, $p_{\mathrm{grid}}(r,t)$ and $p_{\mathrm{PPA}}(r,t)$ are grid and PPA energy prices, $E_{\mathrm{grid}}(r,h,t)$ and $E_{\mathrm{PPA}}(r,h,t)$ are grid and PPA dispatch volumes, $w_h$ is the weighting factor that maps each representative hour h to the corresponding number of actual hours in the year (Section 3.11), ${\pi }_{\mathrm{curt}}$ is the curtailment penalty (Section 3.10), and $\mathrm{Curt}(r,h,t)$ is unserved load. The carbon cost term $C_{\mathrm{carbon}}\left(t\right)$ is given separately in Section 3.7, Equation (10).

The complete set of decision variables optimized by DC-DECOM is summarized in

Table 4. Decision variables in the DC-DECOM optimization.

Variable	Indices	Unit	Range
$\Delta {\mathrm{Cap}}_{\tau }(r,t)$    new capacity	technology $\tau$, region r, year t	MW	$[0,{\mathrm{Deploy}}_{max,\tau }\left(r\right)]$
${\mathrm{Cap}}_{\tau }(r,t)$    cumulative capacity	$\tau ,r,t$	MW	accounting state
$E_s(r,h,t)$    energy supply	source $s\in \{\mathrm{grid},\mathrm{PV},\mathrm{BESS}\mathrm{disch}.,\mathrm{CGS},\mathrm{PPA}\}$, $r,h,t$	MWh	≥0
${\mathrm{SoC}}_{\mathrm{BESS}}(r,h,t)$    battery state-of-charge	$r,h,t$	MWh	$[0,{\mathrm{Cap}}_{\mathrm{BESS}}(r,t)]$
$E_{\mathrm{BESS},\mathrm{ch}}(r,h,t)$    battery charging energy	$r,h,t$	MWh	$[0,E_{\mathrm{BESS},\max }(r,t)]$
$\mathrm{Curt}(r,h,t)$    demand curtailment	$r,h,t$	MWh	≥0

. The model contains approximately 45,000 such variables in total.

3.4. Demand Module

Regional data center electricity demand in Japan at region r (Hokkaido, Tohoku, Kanto, Kansai, or Kyushu), representative hour h, and year t is computed as

$$\mathrm{Load}(r,h,t)=\left[P_{\mathrm{base}}(r,h,t)+P_{\mathrm{AI}}(r,h,t)\right]\times \mathrm{PUE}(\mathrm{tech},r,h,t)$$

(6)

where $P_{\mathrm{base}}(r,h,t)$ represents conventional IT workload (web hosting, enterprise applications, storage; assumed temporally stationary at the 2024 level) and $P_{\mathrm{AI}}(r,h,t)$ represents AI-specific workload (training and inference). The latter grows over time according to

$$P_{\mathrm{AI}}(r,h,t)=P_{\mathrm{AI},0}(r,h)\times {\displaystyle \frac{{(1+g_{\mathrm{AI}})}^{t-t_0}}{{\eta }_{\mathrm{HW}}\left(t\right)\times {\eta }_{\mathrm{arch}}\left(t\right)}}$$

(7)

The AI compute demand growth rate $g_{\mathrm{AI}}$ is set at 25–35% annually. This range is based on the IEA’s projection that AI could represent 35–55% of total data center electricity demand by 2030 [1], corroborating the JST Vol.1 finding that without efficiency measures, IT-related electricity consumption would grow by orders of magnitude [5]. The hardware efficiency improvement rate ${\eta }_{\mathrm{HW}}$ is set at 30% annually, reflecting observed GPU generational improvements: NVIDIA’s Blackwell B200 delivers 2.5× the H100’s inference performance (albeit at approximately 1000 W TDP, up from 700 W for H100) [27], and the B300 achieves 14 petaFLOPS of FP4 compute per chip at approximately 1200 W TDP [54]. Desislavov et al. [57] confirmed sustained efficiency gains in AI inference hardware. The architectural efficiency factor ${\eta }_{\mathrm{arch}}$ captures improvements from algorithmic optimization, model compression, and next-generation computing architectures, modeled as a logistic adoption curve with a saturation level of 5× (reflecting the combined effect of algorithmic optimization and next-generation computing), an inflection point at 2032, and a growth rate parameter of 0.3. The JST Vol.3 “Modest” scenario assumptions—CPU doubling every 2 years, GPU every 1.5 years, memory and switches halving every 2 years, storage improving 10–30× per decade [7]—inform the baseline ${\eta }_{\mathrm{HW}}$ trajectory.

The base year (2024) national IT load of 19 TWh [3] is distributed across regions proportionally to existing data center capacity: Kanto 60%, Kansai 22%, Kyushu 10%, Tohoku 5%, and Hokkaido 3%, as reported in MIC White Papers [10,12] and the Impress Research Institute’s Data Center Survey Report 2025 [72].

3.5. Thermodynamic Cooling Module

A central feature of DC-DECOM is its endogenous treatment of PUE as a function of ambient temperature and cooling technology. This addresses a limitation identified by Brady et al. [44] in their critical assessment of PUE calculation, and by Avgerinou et al. [45] in their analysis of PUE trends. Most existing models use fixed PUE values (e.g., 1.5 for all facilities), which obscures the geographic cooling advantage that is central to the siting optimization in this study.

The air-cooling PUE model is

$${\mathrm{PUE}}_{\mathrm{air}}\left(T\right)=C_{\mathrm{base}}+\alpha ·T+\beta ·H(T-T_{\mathrm{thresh}})·{(T-T_{\mathrm{thresh}})}^2$$

(8)

where $C_{\mathrm{base}}=1.14$ is the base PUE component (IT power distribution, lighting, UPS losses), $\alpha =0.0167$ °C⁻¹ is the linear temperature sensitivity coefficient, $\beta =0.0102$ °C⁻² is the quadratic coefficient for mechanical cooling above the free-cooling threshold, $T_{\mathrm{thresh}}=15$ °C is the free-cooling threshold temperature, and $H(·)$ is the Heaviside step function. The linear term captures the gradual increase in fan and economizer energy with rising temperature, while the quadratic term above 15 °C reflects the nonlinear increase in mechanical chiller energy consumption reported by NTT Facilities [58] and Ni and Bai [53]. The parameters $\{C_{\mathrm{base}},\alpha ,\beta \}$ were calibrated by ordinary least-squares regression against five regional empirical data points (

Table 5. PUE model outputs versus empirical benchmarks.

Location	Mean T (°C)	Air PUE (Model)	Air PUE (Empirical)	Liquid PUE (Model)	Liquid PUE (Empirical)
Sapporo	9.2	1.29	1.2–1.3 [58]	1.07	1.05–1.10 [22]
Sendai	12.8	1.35	1.3–1.4 [18]	1.07	1.05–1.10 [23]
Tokyo	16.5	1.44	1.4–1.6 [18]	1.08	1.05–1.10 [23]
Osaka	17.1	1.47	1.4–1.6 [18]	1.08	1.05–1.10 [23]
Fukuoka	17.3	1.48	1.4–1.6 [18]	1.08	1.05–1.10 [23]

), with $R^2=0.988$. The high $R^2$ should be interpreted with caution given the limited five-point regression basis (three degrees of freedom for three free parameters $\{C_{\mathrm{base}},\alpha ,\beta \}$). To clarify what is empirically estimated and what is prescribed, the functional form—linear up to $T_{\mathrm{thresh}}$, quadratic above—is prescribed from chiller-activation thermodynamics as reported by NTT Facilities [58]; it is not selected from the data. The five empirical points are used to estimate the slope coefficients $\alpha$ and $\beta$ (and the intercept $C_{\mathrm{base}}$), not to validate the functional structure. Approximate 95% confidence intervals on the estimated coefficients are $\alpha \in [0.013,0.020]$°C⁻¹ and $\beta \in [0.008,0.013]$°C⁻²; the resulting PUE uncertainty at Japan’s annual mean temperatures is $\pm 0.04$ for air cooling. This calibration uncertainty is examined through the alternative PUE functional-form tests reported in Section 3.11 (purely linear and purely quadratic specifications), which yield total system costs within $\pm 2.5\%$ of baseline. At Tokyo’s annual mean temperature (16.5 °C), this model yields a PUE of ≈1.44, consistent with the domestic average for newer facilities reported by the JDCC, as summarized in the Uptime Institute survey [18]. At Sapporo’s mean temperature (9.2 °C), it yields a PUE of ≈1.29, reflecting the extended free-cooling advantage. At Fukuoka (17.3 °C), it yields a PUE of ≈1.48. The model is calibrated for Japan’s regional annual-mean temperatures (9–17 °C) and should not be extrapolated to substantially higher temperatures.

The liquid cooling PUE model is

$${\mathrm{PUE}}_{\mathrm{liquid}}\left(T\right)=1.0+C_{\mathrm{pump}}+\gamma ·T$$

(9)

where $C_{\mathrm{pump}}=0.055$ represents the aggregate liquid cooling overhead (pump power, coolant circulation, and heat rejection) and $\gamma =0.0015$ °C⁻¹ is the temperature sensitivity coefficient, an order of magnitude smaller than for air cooling ($\gamma /\alpha =0.09$). This yields a PUE of ≈1.07–1.08 across Japan’s annual mean temperature range, consistent with Haghshenas et al.’s [23] measurements (PUE 1.03–1.10 for immersion cooling) and the MHI demonstration (PUE 1.05) [22]. The near-independence from ambient temperature is a key advantage: liquid cooling effectively neutralizes the geographic cooling penalty, meaning that the siting advantage of cold regions diminishes as liquid cooling adoption increases. To validate cross-consistency, we note that the model yields a PUE of ≈1.58 for air cooling at an annual mean temperature of approximately 18.6 °C, consistent with the Uptime Institute’s reported global average PUE of 1.58 [18] for a mix of facilities in warm-climate regions.

Figure 1 shows the calibrated PUE model curves (lines) for air cooling and liquid cooling across Japan’s five regions, alongside empirical data points (markers) from the Uptime Institute [18], JDCC (as summarized in [18]), NTT Facilities [58], Haghshenas et al. [23], and MHI [22]. The dashed line indicates Japan’s 2030 PUE target of 1.4 [20]. The shaded region between the two curves represents the cooling-energy savings achievable by transitioning from air cooling to liquid cooling.

3.6. Technology Portfolio and Cost Assumptions

The model selects from a portfolio of cooling, generation, storage, and resilience technologies for each region. Cost and performance parameters, drawn from government reports and peer-reviewed literature, are summarized in

Table 6. Technology cost and performance parameters.

Technology	CAPEX	Key Performance	Source
Air cooling (conventional)	Baseline	PUE 1.4–1.6	[18]
Direct liquid cooling (DLC)	+20–30% vs. air	PUE 1.10–1.20	[55]
Immersion cooling	+40–60% vs. air	PUE 1.03–1.08	[22,23]
Utility-scale PV	¥9.9/kWh (LCOE)	CF 13–17% (regional)	[66]
BESS (Li-ion)	¥60k/kWh (2025) → ¥30k/kWh (2035)	Round-trip eff. 90%	[67]
Gas CGS	¥230k/kW	Total eff. 76.2% (elec. 42.3% + heat 33.9%)	[73]
Solar PPA	¥13–16/kWh	20-year contract term	[63]

.

Grid electricity tariffs are set at regional special high-voltage rates: Tokyo ¥19.9/kWh, Kansai ¥16.0/kWh, Kyushu ¥15.5/kWh, and Hokkaido ¥20.0/kWh [68]. On-site PV generation profiles are based on regional solar irradiance data from the Japan Meteorological Agency (JMA). PPA contracts provide long-term renewable energy procurement at contracted prices [63]. BESS dispatch is optimized for charging during low-price/high-RE periods and discharging during peak periods.

3.7. Carbon Flow Module

CO₂ emissions are tracked from two sources: grid electricity consumption (using regional grid emission factors) and on-site fossil fuel combustion (CGS, emergency generators). The carbon cost is computed as

$$C_{\mathrm{carbon}}\left(t\right)=P_{\mathrm{carbon}}\left(t\right)\sum _r\sum _h{w}_h{\displaystyle \frac{\left[E_{\mathrm{grid}}(r,h,t){\mathrm{EF}}_{\mathrm{grid}}(r,t)+E_{\mathrm{CGS}}(r,h,t){\mathrm{EF}}_{\mathrm{fossil}}\right]}{{10}^3}}$$

(10)

where ${\mathrm{EF}}_{\mathrm{grid}}(r,t)$ is the regional grid emission factor declining over time with grid decarbonization per the 7th Strategic Energy Plan [62], ${\mathrm{EF}}_{\mathrm{fossil}}$ is the natural gas emission factor for CGS, and $P_{\mathrm{carbon}}\left(t\right)$ is the carbon price trajectory under the GX-ETS [16,17,50]. The base year national average emission factor is 0.43 kg-CO₂/kWh [59]. Because the emission factors ${\mathrm{EF}}_{\mathrm{grid}}$ and ${\mathrm{EF}}_{\mathrm{fossil}}$ are expressed in kg-CO₂/kWh and energies in kWh, the bracketed expression yields kg-CO₂ per region–hour; division by ${10}^3$ converts this to t-CO₂ for consistency with $P_{\mathrm{carbon}}\left(t\right)$ in ¥/t-CO₂, so that $C_{\mathrm{carbon}}\left(t\right)$ is reported in ¥. Regional differentiation reflects each utility’s generation mix and renewable deployment plans; Hokkaido shows the fastest projected emission-factor decline owing to its abundant renewable resources [60].

For diagnostic interpretation of how data center electricity consumption maps into CO₂ emissions, the contribution of each region–year combination can be decomposed as the product of three factors (this expression is used solely for the wedge decomposition in Section 4.3; the optimization itself uses the source-specific accounting in Equation (10)):

$$E_{{\mathrm{CO}}_2}(r,t)=\mathrm{Load}(r,t)\times {\mathrm{EF}}_{\mathrm{grid}}(r,t)\times \left[1-f_{\mathrm{RE}}(r,t)\right]$$

(11)

where $\mathrm{Load}(r,t)$ is the total facility electricity consumption (IT load × PUE), ${\mathrm{EF}}_{\mathrm{grid}}(r,t)$ is the regional grid emission factor, and $f_{\mathrm{RE}}(r,t)$ is the fraction of electricity sourced from zero-carbon supply (on-site PV, PPA, or non-emitting grid). This formulation shows that emissions can be reduced through three independent channels: (1) reducing total electricity consumption via PUE improvement (i.e., liquid cooling), (2) lowering the grid emission factor through renewable energy procurement and grid decarbonization, and (3) siting in regions where both PUE and emission factors are inherently lower (i.e., geographic optimization). The multiplicative structure implies that the combined effect of simultaneous measures exceeds the sum of individual contributions; this property is leveraged in the scenario-decomposition analysis presented in Section 4.3.

3.8. Constraints

The model is subject to five categories of constraints, each formulated as follows.

(a) Energy balance constraint. At every region r and hourly time step h, total supply must equal total demand:

$$\sum _s{E}_s(r,h,t)+\mathrm{Curt}(r,h,t)=\mathrm{Load}(r,h,t)+E_{\mathrm{BESS},\mathrm{ch}}(r,h,t)\forall r,\forall h,\forall t$$

(12)

where the supply sources are $s\in \{\mathrm{grid},\mathrm{PV},\mathrm{PPA},\mathrm{BESS}\mathrm{discharge},\mathrm{CGS}\}$ and $\mathrm{Curt}(r,h,t)$ is unserved demand. Because the curtailment penalty ${\pi }_{\mathrm{curt}}$ is set sufficiently high (Section 3.10), $\mathrm{Curt}(r,h,t)$ is zero in all optimal solutions.

(b) CO₂ emission cap. In the carbon-constrained scenarios (C-CN, R-CN), total annual emissions across all regions must not exceed a linearly declining cap consistent with the NDC pathway [14] and the Green Growth Strategy’s 2040 ICT carbon-neutrality target [15]:

$$\sum _r{E}_{\mathrm{total}}(r,t)\le E_{\mathrm{cap}}\left(t\right)\forall t$$

(13)

$$E_{\mathrm{cap}}\left(t\right)=E_{2024}\times {\displaystyle \frac{2040-t}{2040-2024}}\mathrm{for}\mathrm{C}\text{-}\mathrm{CN}\mathrm{and}\mathrm{R}\text{-}\mathrm{CN}$$

(14)

where $E_{2024}$ is the base-year total emission level and $E_{\mathrm{cap}}\left(2040\right)=0$. This linear trajectory yields intermediate reduction rates of approximately 47% by 2030 and 66% by 2035 relative to 2024 emissions, broadly consistent with Japan’s NDC milestones of 46% by 2030, 60% by 2035, and 73% by 2040 relative to 2013 levels [14]. The slight discrepancy arises because the model’s base year (2024) already reflects emission reductions achieved since 2013.

(c) Renewable capacity bound. Regional PV and PPA capacity is bounded by the Ministry of the Environment’s REPOS data [60]:

$${\mathrm{Cap}}_{\mathrm{PV}}\left(r\right)+{\mathrm{Cap}}_{\mathrm{PPA}}\left(r\right)\le {\mathrm{RE}}_{max}\left(r\right)\forall r$$

(15)

(d) Resilience constraint (N + 1). Minimum backup power capacity must maintain N + 1 redundancy for critical loads, following the resilience framework of Nishikura et al. [36]:

$$\sum _s{\mathrm{Cap}}_{\mathrm{backup},s}\left(r\right)\ge \left(1+{\displaystyle \frac{1}{N}}\right)\times P_{\mathrm{critical}}\left(r\right)\forall r$$

(16)

(e) Technology deployment rate. Annual new capacity additions are bounded by realistic construction and supply chain constraints, informed by the pace of announced projects such as the SoftBank Tomakomai facility [29]:

$$\Delta {\mathrm{Cap}}_{\mathrm{tech}}(r,t)\le {\mathrm{Deploy}}_{max,\mathrm{tech}}\left(r\right)\forall r,\forall t,\forall \mathrm{tech}$$

(17)

3.9. Scenario Design

Three scenarios, summarized in

Table 7. Scenario definitions and key assumptions.

Parameter	REF	C-CN	R-CN
CO2 constraint	None	Net-zero by 2040 (linear decline)	Net-zero by 2040 (linear decline)
Geographic constraint	Current trends continue	New large-scale DCs in Kanto/Kansai only	All five regions (model optimizes allocation)
Policy incentives	Current policies	Current policies	Regional DC subsidies included
AI demand CAGR	30% (baseline)	30% (baseline)	30% (baseline)
HW efficiency improvement	30%/yr (baseline)	30%/yr (baseline)	30%/yr (baseline)
Carbon price	GX-ETS trajectory	GX-ETS trajectory	GX-ETS trajectory

, are designed to evaluate alternative decarbonization pathways, reflecting the key strategic choice between maintaining metropolitan concentration and pursuing geographic decentralization:

The REF scenario serves as the baseline for measuring the incremental cost of decarbonization. The C-CN scenario tests the cost of achieving net-zero while maintaining the current geographic concentration pattern. The R-CN scenario tests whether geographic arbitrage—exploiting regional differences in climate, renewable energy, and grid carbon intensity—can reduce the total cost of decarbonization. Input data are calibrated against: Wood Mackenzie [3] and JST [5,6,7,8,9] for demand projections; Uptime Institute [18] for PUE benchmarks; REPOS [60] for renewable potential; the GX-ETS price schedule [16,17]; regional grid emission factors [59]; and cloud migration CO₂ reduction estimates from 451 Research [39] and Deloitte Tohmatsu [40]. The years 2024–2025 are treated as calibration years with capacity and demand fixed at observed values; optimization of investment decisions begins from 2026.

3.10. Model Implementation

The model is implemented as a linear program and solved using the PuLP Python library (v2.8.0; Python 3.11) with the CBC (COIN-OR Branch and Cut, v2.10.10) solver. Because the PUE equations [Equations (8) and (9)] contain a Heaviside step function and a quadratic term, PUE values are pre-computed for each region and representative hour using JMA hourly temperature data, and entered as fixed coefficients in the LP. This avoids nonlinearity in the optimization while preserving the temperature-dependent behavior of PUE. The social discount rate is set at $d=0.03$, following the convention adopted in long-term Japanese energy system studies such as Imagawa et al. [35] and the IEA-ETSAP TIMES documentation [69]. A social rather than a private rate is used because the model evaluates infrastructure with multi-decade public-good externalities (national grid load, emissions trajectory, regional development); a private rate would over-discount these long-term welfare effects. Sensitivity to the discount rate is reported in Section 4.6. The curtailment penalty $C_{\mathrm{curtail}}$ is set at ¥500/kWh, reflecting the estimated economic cost of unserved energy for data center operations; in practice, this penalty is sufficiently high that curtailment is never selected in the optimal solution under any scenario. The formulation comprises approximately 45,000 decision variables (covering technology capacity, energy dispatch, and BESS state-of-charge across five regions, 17 annual time steps, and representative hourly profiles) and approximately 120,000 constraints (energy balance, emission caps, capacity bounds, and deployment limits). A single scenario run completes in approximately 90 s on a standard desktop computer (Intel Core i7 CPU; Intel Corporation, Santa Clara, CA, USA; 32 GB RAM). The complete model code and input parameter files are publicly available on GitHub at https://github.com/jtoyoha-ops/dc-decom (accessed on 1 March 2026) and archived on Zenodo (DOI 10.5281/zenodo.20312805).

3.11. Data Preprocessing and Numerical Parameters

To support reproducibility, this subsection documents (i) the data-transformation steps that map raw public inputs into model coefficients, and (ii) the complete set of numerical parameters used in the LP. The corresponding spreadsheet is deposited together with the source code in the Zenodo archive (DOI 10.5281/zenodo.20312805).

Data transformations. Four preprocessing steps are applied to the public inputs:

• Regional load disaggregation. The national 2024 IT load of 19 TWh [3] is allocated to the five regions in proportion to the existing data center capacity shares reported in MIC White Papers [10,12] (Hokkaido 3%, Tohoku 5%, Kanto 60%, Kansai 22%, Kyushu 10%), yielding regional base loads of $\{0.57,0.95,11.40,4.18,1.90\}$ TWh. Subsequent demand growth follows Equation (7) and is allocated using the same shares for REF and C-CN, while R-CN allocates the new growth via the optimization.
• Hourly profile generation. Hourly temperature observations (2020–2023) at the JMA stations of Sapporo, Sendai, Tokyo, Osaka, and Fukuoka are reduced to 96 representative hours per region (24 h × four seasons: DJF, MAM, JJA, SON) by computing the seasonal-diurnal mean. Each representative hour carries a weight equal to the number of actual hours it represents.
• Time-varying parameter interpolation. The grid emission factor ${\mathrm{EF}}_{\mathrm{grid}}(r,t)$ declines linearly from the FY2022 base value [59] to the 7th Strategic Energy Plan target [62] (40–50% renewable share by 2040). The carbon price $P_{\mathrm{carbon}}\left(t\right)$ is piecewise-linear: ¥1700–4300/t-CO₂ for FY2026 [17], escalating linearly to the MRI mid-2030s forecast (¥7000–15,000/t-CO₂) [50]. Hardware-efficiency improvement compounds at 30%/yr (CPU) and 50%/yr (GPU) following JST Vol. 3 estimates [7].
• Currency normalization. All monetary values are expressed in 2024 constant yen; nominal values from cited sources are deflated by the IMF Japan CPI index. The social discount rate $d=0.03$ (Section 3.10) is applied to NPV computation.

Complete parameter list. 

Table 8. Complete numerical parameter list of DC-DECOM, organized by model component.

Component	Parameter	Value	Source/Notes
PUE module (Equations (8) and (9))
	$C_{\mathrm{base}}$	1.14	calibrated; IT/UPS/lighting overhead
	$\alpha$ (linear coeff.)	0.0167 °C−1	calibrated; 95% CI [0.013, 0.020]
	$\beta$ (quadratic coeff.)	0.0102 °C−2	calibrated; 95% CI [0.008, 0.013]
	$T_{\mathrm{thresh}}$	15 °C	free-cooling threshold [58]
	$C_{\mathrm{pump}}$	0.055	liquid cooling overhead [55]
	$\gamma$ (liquid temp. coeff.)	0.0015 °C−1	calibrated [23]
Demand module (Equations (6) and (7))
	$P_{\mathrm{AI},0}$ (2024 AI workload)	1.5 GW	inferred from 19 TWh load & 80% utilization
	$g_{\mathrm{AI}}$ (AI demand CAGR)	30%/yr	JST Vol. 3 Modest [7]
	${\eta }_{\mathrm{HW}}$ (CPU efficiency CAGR)	30%/yr	Moore-Hennessy [7]
	${\eta }_{\mathrm{arch}}$ (architecture factor)	1.05/yr	GPU-specific [27,54]
Carbon module (Equation (10))
	${\mathrm{EF}}_{\mathrm{grid}}$ (FY2022 national avg.)	0.43 kg-CO2/kWh	MOE [59]
	${\mathrm{EF}}_{\mathrm{fossil}}$ (CGS, natural gas)	0.49 kg-CO2/kWh	METI Cost Verification WG (Cogen) [73]
	$P_{\mathrm{carbon}}\left(2026\right)$	¥1700–4300/t-CO2	GX-ETS [16,17]
	$P_{\mathrm{carbon}}\left(2035\right)$	¥7000–15,000/t-CO2	MRI [50]
Constraints (Equations (13)–(17))
	$E_{2024}$ (base-year emissions)	8.2 Mt-CO2/yr	computed from base-year load × EF
	${\mathrm{RE}}_{max}$ (Hokkaido solar)	337 GW	REPOS [60]
	N (resilience redundancy)	1	N+1 standard [36]
	${\mathrm{Deploy}}_{max,\mathrm{liquid}}$	0.5 GW/yr/region	calibrated to SoftBank [29]
	${\mathrm{Deploy}}_{max,\mathrm{PV}}$	1.0 GW/yr/region	FIT historical [63]
Objective function (Equations (1)–(5))
	d (social discount rate)	0.03	Imagawa et al. [35], IEA-ETSAP [69]
	${\pi }_{\mathrm{curt}}$ (curtailment penalty)	¥500/kWh	VOLL proxy
	planning horizon	2024–2040 (17 years)	Green Growth Strategy [15]
	Representative hours/year	96 (24 × 4 seasons)	JMA preprocessing

consolidates every numerical parameter referenced in Equations (1)–(17). Region-specific parameters (electricity tariff, PPA price, solar potential) are in

Table 9. Regional input parameters.

Parameter	Hokkaido	Tohoku	Kanto	Kansai	Kyushu	Source
Annual mean T (°C)	9.2	12.8	16.5	17.1	17.3	JMA [65]
Base-year IT load share (%)	3	5	60	22	10	MIC [10,12]
Grid emission factor (kg-CO2/kWh)	0.49	0.52	0.44	0.36	0.37	MOE [59]
Solar potential (GW)	337	142	99	73	115	REPOS [60]
Grid tariff (¥/kWh)	20.0	18.5	19.9	16.0	15.5	ANRE [68]
PPA price (¥/kWh)	12	13	16	15	13	REI [63]

; technology unit costs are in Table 6.

Sensitivity to alternative PUE functional forms. To verify that the piecewise-quadratic functional choice in Equation (8) does not unduly influence the cost results, two alternative specifications were tested: (i) a purely linear PUE–temperature relationship (no quadratic term), and (ii) a purely quadratic relationship (no piecewise threshold). Re-running the R-CN scenario with each alternative yields total system costs within $\pm 2.5\%$ of the baseline; the C-CN–R-CN cost gap range remains 17–24% (vs. baseline 18–23%). The qualitative conclusion is robust to the specific PUE functional choice within plausible alternatives.

Table 9 summarizes the key region-specific input parameters used across all scenarios.

4. Optimization Simulation Results

4.1. Demand Projections

This section presents the quantitative results of the DC-DECOM optimization across the three scenarios defined in Section 3.9. Results are organized from the demand-side context (Section 4.1) through aggregate cost and emission outcomes (Section 4.2), decomposition of mitigation measures (Section 4.3), technology portfolio (Section 4.4), geographic load redistribution (Section 4.5), and finally sensitivity analysis (Section 4.6).

Figure 2 summarizes data center electricity demand projections from multiple sources used to calibrate the model. The demand trajectory enters the model through Equation (6), with regional disaggregation from MIC base shares and AI-component growth following Equation (7). Historical estimates are based on JST Vol.2/Vol.3 [6,7] and cross-validated against Wood Mackenzie [3]. The JST Modest scenario projects 24 TWh by 2030, while Wood Mackenzie projects 57–66 TWh by 2034. The divergence reflects different assumptions about AI adoption rates and efficiency improvements, with the JST estimate predating the generative AI surge of 2023–2024. The model uses the Wood Mackenzie range as the primary demand trajectory and the JST Modest case as a lower-bound sensitivity.

4.2. Total System Cost and Emissions

Figure 3 presents the annual system cost and CO₂ emission trajectories for the three scenarios over the 2025–2040 period. The annual system cost shown on the left axis is the value of the objective function in Equation (1), evaluated year-by-year as the sum of capital, operating, energy, carbon, and curtailment terms [Equations (2)–(5)]; the CO₂ emissions on the right axis are computed using the source-specific accounting in Equation (10) (excluding multiplication by $P_{\mathrm{carbon}}\left(t\right)$) and aggregated across regions; Equation (11) is used only for the diagnostic wedge decomposition in Section 4.3. The REF scenario yields the lowest annual system cost but produces CO₂ emissions that continue to grow with data center capacity expansion, reaching approximately 11 Mt-CO₂ by 2040—incompatible with Japan’s NDC commitment of 73% reduction by 2040 [14].

The C-CN scenario achieves net-zero by 2040 at a cost premium of ∼35–40% over REF (approximately ¥8.5 trillion NPV for C-CN versus ¥6.2 trillion for REF). This premium is driven primarily by the high cost of procuring sufficient renewable energy within the geographically constrained Kanto and Kansai regions. Kanto’s land-based solar potential of 98,833 MW [60] is limited relative to the concentrated demand, and grid electricity prices in Tokyo (¥19.9/kWh) are among the highest nationally [68]. The C-CN scenario relies heavily on off-site PPA procurement and grid-supplied renewable energy, both of which carry premium costs in land-constrained metropolitan areas.

The R-CN scenario achieves the same net-zero target at 18–23% lower total system cost than C-CN (approximately ¥6.8 trillion for R-CN versus ¥8.5 trillion for C-CN, an NPV saving of ¥1.5–2.0 trillion over 2024–2040). The cost advantage decomposes into four channels (

Table 10. Cost-driver decomposition of the C-CN → R-CN system-cost reduction in Japan (NPV, 2024–2040).

Cost-Driver Channel	Contribution	Mechanism
Cooling-energy reduction	∼35%	Hokkaido/Tohoku PUE advantage [Equations (8) and (9)] reduces facility electricity consumption [Equation (6)]
Lower RE procurement cost	∼30%	Hokkaido PPA price (¥12/kWh) below Kanto (¥16/kWh) for equivalent procurement volume [63]
Lower carbon cost	∼25%	Regional grid emission factor differential reduces $C_{\mathrm{carbon}}\left(t\right)$ [Equation (10)] under the GX-ETS price trajectory
Lower siting-related cost	∼10%	Reduced-form deployment-feasibility effect (regional land and infrastructure differentials), captured implicitly in the deployment-rate parameter rather than as a fully endogenous regional CAPEX coefficient

): cooling-energy reduction from the Hokkaido/Tohoku PUE advantage, lower regional PPA procurement cost, lower carbon cost under the GX-ETS, and lower siting-related cost (a reduced-form deployment-feasibility effect).

The four channels are conceptually independent but their numerical contributions are mildly correlated (e.g., cooling energy and carbon cost both decline together when load shifts from Kanto to Hokkaido), so the percentages in Table 10 are reported as approximate ranges rather than exact values.

4.3. Decomposition of Emission Reductions by Measure

Figure 4 presents a stabilization wedge decomposition of the R-CN scenario. The decomposition is constructed using the multiplicative emission identity in Equation (11): each wedge isolates the contribution of one of the three independent reduction channels (PUE improvement, renewable share, or grid emission factor) plus geographic redistribution and carbon pricing. Five colored wedges fill the growing gap between the REF trajectory (upper dashed boundary) and the R-CN net-emission line (lower red boundary). Each wedge is zero-width in 2024—when R-CN and REF are identical—and expands rightward as the corresponding technology is deployed. Together, the wedges account for the full 11 Mt-CO₂ abatement gap by 2040.

By 2040, the five measures contribute to net-zero achievement in the following proportions: liquid cooling technology (PUE improvement) accounts for approximately 36% of total emission reductions, renewable energy procurement (PPA and on-site PV) for 29%, geographic optimization (load redistribution to low-PUE, low-emission regions) for 22%, grid decarbonization (exogenous decline in regional emission factors per the 7th Strategic Energy Plan [62]) for 11%, and carbon pricing effects (GX-ETS-induced behavioral shifts) for 2%. The dominance of liquid cooling and renewable procurement confirms that technology investment decisions are the primary lever, while geographic optimization provides a substantial additional benefit that cannot be captured under the geographically constrained C-CN scenario.

4.4. Technology Portfolio Evolution

Figure 5 shows how cooling technology shares evolve over 2024–2040 in the R-CN scenario. The model selects liquid cooling (DLC or immersion) for all new high-density AI facilities in both carbon-constrained scenarios. By 2030, liquid cooling accounts for approximately 40% of new installed capacity; by 2035, this exceeds 70%. By 2040, air cooling is retained only for legacy facilities and low-density workloads.

This transition is driven by two reinforcing factors. First, the PUE advantage of liquid cooling (1.07–1.08 vs. 1.29–1.48 for air cooling) directly reduces energy cost and carbon emissions. Second, the physical requirements of next-generation GPU architectures make liquid cooling a necessity rather than an option: NVIDIA’s GB200 NVL72 rack system at 120 kW [27] cannot be adequately cooled by air. The technology selection therefore reflects both economic optimization and physical constraints.

BESS deployment reaches 2–4 GWh (0.5–1.0 GW assuming 4-h duration) nationally by 2040 in the R-CN scenario, serving as an enabler of renewable energy integration. The optimal BESS capacity is driven by the need to match variable PV and wind generation with the relatively flat 24/7 demand profile of data centers. This finding aligns with the complementary role of storage identified in broader energy system models by Imagawa et al. [35] and Nishikura et al. [36].

4.5. Geographic Load Redistribution

Figure 6 compares the 2024 baseline IT load distribution (Kanto 60%, Kansai 22%, Kyushu 10%, Tohoku 5%, and Hokkaido 3%, per MIC White Papers [10,12]) with the 2040 R-CN optimal allocation.

R-CN does not imply wholesale relocation: latency-sensitive workloads (financial services, real-time applications requiring sub-millisecond response) remain in Kanto near major Internet Exchange points, while latency-tolerant workloads (AI model training, batch data processing, backup, archival) are optimally relocated to regional sites where cooling and renewable advantages are greatest.

This workload-differentiated outcome aligns with the direction of the MIC’s Digital Infrastructure expert meetings [48] and with industry practice: SoftBank’s decision to build its 300 MW AI training-focused data center in Tomakomai, Hokkaido [29] exemplifies the model’s recommended strategy. The model’s optimal Hokkaido allocation (27% of national IT load by 2040) matches the scale of announced investments in the region.

4.6. Sensitivity Analysis

Figure 7 presents the tornado chart of one-at-a-time sensitivity analysis results, showing the impact of key parameter variations ($\pm 1$ standard deviation from baseline) on R-CN total system cost.

The most influential parameters are: (1) Carbon price trajectory (±15% cost impact): higher carbon prices accelerate the economic case for both regional decentralization and liquid cooling adoption. (2) Hardware efficiency improvement rate (±12%): faster-than-expected GPU efficiency gains (e.g., 40% annual vs. baseline 30%) reduce total electricity demand and narrow the cost gap between scenarios, as the absolute energy savings from geographic optimization diminish when total demand is lower. (3) PPA price trajectory (±8%): lower renewable energy costs disproportionately benefit the R-CN scenario, where access to cheaper regional PPA contracts is a key advantage. (4) Discount rate (±6%): higher discount rates favor lower-CAPEX conventional technologies, delaying the transition to liquid cooling and renewables.

Figure 8 shows the relationship between carbon price level and the cost advantage of R-CN over C-CN. At the current GX-ETS price range (¥1700–4300/t-CO₂), the R-CN advantage is approximately 12–15%. At ¥10,000/t-CO₂, the advantage exceeds 20%. At ¥20,000/t-CO₂, it approaches 28%. The nonlinear shape reflects the rising marginal cost of carbon abatement in the geographically constrained C-CN scenario.

The OAT analysis presented above varies one parameter at a time. To assess the joint effect of multiple adverse parameter movements, two preliminary combined-scenario stress tests are reported here. (i) Adverse case (high AI demand growth + slow hardware-efficiency improvement + delayed liquid cooling deployment + low carbon price): R-CN remains cost-advantaged over C-CN, but the advantage narrows to approximately 12%. (ii) Favorable case (low AI demand + fast hardware efficiency + early liquid cooling + high carbon price): the R-CN advantage expands to approximately 28%. Across both stress tests, the sign of the R-CN advantage is preserved, supporting the qualitative robustness conclusion summarized in Section 5.1. A more comprehensive stochastic treatment—spanning the joint distribution of all four uncertain parameters via Monte-Carlo sampling—is identified in Section 5.5 as a priority for follow-up work.

In Figure 8, the orange band indicates the GX-ETS FY2026 price range (¥1700–4300/t-CO₂) [17], and the dashed line marks the ¥10,000/t-CO₂ threshold above which the R-CN advantage exceeds 20%, consistent with the MRI carbon price forecast for the mid-2030s [50].

5. Discussion

5.1. Robustness of Findings

Before discussing policy interpretations, we summarize the robustness of the central quantitative findings of this study to the assumptions documented in Section 3 and Section 4. The 18–23% R-CN cost advantage over C-CN is itself reported as a range derived from the one-at-a-time sensitivity analysis (Section 4.6), with the lower bound corresponding to combinations of low carbon-price and faster-than-baseline hardware-efficiency improvement (which compress demand and the value of geographic arbitrage) and the upper bound to combinations of higher carbon-price and slower hardware-efficiency improvement (which amplify the differences in cooling, renewable-procurement, and carbon costs across regions). Within this range, three observations support the qualitative conclusion:

• Sign-stability: The R-CN advantage remains strictly positive across all $\pm 1\sigma$ parameter variations reported in Figure 7. No single parameter excursion within the documented uncertainty bounds reverses the qualitative conclusion that geographic optimization reduces total system cost.
• Driver-decomposition: The advantage is decomposed into three structurally independent channels via Equation (11) (lower PUE, lower grid emission factor, higher renewable share). For the qualitative conclusion to be reversed, all three channels would have to be simultaneously and adversely mis-estimated; the multiplicative structure of the identity makes this jointly unlikely.
• Cross-validation against empirical PUE benchmarks: The dynamic PUE sub-model is corroborated against independent measurements from MHI [22], Haghshenas et al. [23], NTT Facilities [58], and the Uptime Institute [18] (Table 5); all five model outputs lie within the empirical range. This independent corroboration mitigates concern about overfitting on the limited five-point regression basis.

The conclusions are therefore conditional on the model’s assumptions but robust within their documented uncertainty ranges. We acknowledge that combined-scenario stress tests (e.g., simultaneous adverse movements across multiple parameters) and stochastic representations of demand and price uncertainty would further strengthen the quantitative claims; we identify these as priorities for future model extensions in Section 5.5.

5.2. Policy Implications

Subject to the robustness conditions discussed above, the quantitative results of DC-DECOM are consistent with the following policy interpretations, which address the structural tension between AI infrastructure expansion and decarbonization described in the MIC White Papers [10,11,12,13]. We frame these as conditional inferences from the model rather than direct prescriptions; institutional, regulatory, and engineering analyses beyond the scope of this study would be required before implementation. We summarize four policy interpretations as follows:

• Workload-differentiated location subsidies: The 18–23% cost advantage of R-CN over C-CN, conditional on the model assumptions documented in Section 3 and Section 4, is consistent with the government’s geographic decentralization strategy [48] on economic grounds in addition to resilience grounds. The model further suggests that the cost-optimal decentralization pattern is workload-specific rather than uniform; we note, however, that workload-differentiated outcomes in DC-DECOM emerge from a binary latency-sensitive/latency-tolerant split (Section 4.5) rather than a richer workload taxonomy, so the strength of the workload-differentiation conclusion is correspondingly limited. The model results suggest that subsidies for regional data center construction would be most cost-effective if they prioritized facilities designed for latency-tolerant AI training workloads, where the geographic arbitrage is greatest. The SoftBank Tomakomai project [29] exemplifies this approach. This implies that, if the cost-optimal pathway is to be realized, policy incentives could usefully be differentiated: higher subsidy rates for AI training facilities in cold-climate, renewable-rich regions, and lower rates for general-purpose facilities that may need metropolitan proximity.
• Quantifying the carbon premium of AI infrastructure: The model quantifies the additional decarbonization cost attributable to AI-driven demand growth—the “carbon premium” of AI expansion. This directly addresses the policy tension between the Cloud Supply Security Plan [46], which subsidizes GPU deployment and thereby increases electricity demand, and the Green Growth Strategy [15], which targets ICT carbon neutrality by 2040. The JST reports [5,6,7,8,9] document the scale of this tension: without efficiency measures, IT-related electricity consumption could reach 1480 TWh by 2030 [5], far exceeding Japan’s total current electricity consumption. The carbon premium quantified by DC-DECOM can inform the allocation of AI infrastructure subsidies toward complementary renewable energy and efficiency investments, as proposed by Kobayashi et al. [74] in their analysis of carbon neutrality economics.
• Credible carbon price signals: The sensitivity of results to carbon price trajectories (Figure 8) highlights the importance of credible, predictable carbon pricing signals. The current GX-ETS price range of ¥1700–4300/t-CO₂ [17] produces only a modest R-CN advantage of 12–15%. Carbon prices in the range of ¥10,000–20,000/t-CO₂, consistent with the MRI forecast for the mid-2030s [50], would be required for the economic case for optimal decarbonization investments to be clear to private-sector decision-makers, conditional on the model’s other assumptions. This aligns with Green’s [52] analysis in Environmental Research Letters that low carbon prices have limited effectiveness in triggering zero-carbon investments, and with Ding’s [51] finding of heterogeneous carbon cost pass-through across Japan’s regional electricity markets.
• The liquid cooling paradigm shift: The technology portfolio results (Figure 5) indicate that liquid cooling is not merely an efficiency improvement but a structural necessity for AI-era data centers. The Green Growth Strategy’s target of 30% energy efficiency improvement for new data centers by 2030 [15] and the benchmark system’s PUE target of 1.4 [20] are achievable only with widespread liquid cooling adoption. Policy support directed at accelerating liquid cooling deployment—through technology standards, workforce training, and CAPEX subsidies—is therefore likely to be more cost-effective than incremental improvements to air-cooling systems that have reached diminishing returns [18].

Beyond the direct policy implications for Japan, these findings carry broader significance for the global sustainability transition. Data centers are among the fastest-growing sources of electricity demand worldwide [1], and the methodological framework developed here—integrating dynamic PUE modeling, geographic optimization, and carbon pricing within a single cost-minimization model—is transferable to other countries facing similar trade-offs between digital infrastructure expansion and climate commitments. The finding that geographic arbitrage in cooling efficiency and renewable energy access can reduce decarbonization costs by 18–23% is particularly relevant for countries with pronounced regional climate variation, such as China, India, and the United States.

5.3. Model Validation

The PUE sub-model was calibrated by least-squares regression against five regional empirical data points (Table 5), yielding $R^2=0.988$. As a cross-consistency check, the air-cooling model yields PUE ≈ 1.58 at $\overline{T}\approx 18.6$ °C, consistent with the global average PUE of 1.58 reported by the Uptime Institute [18] for a mix of facilities in warm-climate regions. The liquid cooling model yields a PUE of 1.07–1.08 across Japan’s temperature range, corroborating Haghshenas et al.’s [23] measurements and the MHI demonstration [22].

Demand projections were calibrated against multiple independent sources. The base-year (2024) national IT load of 19 TWh is from Wood Mackenzie [3], cross-validated against the JST Vol.2/Vol.3 estimates [6,7]. The growth trajectory (57–66 TWh by 2034) is consistent with the upper range of JST Vol.3 projections when updated for post-2023 generative AI adoption rates.

The carbon price trajectory follows the GX-ETS official schedule [16,17] for the near term (¥1700–4300/t-CO₂ in FY2026) and the Mitsubishi Research Institute forecast [50] for the medium term (¥7000–15,000/t-CO₂ by the mid-2030s). Technology cost parameters are sourced from government reports: PV LCOE from the METI Power Generation Cost Verification Report [66], BESS costs from the NEDO lithium-ion battery technology report [67], and CGS parameters from the METI Power Generation Cost Verification Working Group (Cogeneration and Fuel Cells) [73].

A key limitation of the validation is that full model validation against historical outcomes is not yet possible, since the GX-ETS only became mandatory in April 2026 and the planning horizon extends to 2040. The model’s projections should therefore be interpreted as conditional scenarios rather than point forecasts: they indicate the cost-optimal response to specified policy and technology trajectories, subject to the assumptions documented in Table 6, Table 7, Table 8 and Table 9.

5.4. Academic Contributions and Novelty

DC-DECOM advances beyond prior discrete scenario analyses in several important respects. Earlier studies evaluated pre-specified combinations of cloud migration rates and conventional efficiency improvements, using fixed PUE values and without geographic optimization. Such analyses typically found that only the most aggressive assumptions marginally exceeded emission reduction targets. DC-DECOM endogenizes these choices within a cost optimization framework, allowing the model to determine the optimal combination rather than evaluating pre-specified alternatives. The finding that only the most aggressive prior scenarios marginally exceeded targets corroborates DC-DECOM’s result: net-zero by 2040 requires simultaneous deployment of geographic optimization, liquid cooling, and aggressive renewable procurement—no single measure alone appears sufficient under the modeled assumptions.

DC-DECOM also incorporates the demand projections from the JST report series [5,6,7,8,9], which were not available or not fully utilized in prior modeling work. The JST Vol.1 finding that IT-related electricity consumption could reach 176,200 TWh by 2050 without efficiency measures [5] provides the upper bound that motivates the urgency of the optimization problem. The JST Vol.3 Modest scenario of 24 TWh by 2030 and 500 TWh by 2050 for data centers [7] provides the demand trajectory against which the model’s technology and siting decisions are optimized.

Compared to the global-scale analysis of Masanet et al. [31], DC-DECOM provides Japan-specific regional granularity in terms of climate data, grid emission factors, and policy instruments. Compared to the U.S.-focused LBNL models [2,32], it incorporates carbon pricing as an endogenous cost driver reflecting Japan’s GX-ETS institutional context. Compared to Hao et al.’s [34] multi-data-center operational optimization, DC-DECOM adds the investment dimension (CAPEX decisions over a 16-year horizon) rather than optimizing only operational dispatch. Compared to Koronen et al.’s [33] analysis of European data centers, DC-DECOM incorporates Japan-specific policy instruments and regional energy characteristics.

5.5. Extensibility and Future Prospects

Several limitations should be acknowledged. First, the model uses representative hourly profiles rather than full 8760-h chronological optimization, which may underestimate the value of BESSs in managing renewable intermittency and may not fully capture seasonal variations in cooling load. Second, network latency costs are modeled as a simplified distance-based function rather than a detailed network topology model incorporating Internet Exchange point locations and backbone capacity. Third, Water Usage Effectiveness (WUE) constraints, analyzed by Ristic et al. [75] in Sustainability, are not included; water scarcity may constrain cooling technology choices in some regions. Fourth, data center waste-heat-recovery revenue, analyzed by Wahlroos et al. [76] in Energy, is not modeled; incorporating this could improve the economics of urban data centers. Fifth, the potential impact of next-generation optical networking technologies (e.g., NTT’s IOWN) on energy efficiency, discussed by Ozaki and Iwatsuki [41], is not captured in the current model.

In addition, several geographic and infrastructural factors that affect data-center siting decisions in practice are not formally constrained in DC-DECOM. The R-CN allocation should therefore be interpreted as conditional on these factors not being binding. The principal omitted factors are:

• Land availability and zoning. The R-CN scenario relocates approximately 25–30% of national IT load to Hokkaido and 15–20% to Kyushu by 2040. Hokkaido’s industrial-zone land is generally sufficient at this scale, but local zoning and grid-connection permitting may become binding for individual projects.
• Fiber connectivity and latency. Hokkaido–Kanto trunk lines have a one-way latency of approximately 8–10 ms, adequate for AI training and batch workloads but insufficient for sub-millisecond real-time inference; the model represents this only via the qualitative workload separation in Section 4.5.
• Skilled labor availability. Sapporo and Fukuoka have growing IT clusters but smaller talent pools than Kanto; this is a binding constraint for staffed operations and is not represented.
• Seismic and natural-disaster risk. Hokkaido is in seismic risk zone 3 (lower than Kanto/Tokai), which favors decentralization, but specific tsunami and volcanic-ash exposures are not modeled.
• Water availability. Liquid cooling reduces water use compared with evaporative cooling, but data centers remain non-trivial water consumers; regional water stress is not constrained.
• Workload categories beyond the binary latency split. The model distinguishes only “latency-sensitive” from “latency-tolerant” workloads; finer categories (regulatory data residency, sub-second inference, etc.) would require a richer workload representation.

We do not claim that the cost-optimal R-CN allocation is automatically the policy-implementable allocation; the omitted factors above would be addressed in a multi-criteria extension of DC-DECOM that adds them as soft penalties or hard constraints on a per-project basis.

Future research directions include: (1) extending the geographic resolution from five regions to 47 prefectures to capture intra-regional variation; (2) incorporating WUE constraints and water cost; (3) modeling waste heat recovery as a revenue stream that could partially offset urban data center operating costs; (4) coupling DC-DECOM with a national energy system model to capture feedback effects between data center demand growth and grid capacity planning, building on the methodology of Imagawa et al. [35]; (5) incorporating the impact of IOWN and other next-generation networking technologies on the ${\eta }_{\mathrm{arch}}$ parameter; and (6) validating model projections against empirical data as Japan’s data center decarbonization policies are implemented and facilities such as the SoftBank Tomakomai data center [29] become operational; and (7) extending DC-DECOM to a two-stage stochastic optimization formulation in which AI demand growth, PPA price, and carbon price are treated as random variables with first-stage capacity decisions and second-stage dispatch under realized scenarios, building on the Monte-Carlo framework outlined in Section 4.6.

6. Conclusions

This study introduced DC-DECOM, a multi-regional bottom-up cost optimization model for evaluating decarbonization pathways for Japan’s data center industry toward the 2040 carbon-neutrality target. The model’s key methodological contribution is the endogenous treatment of PUE as a dynamic function of ambient temperature and cooling technology, calibrated against published empirical data from the Uptime Institute, NTT Facilities, Haghshenas et al., and MHI. The model integrates demand projections from the JST report series, technology cost data from government sources, and the GX-ETS carbon pricing trajectory within a unified cost-minimization framework.

Four principal findings emerge from the scenario analysis. First, a regional decentralization strategy (R-CN) achieves the 2040 net-zero target at 18–23% lower total system cost than a metropolitan-concentrated strategy (C-CN), shifting approximately 25–30% of IT load to Hokkaido and 15–20% to Kyushu while retaining latency-sensitive workloads in Kanto. Second, the cost gap decomposes into four channels (Table 10): cooling-energy reduction (∼35%), lower regional renewable-procurement cost (∼30%), lower carbon cost under the GX-ETS (∼25%), and lower siting-related cost (∼10%). Third, liquid cooling is selected as the dominant cooling solution for new high-density facilities across all carbon-constrained scenarios, reflecting both economic optimization and the physical requirements of next-generation GPU architectures. Fourth, carbon prices in the range of ¥10,000–20,000/t-CO₂ would be required to make the economic case for cost-optimal decarbonization investments clear to private-sector actors, substantially above the current GX-ETS price range of ¥1700–4300/t-CO₂. The qualitative conclusion (positive R-CN advantage) is preserved across all OAT ±$1\sigma$ parameter variations and across two combined-scenario stress tests reported in Section 4.6.

These findings embed technology and location choices within a unified optimization framework, incorporate demand projections from the JST report series, and provide quantitative evidence for the policy directions articulated in the MIC White Papers and the Green Growth Strategy. DC-DECOM thus offers an analytical tool for policymakers and industry stakeholders evaluating technology investment and siting strategies in the context of Japan’s AI infrastructure expansion and climate commitments.