Maximizing Renewable Energy Use in Datacenters through Watt-bit Collaboration

1. Background

The proliferation of large-scale data processing, including generative artificial intelligence (AI), has rapidly increased the power demand of datacenters (DCs). Advanced learning and inference processes, such as natural language processing models and image generation models, require enormous computational resources and consume power on a scale orders of magnitude greater than conventional systems. For example, the power consumption of training OpenAI’s GPT-3 language model, which powers ChatGPT, is estimated to be 1287 MWh [1]―equivalent to the electricity generation of a single nuclear power plant in one hour. As generative AI becomes more widespread, DC power demand is expected to grow even further.

Across society, momentum for decarbonization is growing, driving rapid adoption of renewable energy (RE) sources such as solar and wind power. In the 7th Strategic Energy Plan [2], the Japanese government has set a policy to maximize the introduction of RE as a major power source, aiming to increase its share in the power supply mix from 21.8% in fiscal year 2022 to 40–50% by fiscal year 2040. Within the NTT Group, “NTT Green Innovation toward 2040” [3] serves as the environmental and energy vision. To achieve carbon neutrality for mobile (wireless base stations) and DCs by 2030, the Group is actively promoting adoption of RE. This is expected to enable the effective RE utilization to meet the growing power demand of expanding DCs.

However, a problem has emerged where RE cannot be fully used. The potential for RE deployment is concentrated in regions such as Hokkaido, Tohoku, Chugoku, and Kyushu, which are blessed with vast land and favorable natural conditions. Conversely, DCs are clustered around major metropolitan areas, such as Tokyo and Osaka, due to considerations of communication latency and maintenance convenience. This results in a geographical mismatch between the potential for RE deployment and the power demand of DCs. Solar and wind power generation exhibit significant fluctuations in output depending on weather conditions, season, and time of day. In contrast, DCs require consistent operation 24 hours a day, 365 days a year, leading to relatively stable power demand. This creates a temporal mismatch as well.

Due to this geographic and temporal mismatch, output curtailment of RE (suppressing RE generation) is implemented in regions and time periods where power supply, including RE, exceeds demand. This is done to ensure stable operation of the power grid. Figure 1 shows the trend in RE output curtailment in Japan. In fiscal year 2024, over 2 billion kWh of RE was subject to output curtailment, concentrated particularly in Kyushu and Tohoku, regions with high potential for RE deployment. As RE deployment continues to expand, the occurrence of this output curtailment will become an unavoidable issue. Therefore, it is essential to establish technologies and reform systems to resolve this mismatch between RE output and DC power demand, thereby maximizing RE utilization. Currently, plans [4] for developing transmission networks to enable interregional power sharing have been outlined. However, these require enormous investments in the trillions of yen and take many years to implement. They alone cannot achieve maximum RE utilization or stable power supply. A new mechanism is required to resolve this mismatch. This involves not only the supply side but also the demand side, such as DC systems, operating flexibly in response to RE generation conditions to adjust the supply-demand balance.

Fig. 1. Trends in RE output curtailment in Japan.

2. The concept and significance of watt-bit collaboration

Against this background, the concept of “watt-bit collaboration” aims to build a sustainable and efficient social infrastructure by highly integrating energy (watts) and information and communication (bits), thus advancing the development of power and information and communication infrastructure in a unified manner. This concept was proposed in the “GX2040 Vision” presented in February 2025. The Public-Private Advisory Council on Watt-Bit Collaboration, comprising the Ministry of Internal Affairs and Communications (MIC), the Ministry of Economy, Trade and Industry (METI), and power and telecommunication companies, is working to establish policies and implementation approaches for its actualization [5].

On June 12, 2025, as a concrete outcome of these discussions, three key directions were presented, as shown in Table 1. Regarding the first direction, “Addressing Immediate DC Demand,” discussions focused on promoting the siting of large-scale DCs in areas with surplus grid capacity, referred to as “Welcome Zones,” as an immediate countermeasure to power shortages in large-scale DCs in major urban areas such as those represented by cloud-based DCs. Research and development on flexible operation and the establishment of its foundational infrastructure are also being discussed. This includes envisioning the development of next-generation network infrastructure, such as the All-Photonics Network (APN), to achieve both high access performance and energy efficiency for large-scale DCs. The second direction, “Realizing New DC Aggregation Hubs,” envisions establishing core DCs (hubs aggregating multiple large-scale DCs) outside major cities. Discussions focus on selecting new large-scale DC aggregation hubs by comprehensively considering factors such as power infrastructure expansion potential, RE integration feasibility, securing necessary communication infrastructure, transportation accessibility, and disaster risks. The third direction, “Promoting Regional Dispersal and Advancement of DCs,” is positioned as a key measure to enhance the supply-demand balance. Dispersing small-to-medium-scale DCs, such as container-type facilities, in regional areas with high RE potential and improving RE utilization through workload (WL) distribution and control is expected to alleviate grid congestion and reduce output curtailment.

Table 1. Key directions for watt-bit collaboration and implementation timeline.

NTT Space Environment and Energy Laboratories is advancing two approaches as technological development to support the regional decentralization and operational advancement of large- and small-scale DCs: decentralized control of DC processing loads based on RE generation status, and DC placement tailored to RE introduction regions. The following sections introduce the specific research content of these approaches.

3. Approach 1: Integrated control of energy and ICT resources for WL distribution

Adjusting DC power demand according to RE generation enables supply-demand optimization and expanded RE utilization. For example, during times or in regions with excess RE, shifting processing WL from other DCs enables utilization of this energy. Conversely, during shortages, shifting WL to other DCs can increase RE usage. However, flexibly relocating a DC’s WL is not straightforward. Relocation involves multifaceted constraints such as communication quality, cost, and security. Since both power demand and RE generation fluctuate over time, control must be adapted to the power supply-demand conditions.

Thus, to appropriately adjust DC power demand on the basis of RE generation conditions, establishing advanced core technologies and enabling technological collaboration are essential. These include highly accurate forecasting of time-varying DC power demand and RE generation, along with real-time, optimal control of WL allocation between DCs. To address these challenges, multiple NTT laboratories, including our laboratory, are conducting cross-organizational research on integrated energy/information and communication technology (ICT) resource control technology to achieve flexible load control based on RE generation conditions. The overall concept and key technologies are shown in Fig. 2. These key technologies enable the integrated coordination of ICT resources and energy resources such as RE generation and storage batteries across DCs distributed nationwide and connected via the APN. The goal is to effectively use RE and optimize the power supply-demand balance by flexibly controlling DCs’ WL and battery charging/discharging according to RE generation conditions.

Fig. 2. Overview of integrated energy/ICT resource control.

This integrated control consists of three steps: forecasting, optimization, and control. Each step is coordinated and managed by an orchestration platform^*1. First, RE generation and power demand at each DC are predicted with high accuracy on the basis of nationwide weather forecasts and historical data. Next, on the basis of these forecasts, the energy demand per DC—including WL and battery charging/discharging volumes—is optimized to maximize RE utilization. The WL deployment plan and network routes for WL shifting to meet this demand are optimized. Finally, on the basis of the obtained optimization plan, WL shifting is executed in a virtual environment, while battery control is executed independently. This enables DC operation that maximizes RE utilization.

At our laboratory, we are focusing on the optimal energy demand setting technology. This involves determining the type of WLs to shift, their destination, the amount to be shifted, and the timing of the shift, as well as the charge/discharge levels of batteries co-located with DCs, based on forecasts of the next day’s RE generation and DC power demand. Regarding WL shifting, sending data to DCs in regions abundant in RE via communication networks and executing computational processing can shift power demand, thus potentially resolving the geographical mismatch between RE and power demand. However, WLs requiring high real-time performance (e.g., remote desktops or AI inference tasks) cannot be carried over to the next day, so this does not resolve the temporal mismatch. On the other hand, battery charging and discharging effectively resolves the temporal mismatch between RE and power demand such as charging power from RE and discharging it the following day. However, transferring RE stored in batteries to another DC via the power grid is difficult due to the costs of grid reinforcement and transmission losses, thus does not resolve the geographical mismatch. We are conducting research on an energy demand optimization model that combines these two approaches to simultaneously resolve both geographical and temporal mismatches, thus maximizing RE utilization.

This optimization model is based on reinforcement learning, which takes time-series data of RE generation and DC power demand, along with initial battery charge/discharge levels, as inputs. It outputs the optimal amount of WL shifting and battery charge/discharge levels to maximize RE utilization [6]. This model has two key features. First, it defines unique reward functions for the controlled variables—the amount of WL shifting and battery charge/discharge level—and integrates them as a weighted sum, enabling simultaneous optimization of both. The second is that by designing the time window for input data to include forecast values several days ahead, it enables forward-looking decisions. For example, even if the current day is sunny but rain is forecast for the next day, the model can prioritize charging the battery over accepting WL. These features enable the model to output control policies for WL and the battery that maximize RE utilization throughout the year.

An example demonstrating the effectiveness of the optimization model through simulations at three DC sites is presented. First, the optimization model was constructed using tens of thousands of patterns of RE generation data—randomly combining sunny and cloudy day solar radiation for 365 days—and a single representative DC demand curve pattern [7] as training data. Simulation conditions were determined on the basis of actual facilities: solar photovoltaic generation was 1.8 MW, maximum DC power consumption was 1 MW, and battery capacity was 5 MWh. These conditions were identical across all three sites. The three sites were selected for their differing climatic conditions: Hokkaido, Tokyo, and Kyushu. Next, we input the 365-day solar radiation data for the three locations and the aforementioned DC demand curve into the constructed model to optimize the amount of WL shifting and battery charge/discharge levels. Finally, we compared the RE utilization rates under three cases: controlling only the battery, controlling the WL and battery with a rule-based model^*2, and controlling them optimally with this optimization model. The comparison results are shown in Fig. 3. Compared with controlling only the battery, controlling both WL and the battery with the rule-based model alone improved the RE utilization rate by over 15%. Control with the optimization model confirmed an additional improvement of over 8% in the RE utilization rate. Specifically, compared with the rule-based model that takes into account instantaneous RE utilization rate maximization, this optimization model learned to maximize the RE utilization rate throughout the year. This enabled forward-looking control, achieving a further increase in the RE utilization rate.

Fig. 3. RE utilization rate simulation results.

4. Approach 2: Optimal placement of future DCs

As we expand RE toward carbon neutrality, determining where to establish DCs is becoming increasingly complex and challenging. DCs are currently concentrated in major metropolitan areas centered in Tokyo and Osaka, with development plans primarily based on land acquisition feasibility and business viability. However, with an eye toward the future decentralization of DCs, it is necessary to formulate strategic deployment plans that incorporate more diverse perspectives such as the potential for RE introduction, status of power infrastructure, and performance of interregional networks.

Thus, to formulate deployment plans that comprehensively evaluate multiple perspectives, the challenge lies in establishing technology that captures diverse regional conditions using various evaluation axes and quantitatively supports the selection of optimal DC candidate sites. To address these challenges, our laboratory has launched research on DC placement optimization technology to enable strategic DC deployment decisions. While the aforementioned integrated control of energy and ICT resources focuses on dynamic WL optimization, this approach aims for static placement optimization from a medium-to-long-term perspective.

This technology aims to establish a decision support model that objectively evaluates and selects the optimal DC placement pattern from multiple candidate sites. It does this by first scoring evaluation items that vary by region (e.g., land prices, construction unit costs, RE generation potential, power grid congestion levels, communication line quality/latency, and disaster risk) then weighting each item. For instance, initial costs related to land and construction are evaluated using regional land price data and construction unit costs. RE factors are estimated from open meteorological data (e.g., New Energy and Industrial Technology Development Organization) and time-series generation potential data. Grid congestion is quantified using publicly available connection capacities and reverse power flow records. Communication network requirements reflect APN connection latency and regional fiber-optic deployment status. We aim to use these pieces of information to develop designs that contribute to the formulation of realistic deployment plans.

While advancing the development and accuracy improvement of these evaluation items, we aim to design an evaluation model that allows for flexible trade-offs between multiple factors such as RE utilization, communication performance, and total cost. Our goal is to develop a decision support algorithm that can freely adjust evaluation axes and weightings according to use cases and operator needs, thus supporting the optimal DC deployment.

5. Future outlook

The two approaches introduced in this article by our laboratory both aim to optimize the supply-demand balance for RE output and DC power consumption through watt-bit collaboration. They are aimed to maximize RE utilization by preventing waste, with each tackling the geographical and temporal mismatch between RE output and demand through distinct approaches. Moving forward, alongside deepening each technology, establishing an environment conducive to more realistic implementation is required.

Approach 1 relies on the cross-cutting integration of multiple technological elements. For instance, high-precision forecasting technologies for RE and power demand, optimization algorithms, control execution in virtual environments, and orchestration platforms are all essential; the absence of any one makes implementation difficult. Therefore, future efforts will focus on further advancing these individual technological elements while promoting collaboration across multiple research departments and specialized fields to build a practical operational foundation.

Regarding Approach 2, while it is under consideration, numerous challenges remain for implementation. Specifically, the input data spans diverse areas such as RE generation potential, land costs, communication environments, grid congestion, and maintenance personnel distribution. To comprehensively cover these factors, establishing a data-sharing mechanism on the basis of collaboration with regional power companies, telecommunication providers, administrative bodies, and other relevant stakeholders is essential. The objective function for optimization is not uniform; the prioritized factors (RE utilization, initial investment, communication performance, etc.) are significantly impacted by the envisioned use cases and the values of the operators. Therefore, the key going forward is algorithm design that allows for flexible switching of evaluation axes and weights according to the intended use, rather than relying on a single scoring standard.

NTT Space Environment and Energy Laboratories will continue to promote the effective use of RE through the watt-bit collaboration, working to integrate energy and ICT with the aim of achieving a sustainable society. By expanding the application area beyond DCs to the entire communication infrastructure, including wireless base stations, we will support the local production for local consumption of energy and a distributed society, accelerating the greening of society as a whole from both the power and communication perspectives. With an eye on the future of networks and energy, we will work to build a technological foundation that supports a sustainable society in harmony with the environment.

References

[1]	Stanford AI Index Report 2023, https://hai.stanford.edu/ai-index/2023-ai-index-report
[2]	Agency of Natural Resources and Energy, “The 7th Strategic Energy Plan,” Feb. 2025. https://www.enecho.meti.go.jp/en/category/others/basic_plan/pdf/7th_outline.pdf
[3]	Press release issued by NTT, “NTT Group’s New Environment and Energy Vision ‘NTT Green Innovation toward 2040,’” Sept. 28, 2021. https://group.ntt/en/newsrelease/2021/09/28/210928a.html
[4]	Organization for Cross-regional Coordination of Transmission Operators, Japan, “Transmission Infrastructure Expansion Plan,” Mar. 2023 (in Japanese). https://www.occto.or.jp/kouikikeitou/chokihoushin/files/chokihoushin_23_01_03.pdf
[5]	MIC and METI, “Report 1.0 of the Public-Private Advisory Council on Watt-Bit Collaboration,” June 2025 (in Japanese). https://www.soumu.go.jp/main_content/001014454.pdf
[6]	M. Suwabe, M. Kozai, and K. Akiyama, “Long-term Optimization of Data Center Workload Placement and Battery Charge/Discharge Control,” IEICE Society Conference 2024, B-9-03, Saitama, Japan, Sept. 2024.
[7]	M. Macías and J. Guitart, “SLA Negotiation and Enforcement Policies for Revenue Maximization and Client Classification in Cloud Providers,” Future Gen. Comp. Sys., Vol. 41, pp. 19–31, 2014. https://doi.org/10.1016/j.future.2014.03.004