Scalable data centers – Power generation and delivery challenges and solutions

• Converting data centres from air-to-chip (ATC), to liquid-to-chip (LTC) could reduce peak energy demand per kWhr of compute by 6%–14% • Annual energy consumption, carbon emissions, and power usage efficiency (PUE) may be improved by 4%–13% per kWhr of compute by data centre conversion from ATC to LTC • Increasing fluid differential and supply temperatures for LTC systems can eliminate the need for high energy use refrigerant compressor chillers in Australian and Asian climates. • Applying the novel control method to increase fluid differential temperatures from 5°C to 10°C based on I.T. kW measurement for LTC data centres could improve total data centre efficiency between approximately 3%∼6%, reduce PUE from a range of 1.22∼1.25 to 1.18 and potentially reduce central plant energy 18∼28%. The rapid and large-scale development of data centres to manage computational demand is resulting in potential strain on electrical energy and water utilities globally. New methods for minimising the peak power demand and annual energy of data centres that can be adapted to the different ambient conditions of various geographic locations are urgently needed. This study applies a novel advanced method of energy savings by differential temperature control via I.T. kW load measurement, to the conversion of data centres from conventional air-to-chip (ATC) cooling, to new liquid-to-chip (LTC) cooling. The novel method potentially assists elimination of energy-intensive chillers in favour of free cooling, reduces peak power demand and annual energy consumption for temperate to tropical ambient conditions. A new energy model was developed in Design Builder TM and EnergyPlus TM and was validated by onsite measurements at a state-of-the-art data centre in Melbourne, Australia and utilised to predict and optimize the improvement in energy efficiency by controlling differential temperature. The results indicate that converting from ATC cooling to LTC cooling could reduce the total peak power demand by 6%∼14% and reduce the annual energy consumption, carbon emissions and power usage efficiency by 4%∼13%. Application of the advanced control method to increase differential temperature from 5°C to 10°C for LTC systems, predicted reduction of the PUE from a range of 1.22∼1.25 to 1.18, potentially contributing an 18∼28% reduction in central plant energy.

The surge in generative artificial intelligence (AI) may cause growing conflicts between securing electricity supplies and achieving sustainable development goals. Here, we propose off-grid hybrid wind-solar-storage (WSS) systems, which leverage the immense renewable resources in desert areas alongside relatively low-cost fiberoptic connectivity of data centers, to address this challenge. Using a global high-resolution techno-economic optimization model, we demonstrate that well-planned WSS systems can deliver cost-effective 24/7 uninterrupted power, primarily tailored for energy-intensive, latency-tolerant foundation model training. Furthermore, our analysis reveals that regions proximate to load centers can also support latency-sensitive inference tasks. This energy supply is capable of delivering 1 PWh globally in 2030 at levelized costs of around $39/MWh, meeting the forecasted electricity demand for AI by the International Energy Agency. Moreover, even if AI electricity demand increases 10-fold, reaching 10 PWh by 2030, the unit cost increment would be less than 20%. Further uncertainty analysis shows that under extreme investment (3.0×) and cooling (2.0×) cost assumptions for data centers operating with the desert WSS systems, this 10-PWh/yr demand could still be satisfied at competitive cost levels. The desert WSS systems could potentially align computational and clean energy infrastructure in the AI era, as well as simultaneously achieving decarbonization and ecological restoration. Broader context: The rapid expansion of artificial intelligence (AI) poses emerging challenges to electricity supply and climate goals. Addressing this critical energy-computing nexus, we investigate a potential solution: co-locating data centers with off-grid wind-solar-storage systems in desert regions. By utilizing efficient fiberoptic data transmission to circumvent the challenges of long-distance power transport, this approach offers a possible pathway to convert renewable resources into computational power. Our analysis suggests that this strategy could help meet growing AI energy demands in an economically viable and environmentally sustainable manner.

Efficient coordination of heterogeneous AI chips (GPU/NPU/DCU) in data centers is crucial for Large Language Model (LLM) training, but this process is hindered by architectural mismatches, protocol fragmentation, and network partitioning. Existing solutions fail to achieve unified resource management across different chips, resulting in severe resource fragmentation and reduced allreduce efficiency. To overcome these limitations, this paper proposes the unified coordination framework UniOrch, which not only integrates three core functionalities, including hardware abstraction, software standardization, and communication coordination, but also enables the training and inference of large models across heterogeneous AI chips. UniOrch's hardware-agnostic bare-metal cloud eliminates virtualization overhead through Border Gateway Protocol Ethernet Virtual Private Network (BGP EVPN) overlay networks and gateway-based chip integration; its PyTorch-based adaptation layer masks hardware differences and reduces migration costs; the Transformer Collective Communication Library (TCCL) unifies NCCL, HCCL, and OpenMPIprotocols to support seamless hybrid parallel training. Furthermore, the framework ' score scheduling mechanism is the Heterogeneous Hybrid Estimation Model (HHEM), which employs a two stage cost model combining static analysis with dynamic runtime feedback to dynamically allocate computing power based on Transformer task loads, ensuring cross-chip task synchronization, resource pooling, and dynamic allocation. Deployment verification in real-world production environments (e.g., China Construction Bank) shows that UniOrch achieves significant improvements: resource utilization of heterogeneous AI infrastructure is increased by 35%, cross-chip latency is reduced by 42%, accuracy loss in heterogeneous environments is <0.8%.

Efficient data replication in distributed clouds via quantum entanglement algorithms.

Optimal design of hydrogen storage-based hybrid renewable systems: A case study using the particle swarm optimization (PSO) algorithm in Meknes, Morocco

A three-tier energy efficient architecture integrating virtual machine allocation and consolidation leveraging NSGA-II and LSTM for cloud data center

Energy efficient digital Twin–Driven intelligent monitoring and multi-source data fusion optimization for next-generation data center power environments

Thermodynamic optimization of adsorption cooling for low-grade data center waste heat: A calibrated white-box modeling approach

Performance analysis and prediction of single-phase immersion cooling data center

Green leapfrogging in China? Unveiling the impact of new infrastructure construction on synergistic control of pollutants and carbon emissions: Evidence from green data center

Physics-based two-scales reduced order model for real-time thermal management in immersion cooling data centers

Exploring the Demand Response Potential of Liquid-Cooled Data Centers

Extra terra nullius: Off-worlding the externalities of AI, Bitcoin mining and cloud computing with Orbital Data Centres

GRCC: A game explicit rate congestion control for rapid flow convergence in data center networks

Toward a sustainable energy future: Can green data center construction promote energy efficiency?

With the rapid growth of global digitalization, accurate short-term load forecasting for data centers (DCs) is crucial for efficient energy management and carbon reduction. However, existing methods usually treat DC loads as homogeneous, ignoring the distinct dynamics between non-cooling and cooling systems. This study proposes a data–physics hybrid forecasting framework for decoupled DC load prediction. The total load is first decomposed into non-cooling and cooling components according to their operational characteristics. For the non-cooling load forecasting, a DLinear–GRU fusion model is developed to jointly capture the linear trends and nonlinear temporal dependencies. For the cooling load forecasting, a physics-based Equivalent Thermal Parameter (ETP) model is derived, explicitly coupling the non-cooling forecast and ambient temperature variations. The integrated hybrid framework effectively combines data-driven and physics-based modeling. Case studies on measured DC data from North-west China show that the proposed framework achieves higher forecasting accuracy than existing models and exhibits a strong robustness under limited data, maintaining stable performance even with scarce training samples.

U.S. spent nuclear fuel, economic asymmetries, and the data center electricity demand challenge

LLC resonant converter is widely used in big data centers and telecom rectifiers because of its ZVS, ZCS, and high efficiency. These applications require a high voltage step down LLC converter to convert high-voltage input such as 380Vdc to low-voltage output such as 12Vdc. However, with the traditional resonant transformer structure of LLC converter, because of the high voltage step down, even if the secondary windings' turn number is designed only 1, the primary winding's turn number will be about 32. In this article, a novel transformer configuration is proposed that significantly reduces the number of primary winding's turns in high-voltage-step-down LLC resonant converters. The transformer's core is designed with a central primary leg and four separated side secondary legs. The primary winding's turn number is reduced to 32/4 = 8 turns, making PCB-based planar transformer's design viable. Its four secondary windings' eight output terminals are all symmetrically placed on the same side of the transformer, to minimize the output path length and make the four secondary currents sharing. The design method, loss calculation method, and magnetic circuit model are presented. A LLC converter's prototype of 380V/12V/2kW is built to verify the low-winding-turn-ratio planar transformer structure and its magnetic circuit model.

AI-driven low-GWP thermal management strategies: Converging innovations in electric vehicles and data center cooling