Deep underground energy storage solutions for data centers driven by artificial intelligence energy consumption

Abstract

The rapid development of artificial intelligence (AI) has driven an unprecedented surge in electricity consumption by data centers (DCs). Although renewable energy and ground-based energy storage technologies are being widely adopted, they still face inherent limitations in scalability, safety, structure and stability. This review proposes deep underground energy storage (DUES) as a new approach and discusses three specific technologies: hydrogen storage, compressed air energy storage, and flow batteries. Firstly, it is found that DUES can align with AI load curves. The three technologies span storage durations from quarterly to hourly scales and response times from minutes to seconds, matching respectively the sustained high loads of AI model pre-training, the intermittent loads of fine-tuning, and the stochastic loads of inference. Secondly, we further develop a “Source-Grid-Production-Storage-Computing” model, in which DUES serves as the regulating hub bridging the intermittency of renewable energy with the stability of DCs. Subsequently, an economic assessment indicates that the levelized cost of storage for flow batteries ranges 0.340∼0.455 yuan/kWh, while that for compressed air energy storage ranges 0.746∼0.998 yuan/kWh, making them cost-competitive. Although the cost of hydrogen storage is 1.085∼1.658 yuan/kWh, it possesses strategic value. Finally, in view of challenges such as siting and stability of storage facilities, and system conversion efficiency, we recommend launching DUES for DC pilot projects in the short term, advancing multi-scenario applications in the medium term, and achieving industrialization in the long term. In summary, DUES represents a feasible, economically viable, and scalable solution to the energy challenges faced by AI-driven DCs.