Data center cooling in the AI era: cross-scale challenges and system-level outlook

The rapid growth of artificial intelligence (AI) is fundamentally reshaping data center thermal and energy requirements, driven by increasing power densities and heterogeneous workload characteristics. This review provides an integrated cross-scale and temporal perspective that connects AI heat generation with chip-, rack-, and facility-scale thermal management and their interactions with external energy, water, and grid infrastructures. Modern AI workloads, enabled by accelerator-based hardware, can generate highly localized and temporally dynamic heat fluxes, challenging traditional assumptions of uniform and steady-state cooling design. Accordingly, a range of cooling technologies is examined, including air-based systems, liquid cooling approaches, hybrid configurations, and heat-reuse strategies, with emphasis on their performance limits and deployment constraints. The analysis further identifies key cross-cutting barriers to scalability and efficiency, including thermodynamic irreversibility, coupled water–energy–grid dependencies, cross-scale thermal interactions, reliability and maintainability constraints, and technology-specific trade-offs. Building on these observations, future cooling systems are expected to shift toward integrated and adaptive designs that prioritize multi-objective optimization rather than isolated performance improvements. Emerging directions include grid-responsive and water-aware operation, thermal storage-enabled temporal decoupling, co-designed computing and cooling infrastructures, and predictive closed-loop thermal management enabled by digital twins and data-driven control. Collectively, the findings highlight a transition from component-level optimization toward fully integrated system architectures that jointly optimize energy efficiency, resource utilization, operational resilience, and computational performance across spatial and temporal scales.