Interruptible AI training as virtual energy storage for renewable integration in computing centers

Intelligent computing centers are emerging as large electricity consumers whose demand can be temporally misaligned with variable renewable generation. This study proposes a low-carbon dispatch framework that represents checkpoint-enabled interruptible AI training workloads as a baseline-relative virtual energy storage resource. The model distinguishes non-movable, short-term shiftable, continuous, and interruptible computing loads and uses cumulative training progress as a virtual state of charge. In this representation, virtual discharging denotes demand reduction relative to a fixed baseline schedule, whereas virtual charging denotes demand recovery or advancement during periods of abundant renewable generation. The dispatch model jointly considers time-of-use electricity prices, ladder-type carbon trading, and green electricity certificates to align economic and emission incentives. Using workload profiles reconstructed from the Alibaba PAI GPU cluster trace, the framework is evaluated in a representative Northwest China computing-center microgrid case. Under the examined assumptions, the proposed strategy reduces renewable curtailment from 22.82% to 0.69%, lowers daily operating costs by 14.45%, and decreases physical carbon emissions by 30.46%. Benchmark and scenario analyses indicate that workload-side flexibility can complement behind-the-meter storage for renewable matching, although its value depends on workload composition, renewable timing, grid carbon intensity, tariff design, and market conditions. The findings suggest that checkpoint-enabled training can provide an asset-light demand-side flexibility option for renewable-rich computing centers, but its applicability requires careful consideration of service constraints, accounting rules, and regional power-system conditions.