Grid-Intelligent AI Data Centres for Primary Response

Abstract

The rise of artificial intelligence (AI) has driven the growth of GPU-intensive data centres dedicated to AI workloads, which have a greater power demand that stresses grid capacity. While the potential for data centres to provide grid flexibility has been studied previously, it remains unclear whether AI-dedicated data centres can deliver fast-timescale services like primary response, which are highly valuable for the grid and only require short-duration power deviations. To address this gap, this work presents the first power-system-oriented real-world experiments conducted on NVIDIA RTX 2080 Ti and A6000 GPUs running AI training workloads. We show that dynamic power capping—supported by most NVIDIA GPUs—enables rapid power control capable of primary response. Although we observe that power capping introduces transient power deviations, this issue reduces to an acceptable level when many GPUs are capped simultaneously. When aggregated across 50 A6000 GPUs, over 99.7% of the power response remains within the performance bounds of the Dynamic Containment service, a representative U.K. primary response service. To account for varying Quality of Service (QoS) costs and power response accuracy across GPU nodes, we propose a reinforcement learning (RL) controller for efficient real-time power capping. We use GPU hardware measurements to develop a simulation environment for RL training and testing, and demonstrate that the proposed RL controller achieves the highest tracking accuracy at a comparably low QoS cost, compared with benchmarks including two heuristic controllers and an optimisation-based controller with near-perfect information. These findings establish AI data centres as viable, high-value assets for delivering fast-timescale grid services.