AI workload volatility raises data centre power waste

Data Center Knowledge reported that AI training volatility creates sharp GPU demand drops in large clusters. Operators smooth those dips with secondary workloads, including productive tasks that compete with primary training and dummy workloads that waste power. The signal for BTW readers is that AI data centre energy pressure is also a workload-aware power-management problem.

• Bulk-synchronous training creates sharp GPU demand drops across large clusters
• Secondary workloads stabilise power but raise energy and infrastructure costs

The fact

AI data centres use more power partly because bulk-synchronous training creates sharp GPU demand drops across clusters. Operators smooth those dips with secondary workloads: productive tasks that compete for GPU capacity, or dummy workloads that run meaningless calculations to maintain a stable power profile. Oracle uses a millisecond-scale GPU heartbeat to trigger this activity. The practice inflates energy use, increases cooling demand and can delay grid approvals.

The Assessment

AI data centre energy pressure is not only about insufficient generation; it is also about inefficient power behaviour inside the facility. Secondary workloads flatten demand but introduce hidden costs: productive tasks slow primary training, dummy tasks waste electricity, and sustained peak-like operation accelerates equipment wear. The industry is using extra compute to compensate for the lack of workload-aware power management at hardware and orchestration level.

What to Watch

Watch whether GPU vendors and operators develop lower-waste alternatives to secondary workloads, including hardware-level smoothing, predictive scheduling or interconnection rules tied to power volatility.