As artificial intelligence training workloads push datacenter power consumption into the multi-megawatt range, researchers have uncovered an unexpected problem: independent training jobs can spontaneously synchronize their power-draw cycles, creating grid disturbances that scale far worse than previously modeled.
Traditional power-system analysis treats each datacenter as an independent periodic load, assuming aggregate fluctuations shrink with scale. However, when multiple AI training jobs compete for a shared, oversubscribed power budget, a coupling mechanism emerges: load-dependent throttling. Whenever total demand spikes, power caps, voltage regulation, and shared cooling infrastructure all slow computation simultaneously. This creates negative feedback that paradoxically locks independent workloads into synchronous cycles.
The phenomenon mirrors phase-locking in weakly coupled nonlinear oscillators—a classical Kuramoto system—but without the usual clock-based coupling. Instead, the power-management stack itself becomes the synchronization channel. The research reveals three critical findings: First, the coupling is repulsive at low phase lag but becomes attractive when control-loop delays exceed half a cycle, making synchronization mode-selective and dependent on system damping. Second, synchronization onset is hysteretic and first-order, meaning small changes in operating conditions can trigger sudden, large-scale alignment. Third, phase-scattering scheduling—deliberately desynchronizing job phases—raises the stability threshold across all frequency modes simultaneously.
This work reframes the operator's problem: rather than asking how a single datacenter loads the grid, the question becomes whether fleets of co-capped facilities can enter dangerous resonance. The team proposes a falsifiable test using two synchronized jobs to validate predictions, opening a path to practical mitigation.
As datacenters consolidate AI training workloads, grid operators and facility engineers must redesign power-management policies to prevent emergent synchronization. The stakes are high: uncontrolled phase-locking could translate steady datacenter loads into large, grid-destabilizing transients.
