Do Co-Located AI Training Jobs Synchronize? Load-Dependent Throttling as a Coupling Mechanism for Phase-Locking Behind a Shared Power Cap
Large-scale AI training turns computing facilities into multi-megawatt loads whose power draw is periodic: tens of thousands of accelerators step in lockstep between compute-bound phases near peak power and communication-bound phases where they idle. Prior work treats each facility as an exogenous periodic forcing on the grid. We pose the operator's question instead: when many independent training jobs share one oversubscribed power envelope, do their cycles stay independent, so aggregate fluctuation grows as the square root of the number of jobs, or can the power-management stack phase-lock them into linear growth? This is emergent synchronization in a population of nonlinear oscillators - the Kuramoto setting - but classical coupling is absent, since accelerator clocks are decoupled from line frequency. We identify the coupling channel in load-dependent throttling: caps, voltage droop, and shared cooling slow computation exactly when aggregate demand is high. Formalizing the fleet as a generalized Kuramoto system, we obtain three operator-facing statements. Dimension: the coupling is repulsive to leading order and turns attractive only when the control loop's phase lag exceeds half a cycle; protection is mode-selective, so rate diversity is required. Detect: frequency-correlated frustration makes the onset first-order and hysteretic. Mitigate: phase-scattering scheduling raises every mode threshold at once. The prediction is falsifiable by a two-job co-capped measurement, which we specify.