The rapid expansion of artificial intelligence (AI) is driving hyperscale data-centre (DC) growth at a pace that challenges the operation of inverter-based power systems with high renewable penetration. Unlike traditional CPU-oriented DCs, AI training clusters exhibit synchronised, phase-driven execution patterns with abrupt transitions between near-saturated GPU utilisation to near-idle states. These transitions can create large short-timescale demand drops, and when aggregated across many racks or clusters, the effective loss-of-load event can trigger severe over-frequency excursions and reduce stability margins.

To tackle this problem, we propose a stability-constrained unit commitment (UC) framework that coordinates system-level scheduling with DC-side flexibility. In our formulation, each DC can optimise its internal compute operations, while the system operator enforces a period-by-period external power envelope to bound the size of credible disturbances. The model captures two support patterns: (i) envelope-constrained demand scheduling, where worst-case DC trip events are secured through downward reserve and frequency support from synchronous units and grid batteries; and (ii) virtual-synchronous operation, where DC-owned storage is scheduled between grid-forming (GFM) and grid-following (GFL) modes, enabling mode-dependent inertia and damping contribution to stabilise the grid.

The optimisation co-determines generator commitment, renewable generation, upward and downward reserves, controllable DC demand trajectories, and binary GFM/GFL transitions, with associated switching penalties. Beyond standard UC operating constraints (capacity, ramping, power balance), we embed frequency-security constraints directly into scheduling to enforce N-1 generation-loss reserve adequacy, DC-trip reserve adequacy, as well as nadir and RoCoF-oriented inertia requirements. Small-signal stability is also enforced using the generalised short-circuit ratio (gSCR). In addition, DCs’ auxiliary service requirements will be constrained by their active and reactive power when operating in GFM mode, while still satisfying the service promise.

The non-convex frequency and small-signal stability constraints will be approximated by conic relaxation and data-driven linearisation, respectively. The resultant mixed-integer second-order cone program (MISOCP) is solved by open-source solvers such as MOSEK. Using a 14-bus, 24-period case study, we accompany an open-source implementation and quantify the trade-offs among generation cost, quality-of-service penalties, and dynamic stability margins. The results demonstrate that centralised DC dispatch under a single power envelope signal can simultaneously ensure economic optimality, frequency security, and small-signal stability within a tractable single-stage formulation.