Rapid reductions in electricity demand caused by synchronised AI-computing activity may produce over-frequency disturbances, particularly in low-inertia power systems. This paper proposes a two-stage stability-constrained unit commitment (UC) framework that coordinates power-system operation with data-centre (DC) flexibility through a grid-facing import-cap interface. Unlike task-level scheduling models, this approach requires no information about workload arrivals, latency tolerances, or service priorities and is therefore applicable to system operators and facilities with limited workload visibility. Stage I formulates a mixed-integer second-order cone programme co-optimising generation, reserves, inverter-based frequency support, and hourly DC import limits to secure against both under-frequency (generator loss) and over-frequency (DC demand reduction) events. Stage II then assesses each hourly schedule through simplified nonlinear AC redispatch. In the 65-MW low-renewable case, coordination replaces 2.394 MWh of involuntary shedding with contracted import limitation. Under the assumed flexibility price and shedding penalty, the total penalised objective decreases from £282.84k/day to £247.40k/day. The results illustrate how non-firm connection agreements can improve frequency security and connection capability while leaving internal workload control to the DC operator.