Reconfigurable optical links let AI data centers chase cheap, green power: 12.50% lower real-time grid cost and 2.724 points more local renewables

Multi-region AI data centers can move training work toward sites with cheaper electricity or more renewable generation, but a move requires transferring checkpoint state first, and that transfer time decides whether the successor can start before the favorable window closes. Earlier scheduling either treated cross-region migration as an instantaneous virtual link or estimated delay from prescribed transmission speed, distance, or pairwise bandwidth, leaving communication capacity a given parameter rather than a schedulable resource. On a benchmark of five geographically distributed AI data centers with equal completed work and network resources, the fixed-optical case cost $1471.25 in real-time grid purchases and the reconfigurable case $1287.33, with both drawing on the same 65.750 MWh of available renewable energy.

The scheduler splits each training job into checkpoint-separated compute segments, chooses a site and time slot for each segment, then schedules optical circuits for the checkpoint transfers and accepts the compute plan only after delivery is verified to finish inside the window. A compact compute-electricity master problem generates candidate schedules from estimated transfer delays, and an exact multicommodity-flow certification checks checkpoint delivery under shared-link capacity, enlarging only the binding lag and re-solving when a plan fails. Day-ahead planning, event-triggered intraday updates, and one-minute real-time dispatch share the same certification, with the intraday inference profile blending 30% day-ahead and 70% realized data and inference served with priority in every case.

The scheduling targets a portfolio of geographically separated AI data centers managed by one scheduling entity, where each site buys from a regional grid interface and receives local renewable generation, training jobs are checkpoint-separated with the first segment fixed at its origin, and inference is served locally with priority. On the network side, a fixed pool of optical resources is used, and reconfiguration re-pairs or re-directs the same installed circuits rather than adding capacity. The authors propose testing the framework with operator measurements and examining power-flow constraints, reserve participation, and larger portfolios with multiple reconfigurable circuits.

Checkpoint sizes, site renewable profiles, and inference curves are constructed or transformed benchmark assumptions rather than measurements; the electrical model stops at the site grid interface and excludes AC-network and reserve-market constraints; and the decomposition does not claim global optimality. The sensitivity screen shows near-zero cost and renewable differences when added configurability appears without transfer pressure, and some interior grid points are non-monotonic, which the authors attribute to a change in the discrete job-circuit matching regime. Cumulative absolute ramping rises from 17.644 MW to 17.834 MW, an increase of 1.08%.