Model tests optical-network scheduling for AI data centers

The scheduling method jointly considers where training runs, how long checkpoint data take to cross the network and when optical circuits are reconfigured. The researchers used California ISO electricity series, but transformed and redistributed renewable profiles across sites; checkpoint sizes and data-center capacities were also assumptions.

In this modeled case of five AI data centers, 16 training records and matched optical resources, dynamic optical switching reduced real-time grid-purchase cost by 12.50% and increased local renewable-energy utilization by 2.724 percentage points versus static operation. These are not savings measured at operating data centers; the authors say benefits depend on where energy opportunities occur and on network-transfer bottlenecks.