How to Shift AI Inference Workloads Across Data Centers Based on Grid Conditions
Short answer
Because inference is response-time sensitive, it can't simply be paused when the grid is stressed. Instead, it can be moved. LōD's Compute Energy Distribution Network (CEDN) routes inference across data centers in real time, using live grid conditions and energy cost, so load shifts away from constrained or expensive sites without disrupting response time.
Why inference is routed, not curtailed
Training jobs can often be throttled or paused with no customer-facing impact. Inference cannot: users and applications are waiting on responses. Curtailing inference to relieve the grid would put service levels at risk.
Shifting inference to another site keeps the work running while the load on a stressed location drops. The grid sees a reduction at one site; the customer sees no change in service.
What drives a routing decision
- Real-time grid conditions at each site, not a static schedule
- Energy cost at each site, so workloads move to where power is cheaper
- Workload type, because training and inference are handled differently
Where this applies
- Operators with inference capacity at more than one site
- Teams that want to energize more compute per site without waiting on new substation capacity
- Operators who want to respond to grid signals while keeping response times intact
Limitations and scope
- Workload shifting requires capacity at more than one data center. A single-site operator has no destination to route to.
- Routing relies on available headroom at the receiving site; it moves load, it does not create compute capacity.
- Training workloads follow a different path (GPU-level throttling and curtailment) and are covered separately.