AI Inference as Relocatable Electricity Demand: A Latency-Constrained Energy-Geography Framework

Researchers present a framework for optimizing AI inference workload placement across geographically distributed data centers by treating computation as relocatable electricity demand. The model balances latency constraints against energy costs and carbon intensity, revealing that workload flexibility significantly expands execution geography but faces practical friction from migration costs, regulatory limits, and network constraints.

This research addresses a critical emerging challenge in AI infrastructure: the geographic concentration of computational demand and its energy implications. As AI inference becomes ubiquitous across cloud services, the ability to shift workloads between regions based on electricity pricing and carbon intensity offers meaningful optimization opportunities that have received limited academic scrutiny.

The framework's contribution lies in formalizing the trade-off between latency tolerance and energy geography. By modeling inference placement as a constrained optimization problem incorporating electricity prices, marginal carbon intensity, power usage effectiveness, and network latency, the authors provide a methodological foundation for thinking about computation as a flexible electricity load. This perspective mirrors how grid operators have long managed demand response in physical electricity markets, but applies it to digital infrastructure where physical constraints are less binding.

For infrastructure providers and cloud operators, this research validates operational strategies already emerging in practice: routing latency-tolerant workloads toward cheaper electricity or lower-carbon regions. The stylized simulation results demonstrate that heterogeneous latency tolerance naturally stratifies workloads into local, regional, and energy-optimized execution tiers—a finding with direct implications for resource allocation and operational efficiency.

The practical impact depends heavily on the friction factors the paper identifies: migration costs, egress fees, state locality requirements, and regulatory constraints can substantially diminish realized benefits. This suggests that while latency flexibility theoretically unlocks significant energy-geography arbitrage, actual gains will vary widely by workload type and regulatory jurisdiction. The research establishes important baseline metrics for measuring relocatable inference demand and carbon return on latency relaxation.

• →AI inference can function as flexible, relocatable electricity demand when latency constraints permit geographic distribution of computation.
• →An energy-latency frontier exists where relaxing latency budgets progressively expands execution geography and unlocks energy optimization opportunities.
• →Migration frictions, egress costs, state locality, legal constraints, and capacity limits substantially reduce theoretical energy and carbon benefits in practice.
• →Workloads naturally stratify into local, regional, and energy-optimized execution layers based on their latency tolerance thresholds.
• →The framework provides operational metrics for measuring relocatable inference demand and carbon return on latency, enabling quantified optimization.