Powering the AI Inference Wave with EPRI's Ben Sooter - Ep. 292

What It Covers

EPRI's Ben Sooter explains how micro data centers — small, distributed inference facilities of 3–20 megawatts — can be co-located at underutilized electrical substations across the US to meet the coming wave of AI inference demand without overloading transmission grids or requiring new infrastructure investment.

Key Questions Answered

• •Inference vs. Training Power Split: Over the lifetime of an AI model, roughly 80–90% of total compute consumption occurs during inference, not training. Organizations planning energy infrastructure should weight capacity planning heavily toward inference workloads, not just the headline-grabbing training buildouts that currently dominate industry conversation and capital allocation.
• •Substation Co-location Strategy: Thousands of existing electrical substations across the US carry underutilized capacity, typically 3–10 megawatts of headroom. Siting micro data centers directly adjacent to these substations bypasses transmission interconnection queues, reduces permitting timelines, and avoids new steel-in-ground costs — accelerating time-to-power for inference deployments by a meaningful margin.
• •Distributed Aggregation Model: A single substation's 5-megawatt surplus may be insufficient for viable data center economics. EPRI's approach aggregates five nearby sites within a metro region into a 25-megawatt distributed cluster, treating it as one logical project — satisfying both utility grid constraints and the minimum scale thresholds that data center operators require to justify investment.
• •Demand Flexibility as Grid Asset: Substations often hold significantly more capacity than their rated surplus, except during annual peak demand days. Pairing micro data centers with battery storage and backup generation, then engineering load-shedding protocols that reroute compute tasks to other nodes during grid peaks, unlocks that larger envelope without requiring additional grid upgrades.
• •Agentic AI Reshapes Load Forecasting: Early inference load models assumed human-driven usage patterns — daytime peaks, overnight lows. The rapid emergence of autonomous AI agents that run continuous background tasks around the clock invalidates that assumption. Grid planners and data center operators should model inference loads as potentially flat or inverted curves, not standard residential demand profiles.