AI & Data-Center Energy Demand
The question
As AI workloads exploded, utilities and policymakers faced an urgent, poorly-quantified question: how much electricity will AI and data centers actually demand, and what does that mean for a grid that plans in decades? Speculation was plentiful. Defensible numbers were scarce.
Powering Intelligence
I co-authored Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption (EPRI, 2024), which became one of the most widely-referenced analyses on the subject. It projected U.S. data-center load growth from a 2023 baseline of roughly 152 TWh/year — about 4% of national electricity — across four scenarios through 2030:
- A low-growth path reaching about 4.6% of U.S. electricity
- A moderate path near 5.0%
- A high-growth path around 6.8%
- An aggressive path as high as 9.1%
The analysis put concrete bounds on AI’s energy intensity — a single ChatGPT request consumes roughly ten times the electricity of a Google search (about 2.9 Wh versus 0.3 Wh) — and mapped its striking geographic concentration, with 15 states accounting for roughly 80% of U.S. data-center load and Virginia alone drawing about a quarter of its state electricity. It also flagged the end of an era: average data-center efficiency (PUE) had improved steadily for over a decade but is now plateauing just as AI load accelerates.
Crucially, the report didn’t just sound an alarm. It laid out strategies for utilities to absorb the growth: efficiency and load flexibility, tighter developer–utility coordination, and better five-to-ten-year grid-planning tools.
Data Center Heat Reuse
I also contributed to a companion Quick Insight on data-center heat reuse, which examined cooling — 30–40% of a data center’s energy use — as a recovery opportunity rather than pure waste. It highlighted real deployments, including a project in Helsinki that channels data-center waste heat to warm hundreds of thousands of residents while avoiding substantial annual CO₂ emissions, and tracked emerging regulation mandating heat reuse.
Why it matters
Grid planning runs on years-long lead times, so the cost of guessing wrong about AI’s demand is measured in blackouts or stranded investment. Giving planners credible, bounded projections — rather than hype — directly shapes whether the grid can absorb AI’s growth without compromising reliability or climate goals.