AI-driven data center power demand: grid upgrades and flexibility solutions

The article argues that data center operators and utilities must combine flexibility measures and transmission upgrades to meet AI-driven power demand.

• Main action/analysis: Data center operators are implementing flexibility solutions (energy storage, demand response, virtual power plants, behind-the-meter systems, workload scheduling) and technology changes (GPU roadmaps implying 1MW per rack) to reduce grid strain; a Duke University study finds that 0.25% flexibility (≈22 hours/year) could allow the U.S. grid to accommodate 76GW of new data center load. Google has agreements with Indiana Michigan Power and the Tennessee Valley Authority to pause or reduce AI/ML tasks during peak demand as an early example of demand-response for ML workloads.
• Background and infrastructure details: The core constraint is transmission and interconnection, not generation: Dominion Energy’s transmission backlogs will see relief when new infrastructure comes online in 2026, PG&E warns new substation work may take five years or more, and regional operators (outside Texas) say they cannot meet FERC deadlines for critical upgrades; developers build facilities in 2–3 years versus 4–8 years for interconnection, and Goldman Sachs estimates $720 billion of grid spending may be required through 2030 (driving uptake of expensive behind-the-meter solutions).