AI data centers’ erratic power demands during model training are causing physical damage to gas turbines, with the Federal Energy Regulatory Commission acknowledging the risk as severe. The rapid load fluctuations from AI workloads twist and snap turbine generator shafts — equipment designed to run at steady output for days or months — creating a mechanical vulnerability that threatens grid reliability as data center power consumption accelerates.
The problem stems from a fundamental mismatch: modern gas turbines, particularly aeroderivative models favored for their fast-start capability, are engineered for stable baseload or predictable cycling. AI training clusters, however, can swing power draw by tens of megawatts in seconds as workloads shift between compute-intensive phases. That translates into torque spikes on the generator shaft that exceed design margins, a failure mode more commonly associated with grid faults than normal operation.
FERC’s public concern signals that regulators view this as a systemic issue, not an isolated maintenance headache. As utilities lean on gas-fired generation to balance renewables and meet surging data center demand — projected to double or triple in key markets by 2030 — the fleet’s ability to withstand these new duty cycles becomes a planning imperative. The alternative, oversizing turbines or adding mechanical buffers, erodes the economics that made gas the default partner for intermittent generation.
A Tesla-funded study currently in peer review has identified the specific harmonic interactions driving shaft fatigue and proposed a control strategy to smooth demand at the source. If validated, the approach could let data centers modulate their load profiles without sacrificing training throughput, effectively decoupling compute variability from mechanical stress. That would represent a rare win-win: preserving turbine life while avoiding the capital cost of dedicated storage or oversized generation.
Read the full report at Energy Central.