Bit2Watt Attack Turns AI Data Centers Into Cyber-Physical Threats to Local Power Grids

Bit2Watt is a newly described cyber‑physical attack class that leverages large, legitimate GPU workloads in hyperscale AI data centers to create rapid, high‑frequency power modulation that can excite resonances in inverter‑dominated local grids. The report documents two modulation techniques (Synthetic Workload Modulation Attack and LLM Training Modulation Attack), empirical GPU power behavior, inverter resonance measurements, and simulations showing that coordinated GPU modulation (hundreds–thousands of GPUs) can degrade power quality, induce grid instability, damage data‑center power equipment, cause denial‑of‑service to AI clusters, and even provide a covert side‑channel for data exfiltration.