Researchers in China have devised a new technique called Bit2Watt, that demonstrates how AI and GPU workloads in data centers could be used to disrupt nearby power grids.
The study describes a scenario where an attacker pretends to be a normal cloud customer and runs specially designed GPU workloads. Instead of targeting computer systems directly, the workloads decrease or increase power consumption at the same time, putting stress on data centers and the electrical grid.
According to the researchers from Zhejiang University, GPU workloads can change power usage at frequencies above 6,000 Hz, much higher than common household devices. They warn that the rapid changes could cause voltage instability, harmonic distortion, and extra heat.
“Unlike traditional attacks that compromise grid-side devices or communication channels, Bit2Watt operates entirely within the cyber layer as a legal tenant, which could amplify fluctuations, harmonic distortion, and damping degradation, particularly in high-DER-penetration scenarios,” the researchers explain.
In a test involving a 1-megawatt power grid with renewable energy sources, the team found that 1,000 GPUs could create a total harmonic distortion of 46.8%, wasting nearly half of the electrical current and increasing heat output by about 20%. The researchers say this could make the power system unstable and, if protection systems shut down computing loads, could lead to cascading failures and large-scale blackouts.
The researchers also note that the attack could be difficult to detect because it uses legitimate cloud workloads rather than obvious malicious software. They recommend that cloud providers improve workload monitoring, coordinate cybersecurity with power system protections, and use local energy storage to help manage sudden changes in power demand.