The point: Regular GPU workload cycling generates controllable kilohertz-range power oscillations that standard monitoring systems miss and can destabilize data centre power networks.
Researchers at Zhejiang University have demonstrated that cloud users can destabilize data centre power supplies through targeted GPU utilization patterns—without exploits or malware. The technique poses risks to critical infrastructure.
Three scientists from Zhejiang University have presented the Bit2Watt concept for the CHES 2026 conference. The method demonstrates how regular access to graphics processors in cloud infrastructures can induce oscillations in a data centre’s power consumption. GPU power draw depends directly on compute core utilization; rapid switching between full load and idle creates controllable power fluctuations at the outlet. Administrator privileges, exploits or malware are not required for this process.
The researchers describe two implementations: SWMA, a custom-developed CUDA kernel for high-frequency fluctuations, and LTMA, which embeds power oscillations into large language model training through targeted hyperparameter adjustments. In simulations of a local power grid with 1,000 GPUs operating synchronously, harmonic distortion rose to 46.8 percent—well above international standard limits. When modelling a transmission grid, the local disturbances triggered cascading failures and load shedding. The authors note, however, that precise temporal synchronization of thousands of GPUs in production cloud environments represents a significant technical challenge.
A critical gap lies in monitoring: standard data centre telemetry systems record power data at intervals of several hundred milliseconds to one second. The generated oscillations operate in the kilohertz range and are thus barely detected by conventional measurement equipment. At the same time, the researchers document that thermal and electrical feedback effects can trigger circuit breakers. Additionally, fluctuations in electromagnetic radiation can be exploited as a covert channel for near-field data transmission.
To mitigate risk, the researchers recommend multi-layered measures: battery storage to buffer short-term load spikes, high-frequency filters for harmonics, and integrated monitoring of compute load and power supply at higher sampling rates. CISOs should review their power distribution and monitoring systems to determine whether they can detect and respond to such high-frequency anomalies.
Source: www.it-daily.net · Published 22 July 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.