Bottom line: Legitimate GPU usage by cloud tenants can trigger power grid instability by rhythmically modulating data center loads.
Researchers from Zhejiang University have demonstrated an attack in which cloud tenants can deliberately cause power consumption fluctuations in a data center through regular GPU access alone — without exploits or unauthorized access and with direct impact on network stability.
The attack is called Bit2Watt and was developed by researchers at Zhejiang University. The method has been submitted to CHES 2026 (IACR Hardware Security Conference). The core mechanism exploits a cloud tenant’s ability to control GPU loads such that the data center’s power consumption fluctuates in patterns that are measurable and significant.
For a CISO, this represents a novel infrastructure threat scenario: an ostensibly legitimate user — that is, someone with proper GPU authorization — can generate power fluctuations in cloud infrastructure through timing manipulation of computational loads. This requires no classical exploit, no privilege escalation, and no unauthorized access, but only normal usage rights.
The practical threat lies in cascading effects: if data centers are connected to power grids that are critical to stability, these artificially induced load fluctuations can lead to frequency deviations or voltage instability in the public network. Particularly in regions with densely concentrated cloud infrastructure, a coordinated attack could amplify these effects.
For risk assessment purposes, it is relevant to note: this is a supply-chain or tenant-isolation problem that attacks at the hardware level. Classical security models that assume tenant separation at the operating system level do not provide protection here. The research findings were presented at CHES 2026; immediate industrial implementation of countermeasures is not yet documented.
Source: thehackernews.com · Published 21 July 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.7.3.