Bit2Watt Attack Targets Cloud Power Grids Via GPU Access

A novel attack dubbed Bit2Watt enables cloud tenants to manipulate data center power consumption to the point of threatening the stability of the electrical grid, according to research from Zhejiang University. This attack does not require any security exploits or unauthorized access; it leverages ordinary GPU access within a cloud environment. The researchers presented their findings in a paper accepted to CHES 2026, a conference hosted by the International Association for Cryptologic Research (IACR) focusing on hardware security.

The Bit2Watt attack functions by rapidly fluctuating the power draw of GPUs. By strategically increasing and decreasing the computational load on these processors, an attacker can create significant and fast swings in electricity demand. These fluctuations, when amplified across a large number of GPUs within a data center, can exceed the grid's capacity to absorb such rapid changes. The researchers demonstrated this by measuring the power consumption patterns, indicating that the magnitude and speed of these power shifts are sufficient to cause grid instability.

The implications of Bit2Watt are significant for cloud infrastructure providers and energy grid operators. It highlights a previously underestimated vulnerability where legitimate access to hardware resources can be weaponized. Unlike traditional cyberattacks that rely on exploiting software or network vulnerabilities, Bit2Watt exploits the physical characteristics of hardware operation and its interaction with the power grid. This means that even a well-secured cloud environment could be susceptible to such a disruption if a tenant has sufficient GPU resources at their disposal.

The research team from Zhejiang University plans to further investigate the precise thresholds and conditions under which the Bit2Watt attack can cause critical grid failures. Their work underscores the growing need for security measures that consider the physical and resource-level interactions between digital infrastructure and the real-world systems they depend on, such as the power grid. The findings are expected to prompt discussions on new security protocols and monitoring systems within data centers and energy management sectors.