Study models how rented GPUs could disrupt Europe's power grid
A new study models how coordinated power surges from rented cloud GPUs could theoretically crash Europe's grid, exposing a dangerous blind spot between data centre operators and electricity providers.
Researchers at Zhejiang University have demonstrated that a paying cloud customer could weaponise rented graphics processing units to destabilise power grids. In a preprint accepted to a leading hardware-security conference, the team outlines how the rapid cycling of AI workloads can create deliberate, high-frequency power fluctuations.
The attack exploits the basic electrical behaviour of a GPU, which draws current in direct proportion to its computational load. By forcing chips to flip between maximum strain and idle states more than 6,000 times a second, a tenant generates a controllable power wobble at the socket. The researchers note this pulsing can be hidden within legitimate AI training runs, evading detection without requiring any special system access.
A single chip causes negligible disturbance, but the threat multiplies with scale. When the researchers simulated 1,000 GPUs pulsing in perfect synchronisation on a local grid powered by solar and batteries, the system became unstable and wasted nearly half its current. Modelling this scenario on the European transmission network produced a severe result, with a small local disturbance cascading to shed about 81% of the network's load.
That figure represents a worst-case simulation stacked with optimistic assumptions for the attacker, not a guaranteed outcome. The authors concede that synchronising a real fleet of cloud GPUs remains an unsolved challenge, and no live systems were tested.
Despite the theoretical boundaries, the underlying physics are already a recognised concern. In 2025, Microsoft, OpenAI, and Nvidia warned that the synchronised power swings inherent to large AI training jobs can physically damage grid infrastructure. Meta encountered the same risk while training its Llama 3 model.
The grid has also experienced accidental shocks from data centre clusters. A fault near Northern Virginia in July 2024 abruptly knocked 1,500 megawatts off the grid. While regulators downplayed an immediate crisis, they convened a task force to study the systemic risks posed by these massive power consumers.
The Zhejiang team also describes a secondary threat they call Watt2Bit, where the electrical stress intended for the grid bounces back to overheat the servers themselves. This creates a feedback loop, turning a power grid attack into a denial-of-service outage for the cloud provider.
The core issue for European infrastructure planners is that there is no software bug to fix. Standard grid monitoring samples power far too slowly to catch these rapid fluctuations.
The vulnerability is structural. Volatile AI loads are being plugged into grids increasingly reliant on solar inverters, yet the two systems are managed by separate companies using incompatible monitoring tools. As Europe races to build out AI data centres, the seam between the tech sector and the energy transition remains unguarded.
