Bit2Watt: Cloud Tenant Could Destabilize the Power Grid
A cloud tenant could rattle the power grid with nothing but a rented GPU, researchers say.
Chinese researchers assert that a cloud customer can potentially disrupt a power grid without employing malicious tactics like hacking or password theft. All they need is a rented GPU and a workload designed to behave erratically. However, the most concerning findings come from simulations, not actual attacks.
July 21, 2026 – 2:52 pm
Image by: Shutterstock
AI data centers already strain the grid simply by operating. But a new paper poses a darker question: what if a tenant intended to sabotage it?
This concept is outlined in a preprint titled Bit2Watt, accepted to a prominent hardware security conference. The researchers, from Zhejiang University, envision a paying customer as the perpetrator.
How It Works
The key lies in understanding that a GPU’s power consumption varies directly with its computational load. Heavy loading causes current spikes, while idling results in lower current.
By rapidly switching between these states at extremely high frequencies (up to 6,000 times per second), researchers achieved controllable power fluctuations at the wall socket.
They demonstrate two methods: one uses a dedicated workload that cloud providers might be able to detect, and another, more subtle approach hides the pattern within an AI training run, making it harder to identify. Both methods require no special access because tenants already control their own tasks.
The Scary Number (with an Asterisk)
While individual GPUs pose little risk, the potential for bulk attacks is concerning. The paper models 1,000 GPUs synchronously pulsing on a small grid heavily reliant on solar and battery power.
In this worst-case scenario, the simulated grid becomes unstable, wasting nearly half its current and overheating. When applied to a model of Europe’s transmission network, a localized disturbance escalates until it disrupts approximately 81% of the load.
Important Note: As with many reports on cloud attacks, these figures are from a simulation with extreme assumptions, not a forecast. Achieving synchronized pulsing among a large number of cloud GPUs remains a theoretical challenge. No live system was compromised.
Grounded in Real Physics
This concept is not entirely theoretical as the underlying physics is well-documented. In 2025, Microsoft, OpenAI, and Nvidia warned that synchronized power swings from large training jobs could damage grid infrastructure when their rhythms align with utility frequencies. Meta also highlighted this risk during Llama 3’s training.
The grid has already experienced such a disruption accidentally. In July 2024, a fault near data center-dense Northern Virginia caused a sudden loss of approximately 1,500 megawatts from the grid. Regulators initially downplayed it as a non-crisis but later established a task force to study this emerging threat from energy-intensive data centers.
The Feedback Loop and No Bug to Patch
The researchers also propose a feedback attack called Watt2Bit. The same electrical stress that affects the grid can overheat servers, triggering protection mechanisms and leading to offline periods. Thus, a power problem becomes a denial of service.
The challenge is that there is no software bug to patch; it’s a physical limit.