A new research project called Bit2Watt challenges the assumption that cloud workloads are electrically passive. The researchers demonstrated that a tenant with legitimate access to graphics processors could deliberately alternate between intensive computation and near-idle operation, producing high-frequency changes in power consumption.
No malware, administrative privilege or conventional software exploit is required. The attack concept relies on the physical relationship between computational activity and electricity demand.
From GPU scheduling to power modulation
The researchers developed two approaches. One uses a custom CUDA workload that rapidly switches a GPU between high and low utilization. The other hides similar power changes inside a functioning large language model training process by modifying training operations and workload timing.
The second approach is potentially more difficult to identify because it can resemble legitimate artificial intelligence activity. Standard cloud monitoring typically focuses on utilization, temperature, billing and application performance, not high-frequency electrical patterns.
Tests showed that both consumer and data-center GPUs could create measurable oscillations. The researchers then modeled a worst-case scenario in which a large number of physically clustered GPUs operated in close synchronization. Their simulations indicated that such coordination might create harmful distortion or instability in certain grid environments.
Important practical limitations
Bit2Watt should not be interpreted as evidence that attackers can currently shut down a national grid through a cloud account. The most severe results came from simulations using demanding assumptions, including access to many GPUs, tight timing synchronization and favorable electrical conditions. Data-center power conditioning and cloud-provider controls could also weaken the effect.
How operators can respond
- Correlate GPU scheduling telemetry with rack-level and facility-level power measurements.
- Detect repeated utilization patterns that have no clear computational purpose.
- Limit a single tenant's ability to synchronize large numbers of physically adjacent accelerators.
- Use batteries, supercapacitors and harmonic filtering to absorb rapid load changes.
- Create information-sharing procedures between cloud security and electrical engineering teams.
In my view, the most valuable lesson is organizational. Cloud security teams and power engineers traditionally monitor different systems, creating a blind spot between logical workloads and physical infrastructure. Bit2Watt suggests that hyperscale AI facilities should treat electrical behavior as part of tenant risk management, even when every instruction being executed is technically permitted. ([thehackernews.com](https://thehackernews.com/2026/07/new-bit2watt-attack-could-let-cloud.html))