Bit2Watt shows an attacker can destabilize a data center's local power grid purely by scheduling GPU workloads, with no physical or electrical access

Ji, Pan & Xu's "Bit2Watt" (accepted CHES 2026) describes a cyber-physical attack in which an adversary who is nothing more than an ordinary, legitimate cloud tenant induces high-frequency power fluctuations by scheduling GPU compute in patterns timed against the facility's power infrastructure. No malware, privilege escalation, or physical/electrical access to the hardware is required — the attacker "operates entirely within the cyber layer as a legal tenant." The paper validates the mechanism on real GPUs and grid-connected PV inverters, and in a modeled 1,000-GPU, 1-MW local power system with 90% distributed energy resources (DERs), reports that the attack "raises current THD to 46.8% and results in a damping ratio of -0.27" — degradation severe enough to risk triggering protective disconnects and, per the paper's simulations, cascading toward transmission-scale failure.

This reframes a problem the vault has so far treated as a sourcing constraint — can enough power be built or contracted to meet AI compute demand, as in claim-microsoft-constellation-tmi-restart-crane-clean-energy-center and claim-kairos-kp-fhr-molten-salt-triso-reactor, against a transmission bottleneck already running at roughly twice installed capacity — into a security problem: the same tight coupling between compute load and grid stability that makes power scarce also makes it manipulable. Bit2Watt's attacker doesn't compete for power; it weaponizes the demand signal itself.

Bit2Watt's own physical premise — that computation's power draw carries exploitable information/leverage — did not originate with this paper; see claim-power-side-channel-attacks-scaled-chip-to-grid-1999-2026 for the lineage back to 1999 chip-level side-channel analysis.