Power measurements alone weakly constrain hidden compute, study finds
An arXiv paper quantifies how much undeclared computation an adversary could hide from auditors relying on power-draw readings of GPUs.
In plain words
Treaties limiting AI computation need ways to check that parties obey them. One idea is to measure a data centre's electricity draw from outside the chips. The paper calculates how much secret computing could evade such checks, finding power readings on their own are a weak guarantee.
What was announced
The authors derive a closed-form quantity, beta, defined as the largest amount of hidden computation a power trace cannot rule out, expressed as a fraction of declared machine capacity. On NVIDIA A100 GPUs, beta reaches 1.16 in the worst case, and an adversary using a matched-energy schedule still hides at least 0.41. The paper concludes that analogue power measurements alone constrain compute weakly.
The picture improves under a restricted threat model. If the verifier can re-execute the declared work at an observed operating point, beta falls to 0.059 in the maximally restricted case. The authors present this as a quantitative estimate of what analogue measurements can contribute to compute verification, rather than a full verification scheme.
What it does not show
- The evidence is the paper's abstract only; full methods, assumptions and experimental detail are not visible in the supplied passages.
- Measurements cover NVIDIA A100 GPUs as reported in the abstract; results for other hardware are not established by the supplied evidence.
- The re-execution result depends on a restricted threat model, so it does not apply to adversaries unconstrained by those conditions.
For security and IT teams
- Scope
- The findings concern verification of compute limits via analogue off-chip signals such as power draw, not a deployed auditing tool.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.