A new benchmark measures how AI training bends under power cuts

Power, not chips, is becoming the ceiling on how much AI infrastructure operators can build. A new paper measures how much training throughput a job gives up when its power gets cut, so operators can decide which jobs to throttle first.

The authors call it the Power Flexibility Index. They built it from 131 LLM training runs on H200 GPUs, plus 24 more H200 runs and 34 H100 runs used to check the numbers held up, according to the paper. The runs covered dense and MoE architectures, both pretraining and fine-tuning, scaling up to 32 GPUs.

Grid operators and data centers already cut power during peak demand. Someone has to decide which training jobs absorb that cut without wrecking their timelines, and right now that decision is a guess. The index turns it into a number and works as a control primitive, feeding directly into scheduling decisions instead of just describing what already happened.

The throughput cost of a power cut isn't fixed. It swings a lot depending on the job, the authors find. A single index number implies operators can rank jobs by flexibility and throttle down the list in order. High variability means that ranking might not hold from one workload to the next, or one architecture to the next.

The paper also names telemetry signals that predict the index while a job is running, letting a scheduler act on this in real time instead of after the fact. That claim is harder to check from outside, since it depends on how noisy those signals stay once you're not inside a controlled 131-run study.