Power Control is a reinforcement learning agent that manages your GPU cluster's power draw in real time, staying inside your budget while producing more, at 22% lower cost per token, with no hardware changes.
No root access · Week 1 is read-only · Reverts automatically if anything looks wrong
Benchmarked on live A100 hardware, three replications across two machines, comparing the Power Control agent against a safely throttled baseline.
Stylized reconstruction of the measured 20 minute pilot window · with Power Control vs. safely throttled
The 22% cost reduction measured in the pilot holds regardless of scale — it just compounds. Modeled at the same measured efficiency gain, scaled linearly with fleet size.
Anchor: measured 1 MW pilot result ($19,600 / yr, 245 MWh saved). Other scales are a linear projection at the same 22% efficiency gain — replaced with your fleet's real numbers in the pilot.
Across the pilot fleet, measured peak power draw ran to roughly half of combined nameplate rating — capacity most operators are already paying to provision but not using.
Backed by published research, not internal claims. The controller behind Power Control is documented in arXiv 2608.11226 — half-second power telemetry across A100 GPUs, 7B to 72B model training scales, 89.8% fewer power violations and +18.1% output at 7B.
We install a lightweight sidecar agent against your real GPU fleet. It watches, it doesn't act, until you say go.