A verified control layer that holds compute load inside the physical bounds of the grid it runs on. Provably bounded shedding and retargeting under the modeled dynamics. Throughput as measured upside, not as a promise.
Synchronized multi-megawatt load swings from large training and inference fleets stress the transformers, feeders, and utility interconnects they sit behind. The control today is coarse: cap the cluster, trip the breaker, or overbuild the substation. None of it is provably bounded, and the operator finds out at the postmortem.
PLACEHOLDER COPY — replace bracketed fields with deal terms before sending.
A verified controller spine (load shedding, Lyapunov retargeting, telemetry) that carries a formal invariant: under the modeled power dynamics, the controller holds the facility within its declared peak and slew bounds. We pitch the safety proof as the product, scoped to its assumptions. Token-per-kilowatt-hour throughput is reported as measured data on top, not as the guarantee.
We have not found an incumbent shipping a formally verified power-bound control plane. Stated as a current finding, not a certainty.
VFX PLACEHOLDER — final visuals by [VFX dev]. This canvas sketch shows load (signal) against the facility bound (red) and the controller's response.
PLACEHOLDER — replace [Name] fields and bios.
Verified controller spine, building and kernel-checked. Live routing harness across Claude, OpenRouter, and SwitchYard.
First live grid pilot, real benchmark run against incumbents, telemetry hardening to sub-second power feedback.
[name@allodial.io]
[allodial.io]