# What happens to an AI workload when the power goes out?

We ran the evening outage in Watt2Work SIM’s local energy game, looking at battery reserve, the power limit and unmet demand together.

Watt2Work · 2026-11-03 · Build notes · Local prototype
Canonical: https://watt2work.com/blog/sim-power-outage/
Language: en

It is easy to begin an AI system discussion with the model. This time, we are looking at what happens when the electricity stops. SIM’s small energy game is a starting point for exploring the connection between energy and work, one step at a time.

![Watt2Work SIM energy game at hour 19: 80 kW demand, 50 kWh served and 30 kWh unserved.](https://watt2work.com/work/sim-outage-en.jpg)

3 November 2026 · Screenshot from the running local application. Numbers are illustrative scenario outputs, not field measurements or customer returns.

## Question: Does stored energy keep the work running?

Stored energy and the power available in a single hour are different quantities. We wanted this experiment to make that distinction visible: how much demand can a battery serve while the grid is unavailable?

## Experiment: 19 steps in one scenario

We selected “Evening grid outage” in the local application. We kept GPU demand at 80 kW, held the battery for the first 18 steps, then switched to automatic discharge and advanced one hour. The screenshot shows the hour-19 display from that run.

The game defines a grid outage during hours 18–22. Battery capacity is 200 kWh, the discharge power limit is 50 kW and discharge efficiency is 95%. These are illustrative model assumptions, not measurements from an installed system.

## Observation: Energy remains, but some demand is unmet

For this step, the screen shows 50 kWh served and 30 kWh unserved. Even with energy left in the battery, the discharge power limit cannot meet the entire 80 kW demand. This is a calculated result under the selected assumptions.

The work count and cash balance are outputs of the same illustrative model. No real model inference, GPU benchmark, customer revenue or field control took place.

## Next: Find the assumption that changes the decision

Next, we want to change demand and battery policy separately to examine which decisions affect the outcome. Saving inputs, replaying results with the same model version and keeping unknowns visible are part of how we work.

## Traces of the work

- [Scenario record from this run (JSON)](https://watt2work.com/work/sim-outage-2026-11-03.json)