LoadShare finds unused electrical capacity at a building and shows whether AI computing could use it without putting the building or the grid at risk. This page explains the software and is plain about what is still unproven.
Patent pending · Results are simulated, on public NREL data and customer bills where supplied
| Status | What | Evidence |
|---|---|---|
| Working today | Address screening, bill and meter data upload, full report, materials list, model journal | Live on this site and used on real addresses |
| Checked against real data | Reading power bills | All 24 billing readings on one commercial tenant’s scanned Duke Energy bills (three accounts, eight billing periods) matched the bills |
| Checked against real data | Building load estimates | Compared with NREL ResStock and ComStock reference buildings. A 10% safety margin is added to every estimated load |
| Simulated | Computers hosted, hours available, demand-response results, owner income | A year of 15-minute modeling on public data and any bills supplied. No site has been built |
| In development | Site controller that runs at a building | Tested only against simulated equipment so far |
| Not yet tested | Real shutoff times, the hardware cutoff on a live feeder, field performance on a real transformer | These need bench hardware and then a pilot site |
The address gives the location, outline and local weather.
A year of 15-minute building load, from building type and weather, resized to the customer’s own bills or meter data when supplied.
Each 15 minutes, the model checks the transformer against its rating, using an IEEE C57.91 temperature calculation, and adds computers only while the transformer stays within its rating.
The result, grid effect, economics and materials list are produced and the run is written to the journal.
| Location | US Census Geocoder |
| Building outline and type | OpenStreetMap, or what the user enters |
| Weather | NASA POWER hourly temperature |
| Typical building use | U.S. EIA building averages, checked against NREL reference buildings |
| Actual use (optional) | The customer’s bills (PDF or photo) or utility meter data (CSV or Green Button) |
| Transformer size (optional) | What the user enters. Otherwise the smallest standard size that carries the building |
| Hosting capacity | How many AI computers fit, and the share of the year they could run |
| Grid effect | Transformer loading and winding temperature with no computers, uncontrolled computers, and managed computers |
| Demand response | Whether computing stands down for the 10 hottest weekday afternoons of summer, used as modeled grid events |
| Economics | Illustrative owner income and per-computer economics |
| Materials and hookup | A preliminary list of equipment, circuits, feeder size and labor hours for pricing. Not for construction |
| Record | A numbered, fingerprinted journal entry for the run |
These describe the design. They still have to be proven through testing.
Computing gets only what is left after the building’s own use, and it is the first load to go.
In order of precedence: a lockable disconnect, a hardware safety chain that works without software, utility control, then workload control.
Nothing is installed until a site study and the utility confirm it is safe.
Every run is saved with its inputs, data sources, model version and results. Each entry is chained to the one before it, so a later edit would show.
A file is read once and not kept. Only usage figures, utility, rate plan and ZIP code are saved. Bill images and PDFs are read by our AI service provider.
Kept apart from the journal, and used only to send a report and follow up. See the privacy policy.
A screening is an estimate, not a utility interconnection study. The real transformer may be larger, smaller or shared with others, and only the accounts uploaded are counted.