Put idle power to work.
IdlePower · modular edge AI data centers · starting in PG&E, SCE and Con Edison territories
Commercial and industrial sites are wired for their busiest hour. IdlePower puts modular GPU sites into the capacity they don't use, and runs every site under its limit, minute by minute.
Use the power that's already wired in
Megawatts from many small sites
Every site under its limit
A new or bigger service from PG&E took 313 days on average, and 950 when the circuit upstream needed work. A site that's ready now needs neither. The pod runs on room the service already has.
Buyers want megawatts. Sites have kilowatts. The Modulo Orchestrator pools many small sites into one block of capacity, sold in firm and interruptible tiers.
Firm work stays below an 80% line. When the utility asks for less, flexible work steps back first, and Modulo Site AI holds the limit whatever the plan says.
PG&E's average time to deliver a new or bigger service
40 sites of 125 kW, sold as one pool
Illustrative. Blue is the power each site can give to GPU work right now.
Firm work never sits in the top 20% of a site's power
Example shares. Every share is set per site.
- Modulo Digital Twin data
- Modulo Site AI
- Modulo Orchestrator
- Settlement
One system, from finding a site to running the GPUs
The Modulo Orchestrator forecasts each site's spare power, plans GPU work into it and dispatches the battery, chargers and GPUs every minute. Switch the chart above between today and with IdlePower.
Modulo Digital Twin screens the commercial and industrial parcels in a metro and ranks each by the power a pod can use today and the room it can add later, with the evidence behind every number.
The Modulo Marketplace puts hosts, capital, builders, equipment and buyers on the same site record, and settlement pays each party from the meter.
Pods are 125 kW and 250 kW modular edge data centers. A Battery Pod covers peaks and grid events, and both move when a site's needs change.
Up to 160 GPUs in a few parking spaces
Pod 250 holds GPU servers, networking, cooling and controls in one 40 ft container. Pair it with a Battery Pod and the site rides through peaks and grid events.
Start with one pod. Add sites, not substations.
The grid has room for load that can flex
Duke University, Feb 2025
Duke University, Feb 2025
Source: Duke University Nicholas Institute, Rethinking Load Growth, via Latitude Media.
IdlePower works with IEEE 2030.5, OpenADR, OCPP, BACnet, Modbus, SunSpec, Kubernetes, vLLM, NVIDIA DCGM, UL 3141 and UL 9540.
Built for sites that already have the power
If your service is sized for a peak you only reach a few hours a day, the rest of the day that capacity sits unused. A pod puts it to work, and Modulo Site AI makes sure it never pushes you past your limit.
Busy overnight, quiet by day
Chargers sized for the overnight rush leave room through the day.
Wired for the hottest afternoon
Most days never come close to the peak the service was built for.
Big services, uneven loads
Sized for peak season and shift changes, with room the rest of the year.
Room between shifts
Shift and seasonal schedules leave capacity between the peaks.
Many buildings, one gate
Batteries can move between buildings without touching a public road.
Busy for a season
Farms and food processors that run hard at harvest and sit quiet the rest of the year.
Load shapes are illustrative. Hatched amber is capacity the service already has.
Three tiers of GPU capacity
The more your work can flex, the less you pay. Firm work runs straight through grid events; flexible work gives up a little speed or timing when the grid is tight.
- During a grid event
- Runs at full speed
- Where it sits
- Below the 80% line
- Good for
- Voice agents, robot model calls, real-time video
- During a grid event
- Slows by up to 10 to 25% for a few hours
- Where it sits
- In the top band, below the limit
- Good for
- Video alerts, image reads, fine-tuning, embedding refresh
- During a grid event
- Pauses, then finishes by its deadline
- Where it sits
- In the top band, first to step back
- Good for
- Batch inference with 24-hour turnaround
When sites run short of power or GPUs, overflow goes to cloud GPU pools.
Get started
Have a site with spare power?
Send a few bills and photos. We'll tell you what the site could host.
List your siteNeed GPU capacity?
Tell us your models, latency needs and deadlines, and which tier fits.
Reserve capacityWant to finance or build pods?
Commit at the stage that suits you, from a shortlisted site to a live one.
Partner with us