Inference workload
Local or network-scheduled models run on liquid-cooled accelerators.
Compute and thermal infrastructure
Put your AI heat to work.
A liquid-cooled inference appliance that turns compute waste heat into hot water for buildings.
Electricity in
Inference out
Useful heat captured
The principle
Nearly every watt a processor consumes becomes heat.
Conventional servers reject that heat into the air. Compute Boiler moves it into a building's water loop, where it can do useful work.
1 kW
electrical input
Inference
useful computation
~1 kW heat
available before system losses
One electricity bill. Two useful outputs.
The hardware schedules inference around thermal demand, stores heat in water, and falls back gracefully when either side is idle.
Heat exchange
A closed coolant loop transfers processor heat without mixing fluids.
Hot-water store
A buffer tank captures heat for domestic water or space heating.
Stop paying to throw heat away.
Typical inference
Compute, cool, reject.
Electricity runs the model. More electricity moves the resulting heat outdoors.
Compute Boiler
Compute, capture, use.
The same thermal output preheats water and lowers the building's separate heating load.
Built for steady demand.
The best sites need computation and low-temperature heat throughout the day.
Multi-family housing
Shared hot-water loads create a predictable thermal baseline.
Hotels and hospitals
High occupancy keeps tanks cycling and heat recovery valuable.
Pools and fitness
Large water volumes can absorb long, steady compute runs.
District heat
Deploy many units near demand instead of concentrating waste heat.
Laundries and process water
Preheating offsets a direct, measurable operating cost.
Pilot program
Have heat demand? Let's put it to work.
We are mapping early pilot sites, infrastructure partners, and inference workloads.