Heat from compute, run by software
A GPU turns nearly every watt it draws into heat. SilentSwarm puts those GPUs to paid work wherever your units are installed, and tells you how much heat each one delivered.
Looking for GPU capacity rather than heat? SilentSwarm for compute buyers
Why compute belongs in the heating loop
A data centre spends energy to get rid of heat. A building spends energy to make it. A compute heater does both jobs with the same kilowatt-hour: the GPU does paid work, and its waste heat warms the building instead of a cooling tower.
- Same electricity, two products. Every kilowatt-hour runs a computation first and heats the building after. The GPU is an electric heater that does paid work on the side.
- A source the heat pump likes. A liquid-cooled GPU loop runs far warmer than outdoor air in winter. The warmer the source, the less the compressor has to lift.
- Many small sites, one cluster. No basement holds a data centre. SilentSwarm splits a model across machines in different buildings, so many modest units add up to one useful cluster.
How a compute heat pump fits together
You build the unit and the hydraulics. SilentSwarm is the software around it: it brings the unit online, gives it work, meters its energy and listens to the building's heat demand.
- GPU unitDoes AI work inside the building and connects out to the coordinator.
- Warm loopCarries the GPU's heat to the heat pump.
- Heat pumpLifts it to flow temperature and reports the building's heat demand.
- Heating and hot waterWhere the heat ends up.
- CoordinatorSchedules work from compute buyers and meters each unit's energy.
What the platform does today
Everything below is in the software now, not on a slide. What is still missing is listed further down.
Units join with one command
A unit runs one command at the factory or on site, then connects out to your coordinator. No port forwarding and no VPN to set up on the customer's router.
Heat metering per unit
Each unit reports its GPU power with every heartbeat. The coordinator adds it up into kilowatt-hours per unit, per fleet and in total, and keeps the power history.
A heat-demand signal
Your controller tells the platform how much heat the building wants, from 0 to 100 percent. When a job could run on more units than it needs, the units whose buildings want heat get the work first.
Fleet health you can explain
Every unit gets a health grade from uptime, latency and errors, with the reasons behind it. Spare units can be held in reserve for when one fails.
Work that buyers already want
Fine-tuning and running open language models, served through an OpenAI-compatible API - the work AI teams rent data-centre GPUs for today.
Fleets, customers, automation
Pools keep fleets apart, organizations keep customers apart, and every change is in the audit log. What the web interface does is also available over a REST API and to AI agents over MCP.
Live from the SilentSwarm pool
Read from the coordinator that serves this page, not typed into it. Machine numbers are withheld while the pool is small, and the heat figures count GPU power only - the whole machine gives off somewhat more.
Rough numbers for one unit
The arithmetic, with your figures. Drag a slider or type a value, and the year recalculates.
- Heat delivered per year
- -
- Electricity cost per year
- -
- Value of that heat per year
- -
- Compute revenue per year
- -
- Balance per year
- -
An illustration that does the arithmetic with your numbers - not a quote and not a promise of revenue. The defaults are example values; replace them with yours. It counts GPU power only, as the metering does; hardware, installation and the heat pump's own electricity are not in it.
The address bar keeps the values you set, so a link to this page shares them.
What exists, and what does not yet
SilentSwarm is in open beta. A pilot works best when both sides know where the edges are.
Available today
- Energy metering per unit, per fleet and in total, from the GPUs' own power readings.
- The heat-demand API and the placement that favours units whose building wants heat.
- One-command onboarding for units behind any router, with a health grade per unit.
- Open models served through an OpenAI-compatible API and fine-tuned across machines.
Not yet
- Following the demand by capping GPU power. That needs driver-level rights on the unit, which the worker deliberately does not have today.
- Billing and payouts for compute sold from your fleet.
- Calibrated heat metering. The figures come from the GPUs' sensors, not from a certified heat meter.
- Hardware. We write the software; the unit, the hydraulics and the heat pump are yours.
Questions heating makers ask
- How much of a GPU's electricity becomes heat?
- Nearly all of it. The power a GPU draws ends up as heat in the room or the coolant loop, which is why the platform meters each unit's GPU power: the kilowatt-hours of heat have a source.
- How is the heat measured?
- Each unit reports its GPU power with every heartbeat, and the coordinator adds it up into kilowatt-hours per unit, per fleet and in total. The figures come from the GPUs' own sensors, not from a certified heat meter, and count GPU power only.
- Does a unit need port forwarding or a VPN?
- No. A unit connects out to the coordinator, so it works behind an ordinary home router without any change to it.
- Can the platform follow the building's heat demand?
- In part. Your controller reports the demand from 0 to 100 percent, and units whose building wants heat get work first. Capping GPU power to follow the demand more closely needs driver-level rights that the worker deliberately does not have today.
- Who builds the hardware?
- You do. We write the software; the unit, the hydraulics and the heat pump are yours. Running SilentSwarm in a commercial product needs a licence from NexPatch AI.
Building a compute heater? Let's talk.
Tell us about your unit, your timeline and how many sites you have in mind. We answer every message ourselves.
We use your details only to answer you - see the privacy policy. SilentSwarm is source-available under the PolyForm Noncommercial licence; commercial use needs a licence from NexPatch AI, which you can ask about in the same message.