Metered power
A dedicated circuit, a local isolator, and a meter. The number in the report is a measurement.
Architecture
Same computer. Two very different rooms.
In a hotel, the point of the heat is to be captured. In a locker, the point of the heat is to leave the parcels alone. The GPU is managed either way.
Principle · heat
Where the heat is made
A GPU is a heater that also computes. Electrical energy in, heat out, then a preheat loop. One kilowatt-hour at the chip is about one kilowatt-hour of heat. The meter, not the diagram, says what crossed the exchanger.
01
Power in
A metered circuit feeds the cabinet. What the meter records is what the GPU is allowed to draw.
02
GPU
The chip does the compute. Nearly all of that electricity leaves the silicon as heat. That is the source.
03
Cold plate
Liquid lifts the heat off the package. This loop stays on the IT side of the cabinet.
04
Exchanger
A double wall. Coolant and building water meet as temperatures, not as fluids.
05
Water and pool
Preheat for guest hot water, and the pool circuit where the plant can take it. Boilers still finish the setpoint.
Principle · compute
One training run, many rooms
A distributed AI server is not a new hall. A training job is split across nodes that already have an address — a hotel plant room, a locker bay — and the pieces come back as one run. Inference uses the same nodes between training jobs.
Training job
One run, sharded
Scheduler sends a shard out. Gradients and checkpoints return.
H1
Hotel plant
Shard A
L1
Street locker
Shard B
H2
Leisure hotel
Shard C
L1
Building locker
Shard D
Return path — the run stays one job. Parcel data, guest Wi-Fi, and payment networks are not on this path.
Revenue model
Zero cost to build. Server rental covers the expenditure.
The host does not fund construction. Arete builds the node, then rents the GPU. That rental is the income that carries the hardware, the power, and the service. The hotel receives heat and a report. It is not sold a share of the server.
01
Host pays £0 to build
No construction cost for the hotel or the locker operator. They provide the place: a plant room, or one bay.
02
Arete funds the node
Cabinet, GPU, exchanger or sealed bay, meter, and the connection. The hardware stays ours.
03
The server is rented
Training runs and inference are sold as managed GPU time. That rental is the income.
04
Rent covers the cost
Hardware, electricity, and service are carried by the server rental — not by a bill to the host for the build.
The site agreement sets the meter and the term. This is the structure — host build at no construction cost, server rental as the income — not a forecast for a building we have not surveyed.
Hotel and pool
Two loops that do not mix.
IT coolant
Cold plate on the GPU, a pump, and a closed loop inside the cabinet. Drip tray. Smoke and leak detection. Power isolation if either trips.
Double-wall exchanger
Coolant and building water meet only through the exchanger, double-wall where the design requires it.
Building preheat
Onward to the domestic hot-water preheat, and to the pool circuit on sites where that connection is real. Final temperature stays with existing plant. Legionella control stays with the hotel.
Locker bay
One compartment, thermally finished.
Sealed module
The GPU lives in a closed bay with its own coolant. It is not an open card behind a parcel door.
Heat rejection
Heat leaves through that module’s own path. Neighbouring compartments stay at locker ambient. Parcels are not a radiator.
Optional building couple
If a locker already sits against a plant room — some hotels have both — the survey can ask whether heat can be coupled. Street cabinets are designed to reject heat, not to warm a pavement.
Controls on every node
Remote management
BMC, signed updates, disk encryption, secure boot. The cabinet is locked. Tamper raises an alert.
Fail quiet
If cooling, leak detection, or the building side is unhappy, compute stops. The hotel’s water does not.
No guest path
Segregated network. Not guest Wi-Fi, not the property system, not payments, not parcel data.