702 W
Measured IT draw, split by site.
Site A568 W
Site B134 W
RAPL · IPMI · GPU counters
A UK project measuring AI's footprint
What does AI actually do to the machines it runs on?
Every AI answer draws power, and almost all of that power ends up as heat. Alt Horizon is a UK project measuring both on real hardware: from a single prompt, through the GPU, to the electricity meter and the heat in the room.
Measured, not estimated. Every figure comes from a counter or a meter on hardware we run.
Measured
“Summarise this 4-page supplier contract and flag the termination clauses.”
Document summary · qwen3:14b on RTX 5090
Tokens
GPU compute 67.51 tok/s measured · n=279
Energy measured, NVML counter
Carbon × 130.96 g/kWh · DESNZ
Energy cost @ £0.2635/kWh · electricity only
Same prompt, cloud
🔒 energy · 🔒 carbon · 🔒 heat — you get a bill
Energy · 24h
20.1 kWh
Scope 2 · 24h
2.6 kg
Cost · 24h
£5.38
One prompt, measured end to end: qwen3:14b on an RTX 5090, 67.51 tok/s at 210.41 W, averaged over 279 runs (2026-02-28 to 2026-04-29). The job is a 2,000-token document in, a 500-token summary out. Duration, energy, carbon and cost are worked out from those readings, never typed in: carbon at the UK government's 2026 grid factor, cost at 26.35 p/kWh (electricity only). The 24-hour figures are our own machines, 16 of them, as of 16 July 2026. The full walk-through.
Why we're doing this
Cloud AI shows you tokens and a bill. Everything in between is hidden.
The energy, the heat and the load on the building all happen somewhere, just not where you can see them. So we run AI models on our own hardware, where every step is visible, and measure what AI adds to the infrastructure underneath it.
The method
How we measure.
Five layers, each read from real hardware and each checked against the one above it.
-
At the chip
An agent on each machine reads the hardware's own energy counters: the GPU (NVML), the CPU (RAPL) and the server (IPMI).
-
Per answer
We read the GPU's energy counter just before and just after each inference. The difference is that answer's energy. Power readings minutes apart can't see an answer that takes seconds.
-
At the wall
A clamp meter on the circuit measures everything the room draws. Comparing it with the machines' own readings separates the IT load from everything else.
-
As heat
Almost all the electricity a computer uses leaves as heat, so every kWh we measure is also heat going into the room. We track temperatures alongside it.
-
As carbon
Energy becomes CO₂e using the UK government's published grid factor for electricity, updated every year.
Token→
GPU energy→
Machine→
Circuit→
Heat→
kg CO₂e→
£ cost
Cloud AI shows you the first and last step. Running it ourselves lets us measure every one in between.
What we're seeing
From the chip to the room.
A snapshot of our own machines: what they draw, the heat that comes off them, and how much of the circuit they account for.
2,395 BTU/hr
The same 702 W, leaving the machines as heat.
Power in702 W
Heat out≈ 702 W
Power counters · thermal sensors
80%
of the metered circuit is the computers
Whole circuit (clamp meter)877 W
IT, measured702 W
CT clamp · hardware
Readings from Alt Horizon's own machines across two sites, as of 16 July 2026. Heat is power converted at 3.412 BTU/hr per watt.
What it costs
The same job, priced four ways.
Cost isn't the focus of the project, but it's the number most people ask about. Here's the document-summary job priced in the cloud and on our own hardware. Running it yourself isn't the cheapest per token; it's the one where you can see the power, the heat and the carbon behind the price.
| Deployment | Peak power | Total energy (PUE-adj) | Heat output | Carbon / mo | £ / 1M tok | Capability |
|---|---|---|---|---|---|---|
| Cloud · frontier | 🔒 hidden | 🔒 hidden | 🔒 hidden | 🔒 hidden | £6.69 | Highest |
| Cloud · cheapest | 🔒 hidden | 🔒 hidden | 🔒 hidden | 🔒 hidden | £0.12 | Good |
| Self-hosted · 8B | ≈ 320 W | ≈ 210 kWh | ≈ 210 kWh | ≈ 27 kg | £0.60 | Good |
| Hybrid · 60/40 | ≈ 190 W | ≈ 125 kWh | ≈ 125 kWh | ≈ 16 kg | £0.41 | Tunable |
Cloud £/1M-token prices are collected from each provider weekly. Self-hosted rates are worked out from real tokens-per-second on real GPUs, not spec sheets (June 2026), and include the hardware as well as the electricity. The power, energy and heat figures for the self-hosted rows are worked examples, not readings. Heat output equals energy drawn: essentially all the electricity a GPU uses leaves as heat. Carbon at the UK government's 2026 grid factor (0.13096 kg CO₂e/kWh, location-based), gov.uk.
The tooling
Horizon Portal.
It's what we built to take these measurements: a lightweight agent on each machine that reads the power counters and thermal sensors and reports outbound-only, plus a dashboard that brings power, heat and circuit readings together by site. It runs on Windows, macOS, Linux, Proxmox and pfSense.
Notes
The method, step by step.
Part 2
Measuring the energy of a single AI answer
Why we read the energy counter before and after each answer instead of sampling power.
Read → Part 4Where the energy goes: heat
Why every watt a computer draws becomes heat, and how we track it.
Read → Part 6One prompt, end to end
A contract summary followed from the token to the electricity meter.
Read →Get in touch
Interested in the method?
Questions about how we measure, or want to measure your own setup the same way? Email us.
Horizon Portal is available to run on your own machines. Ask us.