How we measure what AI does to infrastructure.

Part 1 · September 2026

What we measure, and where each number comes from

Every machine in the project runs a small agent. It reads the hardware's own sensors and reports outbound-only to a UK-hosted database. Nothing is estimated from a spec sheet; if a machine can't report something, we record it as missing rather than guessing.

• GPU energy from NVIDIA's management library (NVML), which keeps a running total of the joules the card has used.
• CPU energy from RAPL, the processor's own energy counters.
• Whole-server power from IPMI, read from the power supplies on machines that have it.
• Temperatures from the machine's thermal sensors.
• The circuit from a clamp meter on the electricity supply feeding the room.

The agent runs on Windows, macOS, Linux, Proxmox and pfSense, so the picture covers the whole setup rather than just the GPU box: routers, storage and hypervisors all draw power too.

Part 2 · September 2026

Measuring the energy of a single AI answer

The obvious approach is to read the machine's power every few minutes and average it. That works for a server's daily total, but it can't tell you what one AI answer used. Our power readings come every five minutes; an answer takes between two and thirty seconds. A reading either lands in the middle of an answer and exaggerates it, or misses it entirely and reports nothing.

So for AI answers we don't sample power at all. The GPU keeps a running total of the energy it has used. We read that total just before the model starts and again just after it finishes. The difference is exactly the energy that answer took, however short it was.

Two details keep it honest:

• Overlapping answers share the card. The counter covers the whole GPU, so if two answers run at once, both see the same joules. We flag those and split the energy between them, never count it twice.
• We only measure local work. If a model is running on a different machine, reading our own GPU would measure an idle card, so we record nothing instead.

Part 3 · September 2026

From the GPU to the meter on the wall

Chip-level counters tell you what the computers draw. They don't tell you what the building pays for. To close that gap we put a clamp meter on the circuit that feeds our equipment and compare the two.

The remaining fifth is everything the machines can't see about themselves: losses in power supplies and adapters, network gear without its own sensors, and anything else on the same circuit. Where we know about other loads on the circuit, we subtract them. Comparing the two layers is also a check on the method: if the machines ever reported more than the meter, we'd know a reading was wrong.

Part 4 · September 2026

Where the energy goes: heat

A computer doesn't store energy, and very little of it leaves as light, sound or movement. Essentially every watt it draws ends up as heat in the room. That's the first law of thermodynamics, and it means our power measurements are heat measurements too.

We convert at 3.412 BTU/hr per watt, the unit heating and cooling equipment is rated in, and record temperatures alongside so we can see how the heat builds up where the machines sit. An AI workload changes the heat load in a way you can feel: a GPU busy writing answers at around 210 W gives off about a fifth of what a 1 kW electric heater does, for as long as it's working.

What we don't do yet is put a value on that heat. Whether it can usefully warm a space, or how much cooling it costs to remove, depends on the building, and that's a question for later.

Part 5 · September 2026

Turning electricity into carbon

To turn kilowatt-hours into carbon we use the UK government's published conversion factor for grid electricity (DESNZ, location-based, Scope 2). For 2026 it is 0.13096 kg CO₂e per kWh.

The factor is published every June and falls most years as the grid adds more renewable power, so we read each year's value from the government's own data file rather than carrying the last one forward. If a year's figure isn't out yet, we use the latest one and label which year it came from.

On our own machines that means 20.1 kWh over 24 hours works out at about 2.6 kg CO₂e.

Part 6 · September 2026

One prompt, end to end

Putting the layers together, here is one job followed all the way through: summarising a 4-page supplier contract and flagging the termination clauses. That's 2,000 tokens in and 500 tokens out, run on qwen3:14b on an RTX 5090.

Across 279 runs between 2026-02-28 and 2026-04-29, the card generated 67.51 tokens a second at an average draw of 210.41 W.

One answer is tiny. The point is what it adds up to: a thousand of these is 0.43 kWh, about 57 g of CO₂e and 11p of electricity, and every one of those watt-hours becomes heat in the room.

These figures cover the GPU while it writes the answer; reading the 2,000-token document is quick by comparison and isn't included. Electricity is priced at 26.35 p/kWh. The cost of the hardware itself is covered in what it costs on the homepage.