EIP / 01System analysis · 2026 benchmarks

AI energy comparison

Inference type

Everyday useKitchen / Transparent electrical calculation

Make one mug of tea

Energy demand25.8 Wh

AI equivalent

287

AI images

≈0.09 Wh/image · SANA 1.5 1.6B · B200 minimum-energy configuration · GPU energy only

Everyday-energy scale

Logarithmic range / 20 Wh — 2.4 kWh

20 Wh501002505001 kWh2.4 kWh

Working assumption

Heat 250 mL of water from 20–100°C in a 90%-efficient electric kettle.

EIP / 02Framing · Scale of the question

Read before quoting a number

The unit gets smaller while the system gets bigger.

AI’s environmental problem is real, but the debate is often aimed at the wrong scale. Efficient modern inference can make a single image or text response surprisingly cheap in energy terms, while total AI electricity demand can still grow enormously as usage, model capability, and data-centre capacity expand. That means the serious environmental question is increasingly not whether an ordinary person should feel guilty for generating an image, but how efficiently AI infrastructure runs, what energy sources power it, where new generation and grid capacity come from, and who is accountable for those choices. The unit can get smaller while the system gets bigger. So the conversation should move upstream: don’t police the prompt; govern the megawatts.

Ten working positions

Per-request AI energy can be very small, especially on efficient modern hardware.
Small per-request energy does not mean small total energy use.
AI demand can rise faster than efficiency improves.
The real environmental scale problem is data centres, not individual prompts.
What powers those data centres matters as much as how much electricity they use.
The same workload can have very different emissions depending on the grid and generation mix.
Users do not choose data-centre siting, cooling systems, grid connections, or power procurement.
Environmental accountability should focus increasingly on infrastructure decisions.
Individual use still contributes to demand, but it is not the whole system.
The key question is not “Should people generate?” but “What are we building to power all this generation?”

Core frame

Don’t police the prompt. Govern the megawatts.