When Will AI Pay for Itself?
What AI costs, and what it is starting to give back.
The ledger
Costs on the left, benefits on the right. Each card opens to show the evidence, its source and its badge, and a line saying what would change our mind. The costs column is not softened, and the benefits column does not include promises: every benefit is a measured deployment or a peer-reviewed result.
Benefits, as they land
- Sweden: AI finds 29% more breast cancers with 44% less radiologist reading (randomised, 105,934 women)
- NHS: stroke thrombectomy rates doubled across 28 hospitals using AI scan reading
- Johns Hopkins: 18.7% lower sepsis mortality when clinicians confirm the AI alert within three hours
- The first AI-designed drug cleared its Phase IIa safety trial, and also improved lung capacity
- AlphaFold: 200 million protein structures, 2 million users, a Nobel Prize
- Europe's AI weather model uses 1,000 times less energy per forecast
- Google's storm model beat human hurricane forecasters at 12-72 hours in 2025
- Five-day flood warnings, free, for river basins in 150+ countries
- AI recovers 0.7% of Google's worldwide compute, every day
- Britain's grid runs an AI solar forecast 2.8x more accurate than before, saving an estimated 300,000 t CO2 a year
- 164 French wastewater plants: 5.4% less electricity, audited, AI's own power under 1% of the saving
- EirGrid: AI forecasts reserve needs 38 hours ahead, with a potential 15% cost cut
- London: AI listening to water pipes found leaks worth 2,376 million litres in a year
- 9.9 billion African trees counted one by one from space
- Deforestation fell 18% where agencies acted on AI satellite alerts
- Three-quarters of industrial fishing vessels were invisible until AI found them
- California's AI cameras spotted 900+ wildfires before anyone rang 911
- Ireland's 11,000 raised bogs mapped for the first time; their condition predicted with 84% accuracy
- Five million US acres sprayed weed-by-weed: herbicide nearly halved
- IKEA halved its food waste: 20 million meals saved
- Meta's data centres stand on AI-designed concrete with 35-40% less carbon
- Steel: 500,000 lb of alloys not mined thanks to AI mix control
- IEA: today's AI applications could cut 1.4 billion tonnes of CO2 in 2035, three to four times data-centre emissions
- 3,500 methane super-emitter alerts sent; 40+ mitigations confirmed; nine in ten still unanswered
- Storms Darragh and Éowyn: 2,533 damaged ESB components found by AI in under 48 hours
Costs9
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.Projection: A scenario, not a measurement.
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.Vendor: The company's own figure, unaudited.
SDG 6.4
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.Vendor: The company's own figure, unaudited.
SDG 13.2
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.Projection: A scenario, not a measurement.
SDG 12.5 · partial
- Two sources: Two independent organisations agree.
SDG 6.4
SDG 13.2 · partial
- Two sources: Two independent organisations agree.Vendor: The company's own figure, unaudited.
SDG 7.3 · partial
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.
SDG 16.b · partial
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.
SDG 16.2 · partial
Benefits11
- Two sources: Two independent organisations agree.(a) peer-reviewed
SDG 3.4 · partial
- Two sources: Two independent organisations agree.One source: A single primary document, usually an official statistic or a peer-reviewed paper.(b) real deployment
SDG 13.1 · partial
- Two sources: Two independent organisations agree.Vendor: The company's own figure, unaudited.
SDG 7.3 · partial
- Vendor: The company's own figure, unaudited.(b) real deployment
SDG 6.4
- Two sources: Two independent organisations agree.(a) peer-reviewed
SDG 15.2
SDG 14.4 · partial
- Vendor: The company's own figure, unaudited.(b) real deployment
SDG 12.3
SDG 2.4 · partial
- Two sources: Two independent organisations agree.Vendor: The company's own figure, unaudited.(b) real deployment
SDG 9.4
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.Projection: A scenario, not a measurement.
SDG 13.2 · partial
SDG 11.5 · partial
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.(a) peer-reviewed
SDG 1.3 · partial
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.(a) peer-reviewed
SDG 4.1 · partial
- One source: A single primary document, usually an official statistic or a peer-reviewed paper.(a) peer-reviewed
SDG 8.2 · partial
Personal scale calculator
Send this many prompts a day, and here is the same habit translated three ways - each source measures something slightly different, so the honest answer is a range, not one number.
≈ 7,300 a year at this rate
Google view
text prompt, on-site water, vendor, median
1.75 kWh a year
18 kettle boils
117 min in a 900 W microwave
1.9 litres of water
Vendor: The company's own figure, unaudited.Source (opens in a new tab)Full-life view
Mistral, audited, 400-token answer, includes hardware and upstream
8.32 kg CO2e a year
Energy not disclosed
329 litres of water
One source: A single primary document, usually an official statistic or a peer-reviewed paper.Source (opens in a new tab)2023-era estimate
Li et al., GPT-3, includes power-station water
29.2 kWh a year
292 kettle boils
1,947 min in a 900 W microwave
123 litres of water
Two sources: Two independent organisations agree.Source (opens in a new tab)
Our constants
- boiling one litre in a kettle ≈ 100 Wh.raising 1 L from 20 °C to 100 °C takes 93 Wh, plus typical kettle losses - physics, not a dossier figure.
- a 900 W microwave ≈ 0.25 Wh per second.One source: A single primary document, usually an official statistic or a peer-reviewed paper.Source (opens in a new tab)
- a smartphone charge ≈ 22 Wh (Luccioni 2024 uses 0.022 kWh).Two sources: Two independent organisations agree.Source (opens in a new tab)
- an image generation ≈ 2.9 Wh (Luccioni 2024 mean).Two sources: Two independent organisations agree.Source (opens in a new tab)
Worth remembering
A single generated image costs about twelve text prompts. A reasoning or video task can cost hundreds to thousands of times more (IEA 2026).
The balance
Put the two columns on one scale and this is what the evidence supports today. This scale weighs the climate case, in tonnes of CO2, because that's the one place the evidence reduces to a single comparable unit; the social-goal cards on jobs, bias, learning and poverty don't, so they sit outside it - not because they don't matter, but because nothing here converts a wage into a tonne of carbon.
Left pan - Data-centre emissions 2035: 300-500 Mt CO2 (IEA)
Right pan - AI-enabled reductions 2035: up to 1,400 Mt CO2 (IEA Widespread Adoption Case)
| Pan | Reads |
|---|---|
| Data-centre emissions | Data-centre emissions 2035: 300-500 Mt CO2 (IEA) |
| AI-enabled reductions | AI-enabled reductions 2035: up to 1,400 Mt CO2 (IEA Widespread Adoption Case) |
What has to be true for the benefits to win
Efficiency keeps compounding.
The 10x-a-year fall in energy per task has to continue for years, and history says such curves flatten (Koomey's law slowed from a doubling every 1.6 years to every 2.6 after 2000).
One source: A single primary document, usually an official statistic or a peer-reviewed paper.The grid gets cleaner faster than AI grows.
Google's hourly carbon-free share was 66% in 2024 against a 100% goal for 2030; 143 GW of gas is planned beside US data centres.
Two sources: Two independent organisations agree.One source: A single primary document, usually an official statistic or a peer-reviewed paper.Projection: A scenario, not a measurement.The benefits get adopted, not just invented.
The IEA's 1.4 Gt depends on buildings, factories, grids and vehicles using tools that already exist. Nine in ten methane alerts currently go unanswered.
Two sources: Two independent organisations agree.Projection: A scenario, not a measurement.Local costs get priced in.
Ireland's new connection rules and Arizona's water ordinances are the first signs of this; the analogy with the railways is only fair if the third parties who pay are named.
The three biggest uncertainties
Whether 2035 demand is 700 TWh or 1,700 TWh (the IEA's own range).
Projection: A scenario, not a measurement.Whether the frontier models that don't follow the price curve stay a small share of use.
Whether any AI-discovered drug completes Phase III: none has yet.
Verdict
The costs are real, front-loaded and rising in absolute terms, and they fall hardest on the places where data centres cluster, Ireland among them. The benefits are no longer promises: randomised trials, operational weather models, one causally audited industrial saving and counted trees. The cost of each unit of AI is collapsing faster than any technology on record. On the evidence here, the benefits are starting to outweigh the costs, and the direction of travel favours them, but the ledger is not yet closed and it will not close on its own. The social-goal cards on jobs, bias, learning and poverty don't reduce to a single number the way carbon does, so they aren't on this scale. They're weighed card by card, on the ledger. The railways paid off because the lines stayed up after the money was lost. AI will pay off if the efficiency curve holds, the grid cleans up and the tools get used.
