Skip to content

When Will AI Pay for Itself?

What AI costs, and what it is starting to give back.

Joe Houghton · UCD Smurfit · Houghton Consulting
Evidence checked 23 September 2026

In the 1840s Britain spent almost 7% of everything it earned on railways, most investors lost their money, and the lines are still carrying passengers today. AI is in its railway years: enormous, wasteful, speculative spending, with real damage to water, power, workers and communities on one side and real gains in medicine, science, learning and energy on the other. This page puts both sides on one ledger, with every number sourced, so you can judge the balance yourself.

Cover: Andrew J. Russell, Union Pacific construction train, Bear River, c. 1868, Beinecke Rare Book and Manuscript Library, Yale. Public domain. Source

One number, to begin with

Price of a GPT-3.5-level answerTwo measured prices on a logarithmic scale: $20.00 per million tokens in Nov 2022 and $0.07 in Oct 2024.$20$2$0.20$20.00$0.07Nov 2022Oct 2024

The price of a GPT-3.5-level answer fell 280-fold in 23 months, from $20.00 to $0.07 per million tokens. Two measured prices; the path between them is not plotted.Source: Stanford AI Index 2025 / Epoch AITwo sources: Two independent organisations agree.Epoch's own warning: the steepest falls are recent and may not persist. A 2026 replication finds about 5x a year since 2024.

Then, the cutting. Now, the hall.

Tring Cutting, on the London and Birmingham Railway, as John Cooke Bourne drew it. The hall is CERN's data centre, standing in here for the halls now being built for AI. This page asks whether the second build-out ends the way the first did.

A vast cutting gouged through chalk hills, its sloped walls lined with barrow runs and tiny figures of labourers.
1839Then
John Cooke Bourne, Tring Cutting, 1839. Public domain. Source
Long rows of low server cabinets under a bright ribbed ceiling, receding to the far wall of a data hall.
2024Now
Mramoeba, Data storage centre, CERN, 2024. CC0. Source

The curve

The cost of a fixed amount of "intelligence" has fallen faster than any commodity in the PC or dotcom eras: roughly ten-fold a year on the cautious reading, far faster on some tasks. The energy per task is falling almost as fast. Total spending and total electricity are rising anyway, because the amount of computing bought grows faster than the price falls. All four charts below say the same thing from different angles.

The price of intelligence

The price of intelligenceAll three series fall on a logarithmic scale: a GPT-3.5-level answer, GPT-4-level capability, and the GPT-4 lineage's list price.$20$2$0.20GPT-3.5GPT-4GPT-4 lineageNov 2022Nov 2023May 2024Apr 2025
  • Price of a GPT-3.5-level answer
  • GPT-4-level capability
  • GPT-4 lineage list price

Frontier reasoning models are priced about 30x above this curve and have not followed it.

Three measures of the same collapse: a mid-tier model's answer, GPT-4-level capability, and OpenAI's own list prices for the GPT-4 lineage. Frontier reasoning models are the exception - priced well above this curve and, so far, not following it down.Source: Stanford AI Index 2025 / Epoch AI, Epoch AI, Mar 2025, OpenAI list prices via arXiv 2603.28576Two sources: Two independent organisations agree.One source: A single primary document, usually an official statistic or a peer-reviewed paper.Epoch's own warning: the steepest falls are recent and may not persist. A 2026 replication finds about 5x a year since 2024.

Energy: efficiency up, demand up

Per-unit efficiency is improving faster than almost any technology in energy history. Total consumption is rising anyway.

Compute up, electricity barely

Data-centre compute vs electricity, 2010-2018+0%+200%+400%+600%+550%+6%Data-centre computedata-centre electricity

Data-centre compute rose 550% from 2010 to 2018; electricity for the same data centres rose just 6%.Source: Masanet et al., Science, 2020Two sources: Two independent organisations agree.

Cooling overhead, industry average

Industry-average PUE, 2007 to 20251.01.52.02.51.091.0820072025

Industry-average PUE fell from 2.50 in 2007 to 1.54 in 2025. Google's own fleet and Dublin campus already run below it.Source: Uptime InstituteTwo sources: Two independent organisations agree.

Energy per prompt, falling fast

Energy per median Gemini text prompt2 Wh0.2 WhMay 2024 ~7.9 Wh (implied by 33x)May 2025 0.24 WhMay 2024May 2025

Energy per median Gemini text prompt fell from about 7.9 Wh (implied by Google's own 33x figure) to a measured 0.24 Wh in a year.Source: GoogleVendor: The company's own figure, unaudited.

Total electricity, rising anyway

Google's total data-centre electricity0255030.8 TWh42.4 TWh20242025

Google's own data-centre electricity still rose 37% in the same year efficiency improved 33-fold - more computing bought outpaces the price and energy falls.Source: GoogleTwo sources: Two independent organisations agree.

Where it goes

Where global data-centre electricity goesMeasured electricity rising from 205 TWh in 2018 to 485 TWh in 2025, a single projected central case of 950 TWh in 2030, then four 2035 scenarios ranging from 700 to 1,700 TWh.05001,0001,500205 TWh415 TWh485 TWh950 TWh (central case)Headwinds, 700 TWhHigh Efficiency, 970 TWhBase, 1,200 TWhLift-Off, 1,700 TWh201820242025203020353% of world electricity

Japan's total electricity today ≈ 2030 central case

Where the emissions go, and where they could come off

Data-centre CO2, and the Widespread Adoption reduction2024180 Mt2035, Base300 Mt2035, Lift-Off500 Mt1,400 Mt avoidedWidespread Adoption Case

the 1,400 Mt Widespread Adoption reduction

Source: IEAProjection: A scenario, not a measurement.

Every serious forecaster has data-centre electricity roughly doubling by 2030. The argument is about 2035, where the IEA's own cases differ 2.4-fold.Source: Masanet, IEATwo 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.

Ireland: data centres' share of metered electricity

Data centres' share of Ireland's metered electricity, 2015 to 2025A bar chart rising from 5% in 2015 to 23% in 2025, above both the urban and rural household reference lines.0%5%10%15%20%25%urban households 18% (2025)rural households 9% (2025)EirGrid stops new Dublin-area connections until at least 2028CRU requires new data centres to bring their own generation or storage201520172019202120232025

Up 518% in ten years. Data centres now use more electricity than every urban home in the State combined.Source: CSO, Data Centres Metered Electricity Consumption 2025One source: A single primary document, usually an official statistic or a peer-reviewed paper.

Ten years, one line

  1. 2016

    DeepMind cuts cooling energy at one Google site (40%, never audited)

    Vendor: The company's own figure, unaudited.Source

  2. 2017

    The Transformer paper; training cost about $930

    One source: A single primary document, usually an official statistic or a peer-reviewed paper.Source

  3. 2018

    Data centres use ~205 TWh, 1% of world electricity, only 6% above 2010

    Two sources: Two independent organisations agree.Source

  4. 2020

    GPT-3 trained (552 t CO2e); API at $60 per million tokens

    Two sources: Two independent organisations agree.One source: A single primary document, usually an official statistic or a peer-reviewed paper.Source

  5. 2022

    ChatGPT launches; GPT-3.5-level at $20 per million tokens

    Two sources: Two independent organisations agree.Source

  6. 2023

    US data centres 4.4% of US electricity; Ireland 21%

    Two sources: Two independent organisations agree.One source: A single primary document, usually an official statistic or a peer-reviewed paper.Source

  7. 2024

    GPT-3.5-level at $0.07; global data centres 415 TWh; 3.7 million GPUs shipped; IEA counts data centres among the fastest-growing emitters

    Two sources: Two independent organisations agree.One source: A single primary document, usually an official statistic or a peer-reviewed paper.Source

  8. 2025

    IEA Energy and AI; Google reports 0.24 Wh per prompt and 33x gain; data centres 485 TWh; MASAI final results; ECMWF's AI forecast goes operational; Ireland 23%

    One source: A single primary document, usually an official statistic or a peer-reviewed paper.Vendor: The company's own figure, unaudited.Source

  9. 2026

    IEA revises 2030 to 950 TWh; hyperscalers guide $720-745bn of capex; CRU's new Irish connection rules take effect

    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.Source

Ireland, where the costs cluster and the benefits are starting