One prompt is not one number

Almost every argument about AI's footprint compares a single number for “AI” against a single number for something else. Both sides are wrong, because AI spans four orders of magnitude on its own. A short text answer and an agentic coding session are not the same activity, any more than a text message and a video call are.

So here is one axis. Everything sits at its measured position, with its real uncertainty shown. Tap anything to see the source.

Note on scope: the email figures include the electricity of both people's devices while they write and read. Google's AI figure explicitly excludes your device. The two are not perfectly like-for-like, and the email numbers are the more generous of the two.

Grams of CO₂-equivalent · each gridline is 10×

Finding 01The decade that matters is 100 g to 1 kg

A short text prompt really is negligible — 0.03 g, five drops of water, a microwave running for one second. That number is measured, published, and about as good as we have. It is also the number people quote when they want the conversation to end.

The honest version: as soon as you switch on a reasoning model or hand a task to an agent, you move up two to four decades and land in the same band as a cup of coffee, a mile of driving, and a hot shower.

That is the IEA's own framing. Efficiency per task is improving faster than almost anything in energy history, and simultaneously the tasks people ask for are getting far more expensive. Both are true. For a room full of engineers who will be running agents, not chatting, the second half is the half that applies to them.

Finding 02Text AI is not additional to normal internet use

This was the specific question worth answering, and the IEA answered it directly in April 2026: if every conventional internet search on earth were performed instead as a simple AI text query, it would add under 4 TWh a year — less than 1% of what data centres already consume.

Streaming makes the same point from the other direction. One hour of video is around 36 g on the IEA's central estimate, with credible studies spanning 36 to 440 g. And the largest single contributor is usually not the data centre at all — it is the screen in front of you. The same hour on a phone emits five to ten times less than on a large television. Most people's mental model of where digital impact lives is simply in the wrong place.

How many median text prompts equal one of these

One of these	≈ text prompts
1 spam email your filter catches	1
1 short email, laptop to laptop	10
1 long email, 13 min at the keyboard	570
1 hour of streaming video	1,200
1 cup of coffee	7,000
1 mile in a petrol car	13,000
1 load in the tumble dryer	38,000
1 beef burger	100,000
1 new smartphone (manufacturing)	2,300,000
Your refrigerator, running one year	6,300,000
One transatlantic round trip	53,000,000
Flip it: 100 text prompts a day for a year	≈ ⅓ burger

Finding 03Email is where this stops being about data

Email is worth putting on the scale because it breaks the assumption underneath most of these conversations. A short email sent laptop to laptop is 0.3 g. A long one — ten minutes to write, three to read — is 17 g. Same protocol, same infrastructure, same handful of kilobytes. The difference is fifty-seven fold, and essentially all of it is two people's laptops being switched on while a human sits there.

The attachment is not the variable that matters. Time at the machine is the variable that matters.

Which is why there is no honest figure on this chart for “one email with an attachment.” The number people quote — 50 g — comes from the 2010 first edition of How Bad Are Bananas?. Berners-Lee revised his entire email range down to 0.03–26 g in the 2020 edition and restructured it around minutes rather than megabytes. The old figure is on the chart in ochre, struck through, because it is still in circulation fifteen years later and someone in the room will have seen it.

There is a real implication here for how you frame AI, and it points in both directions at once. If the dominant cost of knowledge work is a person and their laptop being switched on, then anything that changes how long that takes matters more than the bytes moved — a prompt that saves a quarter of an hour of drafting, or an agent that runs for two hours while nobody is watching. That is a genuinely open question about a specific workflow, not a slogan, and it is worth putting to the room as a question rather than answering for them.

Finding 04What this chart does not say

• Per-person is the wrong lens for the real concern. Data centres are just over 1% of global electricity and about 0.5% of global CO₂, heading for roughly 3% of electricity by 2030 as consumption doubles from 485 to 950 TWh. Small globally — but around 21% of Ireland's electricity, 26% of Virginia's, and 79% of Dublin's. The legitimate ethical argument is about specific grids, water basins, and electricity bills, not about whether you personally sent too many prompts.
• Training and video generation are excluded. Google's per-prompt figure covers text inference only. Image and video generation are materially more expensive and have no comparable published measurement.
• The published numbers have contested edges. Google's water figure counts data-centre cooling but not the water consumed generating the electricity, and its carbon figure is market-based rather than location-based. Critics, notably Shaolei Ren at UC Riverside, argue this understates the real footprint. The number is the best available, not the last word.
• Not every item measures the same boundary. The email figures count both people's devices; Google's AI figure does not count yours. Streaming figures include your screen, which is usually the largest single term. Where a comparison looks decisive, check whether the two sides drew the box in the same place.
• Scary numbers deserve an audit trail. Two of the most-quoted figures in this debate are retired. “30 minutes of Netflix = 1.6 kg CO₂” came from a 2019 Shift Project report and is roughly 90× too high. “An email with an attachment = 50 g” is from a 2010 book its own author revised downward in 2020. Both are still circulating. Check the provenance of any figure here — including the ones on this page.

SourcesWhere every number comes from

AI per prompt — Google, “Measuring the environmental impact of AI inference,” Aug 2025: median Gemini Apps text prompt 0.24 Wh, 0.03 gCO₂e, 0.26 mL water.
Reasoning & agentic multipliers — IEA, Key Questions on Energy and AI, Apr 2026: video, reasoning and agentic tasks “can consume hundreds or thousands of times more energy per query than simple text generation.” Agentic session sized on EcoLogits' 100,000-token application-development benchmark.
Data centre totals — IEA, Energy and AI (2025) and Key Questions on Energy and AI (2026); Carbon Brief analysis of IEA data for sector shares.
Streaming — IEA / Carbon Brief, “Factcheck: what is the carbon footprint of streaming video on Netflix?” (36 g/hr central, revised Nov 2020); Carbon Trust 2021 (55 g/hr Europe); Borderstep 2020 (100–175 g/hr).
Email — Mike Berners-Lee, How Bad Are Bananas? second edition (2020), as tabulated by the Carbon Literacy Project: spam caught by filter 0.03 g; short email 0.2 g phone-to-phone and 0.3 g laptop-to-laptop; long email (10 min writing, 3 min reading) 17 g; blast to 100 people 26 g. The 50 g “email with attachment” figure is from the superseded 2010 first edition. Note that these figures include both users' device electricity, whereas Google's AI figure excludes end-user devices — so the email side of the comparison is the more generous one.
Everyday items — EPA (0.40 kg/mile driving; WaterSense); ENERGY STAR (dryer ≈ 3 kWh, refrigerator 350–600 kWh/yr); EIA (US household 10,500 kWh/yr); Poore & Nemecek 2018 (beef burger ≈ 3 kg); Apple environmental reports (smartphone ≈ 70 kg); Wynes & Nicholas 2017 (transatlantic round trip ≈ 1.6 t).
Grid conversion — electricity-based items costed at the US average grid intensity, ≈ 380 gCO₂e/kWh (Ember 2024 / Our World in Data). On a cleaner grid every electricity-based item on this chart falls, AI included.