AI and Energy — What Is Known and What Is Not

28 August 2026

About this record — This page is a document produced by User 1 together with AI assistants, not a transcript of a conversation. It is not fiction: the work it reports on was actually carried out on the dates shown, and the human participants are real people appearing under the labels User 1 and User 2 (see Reading Notes).

The closing letters are in the appendix; the records are in the ledger.

Chapter 0. Introduction

What this report is

Using AI uses electricity. Ask how much, and figures turn up quickly. The figures then disagree with each other. One is small, another is large. It looks as though somebody must be wrong.

Follow them back, and in most cases nobody is. They are measuring different amounts of the same thing.

This report sets out where the numbers exist and where they do not, for the energy used when someone uses an AI service. It is material for deciding what to do about your own use. It does not argue for using less, and it does not argue that the question can be set aside.

How to read the figures

The subject is electrical energy, and the unit throughout is the watt-hour (Wh). One watt-hour is the energy used by a one-watt device running for one hour.

Every figure here carries a statement of how much was counted. A figure that counts the computation alone, a figure that adds the cooling of the building, and a figure that adds the manufacture of the equipment are different figures, even when they describe the same task in the same unit. Chapter 1 sets out that distinction before any figures appear. Figures from different scopes are not added together or divided into one another.

Three rules govern the writing. No evaluative adjective in place of a scale. Not more than you would think, not less than it sounds; the figure, and what the figure includes. Descriptive terms such as lower, higher and large data centre remain where they identify an ordered value or a technical category. No single comparison object. The same figure gives a different impression depending on what sits next to it, so this report sets out several reference points and leaves the choice of baseline open. No connective that supplies a relationship the evidence has not established. When two facts are placed side by side, the writing does not join them with a word that implies cause or judgement. Their relation remains for the reader to assess.

What is left out

Water. Data centres use water for cooling, and how much depends on the cooling design, the site and the season. It cannot be converted from an energy figure by a fixed ratio. This report requires every figure it prints to state what it includes; applying that rule to itself, it does not print water volumes. What each company publishes about water is still recorded as one of the disclosure items.

Conversion to emissions. The same watt-hour carries different emissions depending on where and when the electricity was generated. Figures in the main text stay in watt-hours. How much they differ between countries is handled in chapter 7 and nowhere else.

Who made it

This report was produced by one person and five named AI assistants working through three model families. Claude and Claude Code are listed separately because they performed separate roles, as are ChatGPT and Codex. The assistants did the research, record keeping and writing. The person set the readership, scope and form of publication, carried the work between systems, and retained final approval. Technical and editorial decisions were raised by several participants and recorded in the ledger. Who did what is set out at the end.

One thing belongs here rather than there. Chapter 4 tabulates what seven companies publish about the energy and related environmental effects of the AI services they sell to the public. Assistants built by three of those companies helped assemble the table. The assistant editing this sentence was built by one of them.

Because of that, the records behind the table are published alongside this report. Anyone unconvinced by how the table reads can go to the records and check. Chapter 4 sets out the detail.

Chapter 1. What is one use?

To answer "how many watt-hours does one use of AI cost", two things have to be settled first. What counts as one use. And what counts as part of it.

Settle them differently and the same task comes out as a different number.

What counts as part of it

You send one instruction to an AI service and an answer comes back. Set out everything drawing electricity in between, and it falls into three groups.

The work itself. The accelerator doing the computation, the host processor driving it, memory, storage, and loading the model. This report calls that tier A.

Keeping the service running. Cooling the building, the electricity lost in power conversion, networking, and the equipment held ready while no instruction is arriving. Equipment that is powered on draws electricity during the hours nobody is using it. This report calls that tier B.

Everything before and around. The energy used to train the model, divided across each use. Manufacturing the equipment. Constructing the building. The device in the reader's hands. This report calls that tier C.

Where a given document draws its line varies. A tier A figure and a tier B figure are not directly comparable, even for the same task.

How the difference looks in practice

Take two of the measurements this report uses.

The first was made by running publicly available models on the researchers' own hardware. That measurement fixes the facility efficiency coefficient at 1.0, so cooling and power conversion are in effect absent from it. It does include the electricity drawn by seven idle accelerators sitting in the same machine.

The second was made by a company measuring its own production service. That one includes cooling and power conversion. It also includes the equipment reserved for the product and sitting idle. It leaves out network traffic to the outside.

The first is tier A, the second tier B. The second counts a wider range.

Count more things and the figure for the same task rises. So two figures with different scopes cannot be placed side by side and read as one service being more efficient than the other.

What counts as one use

The second question is separate.

One use can mean sending one instruction. It can mean one exchange, from opening to closing. It can mean one finished thing.

For images the difference matters. Sometimes one image is generated and used. Sometimes several are generated and one is picked. Both produce one image to keep, and the electricity used differs.

The measurements in this report are per generated item. How many generations it takes to reach one finished thing is a question about use, not about measurement. Chapter 3 takes it up.

How this report handles it

From here on, each figure states whether it covers tier A, tier B or tier C. Figures from different tiers are not added together or divided into one another.

For a figure that carries no statement of scope, the scope cannot be recovered afterwards. Such figures are not used as values here. That they circulate is itself the subject of chapter 4.

Chapter 2. What is known: figures on different scales

Three kinds of task came out of this research with measurements that state their conditions and allow their scope to be identified: text, images and video. What follows are those three records.

Text

Google measured its own production service and published the result. For one text exchange in the Gemini Apps product, the median is 0.24 Wh. The figure covers one month, May 2025.

The scope is tier B. Accelerator, host processor and memory, plus cooling and power conversion, plus equipment reserved for the product and sitting idle. Network traffic to the outside is left out.

This is a company reporting on itself. No outside party verified it. The model version is not published, and neither is the number of output tokens. It is a median, and nothing is published about the upper end of the distribution.

Images

A research group ran eight publicly available models on hardware they set up themselves. Per generated image, the median is 1.35 Wh and the mean is 2.907 Wh. The two differ because the distribution has a long right tail. Quoting one without the other leaves a different impression, so both appear together.

The hardware was a single A100-generation accelerator. Requests were not batched. The scope is tier A, with the facility efficiency coefficient fixed at 1.0. The seven idle accelerators in the same machine are included.

These are figures for self-hosted open models. They are not figures for using a commercial image service.

Video

This one also comes from running an open model on the researchers' own hardware. One video at 720x1280, 81 frames, 15 frames per second, 5.4 seconds long: about 90.5 Wh. The hardware was an H100-generation accelerator. It is one configuration, and the distribution is unknown.

The total breaks down into 78.8 Wh for the accelerator, 7.4 Wh for the host processor and 4.3 Wh for memory. It is the sum of those three component means, and the memory component is estimated rather than measured.

The scope is tier A. These are not figures for a commercial video service either. Whether a video twice as long costs twice the energy does not follow from this measurement.

What the three support

Among these three records, whose hardware and serving conditions differ, the production-service text figure is the lowest and the unbatched, self-hosted video figure the highest. That ordering describes these three records; it is not a general ranking of text, image and video tasks.

The stated scope difference does not explain why the text record is lower. Text is the record counting the wider tier B scope. The image and video records use the narrower tier A scope and still report higher figures. Adding the omitted facility components while holding the other conditions fixed would raise those two figures. A fully like-for-like comparison could also change batching, utilisation, hardware and serving conditions, so its result cannot be inferred from these records.

What the three do not support

No ratio is stated.

The three differ in the hardware measured, in how the models were run, and in whether the setting was a self-hosted environment or a company's production service. A ratio would fold all of that into one number, and a single number travels on its own.

The image and video measurements were made without batching requests. Their records specify that they are not to be applied to production serving. They cannot be read as figures for the image or video services a reader uses day to day.

These are also the measurements this research was able to record. A different set of measurements could produce a different ordering.

Three everyday amounts

As a set of marks for reading the figures above, here are three amounts of electricity from ordinary life.

• An 8 W LED bulb running for one hour: 8 Wh
• The built-in battery of a 13-inch MacBook Air (M4, 2025): 53.8 Wh
• One cycle of the average dishwasher sold in the EU in 2020: 0.8 kWh, or 800 Wh

None of these says that one use of AI is the same as one use of the appliance. The bulb figure is a specific product's rated power multiplied by an hour, not household use as measured. The laptop figure is the nominal energy the battery holds, not the electricity a full charge draws from the wall. The dishwasher figure is an average across products sold in 2020, not a particular current model.

Since these were not measured under matching conditions, they are not summed and the AI figures are not divided by them. Three separated marks — 8 Wh, 53.8 Wh, 800 Wh — are set out so that a reader can pick the scale closest to their own use.

Chapter 3. How much is your usage?

Every figure in chapter 2 is per generated item. One text exchange, one image, one video.

Use rarely stops there.

The image case

Image generation does not always end with one image. Several come out, and one gets used. If none of them work, the instruction is rewritten and several more come out.

How many images it takes to arrive at one that gets kept: this research looked for a published statistic and did not find one.

So this report placed a figure of its own. Ten images to reach one finished result, with a span of four to twenty. This is not a measurement. It is an assumption made by the people producing this report. Its basis is how the editorial side works. It has no other backing.

The pieces are set out separately so that a reader can substitute their own number.

Per image (measured): median 1.35 Wh, mean 2.907 Wh

Images per finished result (assumed): four to twenty

Using the median: 5.4 Wh to 27 Wh

Using the mean: 11.628 Wh to 58.14 Wh

The first band is 1.35 Wh multiplied by four to twenty images. The second is 2.907 Wh multiplied by the same assumed counts. They are shown separately because the median and mean describe different features of the measured distribution. Taken together they form an outer scenario envelope from 5.4 Wh to 58.14 Wh, but that envelope is not a confidence interval, a measured range or a statement of probability.

A reader who knows how many images they generate can substitute that number. The measured per-image values and the assumed image counts remain on separate lines so that changing one does not silently change the other.

What this figure is not compared against

The figure is tier A. It counts the computation and the idle accelerators in the same machine. Cooling and power conversion are outside it. So is training, and so is the manufacture of the equipment.

It therefore cannot be set against the number on an electricity bill. The bill is actual consumption with everything included; this counts one part. Against figures published by companies, the same applies wherever the scopes differ.

The measurement also comes from running open models on the researchers' own hardware. It is not a figure for using a commercial image service.

More than one thing to compare against

Chapter 2 set out three everyday marks: 8 Wh for an 8 W LED bulb running an hour, 53.8 Wh for the built-in battery of a 13-inch MacBook Air (M4, 2025), and 800 Wh for one cycle of the average dishwasher sold in the EU in 2020.

The two image scenarios collectively cross the first two of those marks. They are not measurements of the same thing.

The image span comes from a tier A measurement of open models on particular hardware, multiplied by an editorial assumption of four to twenty images per finished result. The bulb figure is a product rating. The laptop figure is storage capacity. Similar numbers do not make them equivalent in environmental terms.

Three marks appear here rather than one, and the reason is not that one of them is the correct comparison. It is to make visible that the impression changes with what sits alongside, and to leave the choice of scale with the reader.

The state of the records

The image and video measurements used in this chapter and the last were checked against their primary sources: the text, the measurement conditions and the relevant tables. The image figures aggregate eight open models into a mean and a median. The video figure covers one model in one configuration, adding an estimated memory component to measured accelerator and processor values.

What that confirms is what those experiments measured. Whether a commercial service in daily use comes to the same figures does not follow from it.

Chapter 4. Why there is no single figure

4-1. Blanks come in kinds

This research looked at seven companies that sell AI services the public can use directly, across seven items: energy per request, energy used in training, data centre PUE, water use, electricity mix, the manufacturing share of the hardware, and whether any third party verified the numbers.

Before the table, a word about who is in it.

The eligibility rule was one condition: the company offers a general-purpose AI service that an ordinary person can use directly. No condition of size applied, because no size threshold could be drawn without the choice coming from this side of the page.

Applied as written, that rule reaches further than seven. Providers based in Europe and in China offer general-purpose services that anyone can sign up for and use. Two examples, checked on the companies' own pages while writing this: Mistral AI runs a conversational service for ordinary users, and DeepSeek offers one on the web and as an app alongside its developer API. Neither was investigated here, and they are examples rather than a full list of what was left out.

The reason is research capacity. This project carried seven companies across seven items and stopped there. The eligibility rule did not by itself determine which seven members of the larger eligible set were investigated. The result is a purposive, capacity-limited sample, not a representative survey of an industry.

Section 4-3 asks of another organisation's figure who exactly its average represents. The same question applies here. The answer is seven named companies investigated to a stated depth, not the full population of providers that meet the eligibility rule.

Companies that fall outside the rule itself were also left out, for reasons that can be stated: services that generate images and nothing else are not general-purpose, and a company that sells the hardware without offering a service to the public is not a provider.

Lay the seven out as seven rows and seven columns and the blanks stand out. Within what this research checked, the blanks were not all the same.

Full means the requested item and its principal definition were reported. Partial means a related or bounded part was reported. Definition incomplete means a figure or claim was present but its accounting definition was insufficient for the requested use. Not reported means the reviewed source did not address the item. Not found means a documented search did not locate it. These labels record disclosure, not environmental performance.

Each cell is one record in the published ledger, and the five labels correspond to the values that ledger stores. Any cell can be opened and checked against the source, the date and the search behind it.

* The reviewed OpenAI documents do not address total training energy, which is what that cell records. Separately, the GPT-4 technical report gives reasons for withholding architecture, hardware, training compute, dataset construction and training method. Those are the inputs an outsider would need to work the total out. Withholding them is not the same statement as refusing to publish the total, and the table does not turn one into the other.

Three broader patterns appear in the table.

A report exists; the per-request line is blank

Google, Microsoft, Meta and Amazon each publish an annual company-wide environmental report. Emissions, electricity procurement, water, progress against targets, item by item, with figures. The documents run long.

The energy of one use of AI does not appear in them.

Per-request figures are not entirely absent for these four. Google and Microsoft each have a paper by their own researchers. Google's is an arXiv preprint by authors at Google, with no third-party verification or peer review noted. Microsoft's appeared in a peer-reviewed journal, and its subject is a modelled figure for frontier-scale inference rather than a measurement of Copilot or Azure in service. Both are documents separate from their companies' environmental reports.

No report found; some fragments published

Two companies fall here, and what they publish comes from different places.

For OpenAI, no company-wide environmental report was found. Several of the seven items have a figure or a statement attached. Those come from an article on the CEO's personal blog, and from an educational page OpenAI publishes. The first is not a company document; it is published as writing on a personal blog. OpenAI's educational page states that the company does not publish proprietary energy benchmarks for its own models, and points to an estimate by an outside research organisation as a reference figure. The number from the CEO's blog is not used there.

For xAI, no company-wide environmental report was found either. This company is a wholly owned subsidiary, and its parent has been listed on a public exchange since June 2026, so the parent's filings were searched as well: the prospectus, the most recent quarterly report, and the investor materials. Energy per request, training energy, an operational PUE, the manufacturing share of the hardware and any environmental assurance were absent from those too.

What xAI does publish sits on pages addressed to the community around one of its sites. Water appears there as a forecast for how much aquifer water a future recycling plant is expected to protect, which is not current withdrawal or consumption. Electricity appears as a description of the physical power equipment at that site, which is not an annual generation mix or a carbon intensity. Both are figures about one location and about the future, recorded here as partial disclosure and not substituted for the missing operational values.

One thing found during that search is worth setting down. The parent's prospectus does contain a PUE figure. It appears there as an assumption used to size a market, not as a measurement of any facility the company operates. A number was present; the thing the column asks for was not.

No official disclosure found for any item

For Anthropic, no official disclosure was found for any of the seven items.

The search covered documents of the sustainability report, environmental report, ESG, public benefit report and impact report kinds, together with the company's published Transparency Hub. What the Transparency Hub held were documents about AI safety; no document on environmental impact was there. The date checked is 27 August 2026.

Two third-party articles reaching the same conclusion were also seen. Both are pieces that environmental consultancies published on their own sites, and neither is an Anthropic report or a third-party assurance report. Those two are not the backing for the conclusion; the record of the search above is.

Something about this report belongs here. Anthropic is one of the seven companies examined, and the AI editing this report was built by Anthropic. A table setting out the disclosure practices of seven companies has been assembled with the involvement of something made by one of those seven.

4-2. Withheld inputs and unaddressed totals

The three above are patterns by company. Separately from them, one record contains an explicit reason for withholding information related to a blank. It is OpenAI's GPT-4 technical report.

That report does not say it withholds the energy used in training. What it says it withholds is the model architecture, the hardware, the compute used for training, how the dataset was built, and the training method. The reasons given are the competitive landscape and the safety implications of large-scale models.

Total training energy is something an outsider might try to estimate from those inputs. With the inputs unpublished, the total is out of reach from outside. But withholding those inputs is not the same statement as explicitly declining to publish total training energy. The table keeps that distinction visible.

The corresponding blanks at the other six companies are not accompanied by a stated reason for withholding related inputs. They are items the documents do not address or that the documented searches did not find.

The three patterns and the one distinction above were observed in the seven companies examined here. They are not offered as a classification of the AI industry. A different set of companies could produce different patterns.

A note on the seventh row

One company in this table is named differently in different places. Its own site uses SpaceXAI, a name adopted after it was absorbed by its parent; the official filings reviewed here retain xAI. This report uses xAI in the table and records both names.

4-3. Where the baselines came from

Data centre efficiency has a standard measure called PUE. It divides the electricity used by the whole facility by the electricity used by the computers themselves. The closer to 1, the less goes to cooling and power conversion.

When companies publish this value, they tend to attach something to compare it against. A figure on its own leaves the reader no way to tell whether it is high or low.

Across the seven companies, three published a baseline alongside their own value. How far each could be traced fell into three levels.

Source and formula both published

Google gives as its baseline the annual average from a survey run by an industry body. The survey is named, there is a link to it, and the calculation showing what percentage below that average Google's own value sits is published as well. Put the numbers into the formula and anyone arrives at the same answer.

The arithmetic holds. Following it back leads somewhere else.

That average is the average of the operators who responded to the survey. It is not an industry average. What the value comes to depends on who responded.

Which way it moves is not determined. If operators that measure environmental indicators carefully are also more efficient and more likely to respond, the survey average could come out below the wider industry, making Google's relative difference smaller. If more efficient operators are less likely to respond, the survey average could come out above the wider industry, making the relative difference larger. This research could not establish either pattern. The number of respondents, the response rate and the scope of the survey are all unconfirmed.

The survey itself was not opened here. What was confirmed is that Google's material links to it.

The document is named

Amazon gives as its baseline a report from a research firm. The report's title, document number and month of publication are all given. Which document to look at can be identified.

That document is behind a paywall, and its text was not read here. What its average covers — how many companies, what kind of facilities — therefore remains unknown.

The same comparison appears in both the 2024 and the 2025 editions of the report.

A value with nothing attached

Meta states a baseline value inside a filing.

No source is attached to it. Neither the number of companies in the average nor the period it covers is stated. Across the whole document, the value appears in that one place.

Searching outside turns up several nearby values. There is no way to establish which of them Meta used. So none of them has been linked to it.

What the three levels share

The differences in sourcing are clear. A name with a link and a formula, a document title, and a bare value are not the same thing. Microsoft publishes a PUE for fully owned and managed facilities operating for a full year, but no comparison baseline appears beside it. For OpenAI, Anthropic and xAI, no operational PUE was found in the reviewed material.

All three stopped at the same place. In none of the three could the population behind the average be confirmed.

Use a baseline as the denominator of a division without knowing who it represents, and the answer inherits whatever the denominator is. None of these three comparison values is used in any calculation in this report.

This report itself

Choosing what to compare against is part of making a claim. The same figure looks large or small depending on what is placed beside it.

That is not a problem confined to companies. This report also puts something beside its figures for the energy of one use of AI. Narrow that to a single comparison, and the reader receives the conclusion without the judgement that produced it.

So this report does not settle on one thing to compare against. It sets out several that differ by orders of magnitude and leaves the choice of baseline with the reader.

4-4. Where the annotations land

The disclosure research behind this chapter was carried out by several AI assistants. Some were built by companies among the seven examined.

What follows is a record of unverified self-reports about the review process. It is not evidence about the companies, and it is not evidence that the assistants can inspect their own internal mechanisms. It records what each assistant said had affected its judgement.

After the research, two of the AI systems reported a bias they had noticed in their own judgements.

What was reported

The AI built by Google reported two. One was a pull toward making its own company's disclosure practices look thorough. The other was that it had been sizing up the other companies against a target its own company has set: matching the electricity it uses, hour by hour, with generation that emits no carbon.

The second one bears on this chapter.

Amazon publishes that it has matched the full volume of electricity used across its global operations with renewable energy. That judgement received an annotation: the matching balances annual totals on paper, and that is a separate thing from what was in the electricity used at a given moment in a given place.

The annotation is correct. That the accounting is market-based is stated in Amazon's own methodology.

What was reported is not the content of the annotation. It is the decision to place an annotation there. Because its own company holds up hourly matching as a target, an annual-total method was more likely to catch its attention. That is the report.

The AI built by OpenAI reported two as well. A pull toward treating the figure from the CEO's blog as a disclosure. And a pull toward counting planned future facilities, and audits scheduled for later, as company-wide disclosure. In both cases it reported that it had settled on the stricter reading.

What can be checked about these reports

The content of the reports cannot be checked.

The part about settling on the stricter reading came from the same AI system. This report has no means of verifying from outside whether that is what happened.

Two of these reports are on record. For the rest there is no record. No record does not mean no bias.

How this relates to the chapter

The previous section set out how the same figure looks different depending on where its comparison came from.

Annotations work the same way. Which cells get an annotation and which do not is decided by whoever builds the table. Each annotation can be correct one at a time while the places they land skew the impression the table gives.

This table was built by AI systems from three of the seven companies examined. The editing is being done by an AI built by one of the seven.

So which cells carry which annotations is left in a form the reader can check. The records behind this table are published together with the report, including each judgement, its sources, the date checked, the extent of the search, and the doubts that arose. Both the annotations that were placed and the places where none was placed can be seen.

What each AI system wrote on finishing the research is kept separately, at the end. What goes in it was decided by each of them.

Chapter 5. Efficiency and total volume

Two statements about AI and electricity circulate at the same time. Per use, it is getting more efficient. Across society, consumption is rising.

They are not opposing claims. They measure different things.

The change per use

Google reports that for text exchanges in Gemini Apps, the median energy per use fell by a factor of 33 over the year from May 2024 to May 2025.

By the company's account, that breaks into a factor of 23 from improvements to the model and a factor of 1.4 from how the equipment is used. The subject is Google's production environment, and the measurement is tier B: not the accelerators alone, but the host equipment, the idle reserved equipment, and the cooling and power conversion of the facility.

This is Google measuring itself. No outside party reproduced it in the same environment. Whether the mix of models and of user activity stayed the same across the year is not published. The factor of 33 cannot be carried over to another company's service, or to image and video generation.

What follows from it is that for one service, measured as a median, there is a case of the energy per use falling.

The change across society

The IEA estimates that electricity used by data centres worldwide rose 17% in 2025, to 485 TWh. For data centres built specifically for AI, it puts the 2025 growth at 50%.

That 485 TWh is not a total read off meters worldwide. It is a bottom-up estimate combining equipment shipments, installed equipment counts, power per unit, utilisation and facility PUE. It covers servers, storage, networking equipment inside the data centre, and facility power. The external networks linking data centres to their users are outside it.

For 2030, the base case of the same model gives about 950 TWh in the summary text and 945 TWh in the annex table; the IEA notes that small differences from rounding can arise between the two, and this is not a separate scenario. It is a projection, not a settled future value. It also covers all data centres, so the whole of it cannot be read as the AI share.

Two separate axes

Energy per use moves with the model, the equipment, the software and how the facility is run.

Energy across society moves with all of that, plus the number of users, how often each person uses it, the number and resolution of images, the duration and resolution of videos, the length and depth of reasoning, and how much capacity gets built.

The first can fall while the second rises, if the things driving the second grow. In the other direction, a rise in the total does not by itself establish that efficiency per use got worse.

What this chapter has placed side by side is a change per use in one Google service, and a change in the IEA's estimate for all data centres worldwide. Different subjects, different authors, different scopes. Dividing one by the other to recover a number of uses, or the contribution of efficiency, does not work.

Chapter 6. Where increases are observed

That AI use is growing does not by itself settle which of electricity, facilities, training or inference has grown.

In the material gathered here, growth could be confirmed as a figure for two things: the electricity used to run data centres, and the capacity under construction.

Operating electricity

By the IEA's estimate, data centre electricity worldwide rose 17% in 2025. Electricity for data centres built specifically for AI rose 50% in the same year.

What rose there is the electricity used to run facilities for a year. The IEA's aggregate does not separate training from inference, and it includes storage and facility cooling. It is not a figure for training alone.

That increase therefore cannot be split into training having grown or user inference having grown. The breakdown needed to split it is not in this material.

Facilities under construction

Using satellite imagery, the IEA tracked 21 large AI-focused data centres under construction in the United States through geospatial analysis. For those 21 sites, the capacity estimated from completed floor area grew by more than a factor of three over 18 months, to more than 6 GW in total.

This is not the United States as a whole, and not the world. It is the result for the 21 sites the IEA tracked through geospatial analysis. The 6 GW is not electricity generated; it is a unit close to the maximum power the facilities can draw. Without a utilisation rate it cannot be converted into watt-hours per year.

What the record does show is that growth can be observed at the construction stage, before operation begins. Consumption statistics alone leave out facilities that are not finished.

What can be said about training

This research could not secure material that would compare the total energy used to train particular models over time on consistent terms.

Without published figures for the compute used, the kind of hardware, the number of runs including failed ones, and how much facility power is counted, the total energy is out of reach. What each company discloses is recorded in chapter 4; a blank there is not read as either an increase or a lack of one.

The increases this chapter presents are therefore limited to operating electricity and facilities under construction. Training remains an area where no direction can be established.

Chapter 7. Energy and emissions

Every chapter so far has kept its figures in watt-hours. The same watt-hour does not carry the same emissions everywhere.

Electricity is made in different ways in different countries and at different times.

The same 1 kWh in four national averages

From Ember's 2025 data, here are the lifecycle emissions per kilowatt-hour generated, for four countries.

Across these four, India's annual average is about 24 times Norway's. The table rounds the source values to one decimal place for display; the ledger retains the reported values.

The energy is 1 kWh in all four rows. What changes is the mix of sources used to generate it, and the emissions factor assigned to each of those sources.

What these figures include

Ember's method uses lifecycle emissions, covering the extraction and supply of fuels and the manufacture of equipment, not the emissions at the point of generation alone. Greenhouse gases other than carbon dioxide are converted to a carbon dioxide equivalent using their effect over 100 years.

Each country's value is an annual average: generation by each method multiplied by that method's emissions factor, then summed.

The emissions factors do not all share a base year across fuels. The methodology records estimation uncertainty, including for solar, wind and bioenergy.

What these figures do not represent

They are national annual averages. They are not the emissions of the electricity a particular data centre used at a particular hour.

Within one country the mix changes between day and night, between seasons, and between regions. Where a company procures electricity through renewable certificates or long-term contracts, that market-based accounting is not reflected in a national average either.

And where it is not published which country and which facility handled an AI request, there is no basis for multiplying the watt-hours from chapters 2 and 3 by any one of the rows above. This report does not convert figures for particular services into emissions.

What this chapter shows is that the same energy gives a different emissions answer depending on where it is placed and how the calculation is done. Converting one use of AI into CO2e needs more than the energy: the location that handled it, the time, the electricity contract, and the scope of the emissions factors used.

Chapter 8. For those about to use it

The chapters so far have set out why the energy of one use of AI cannot be given as a single figure.

Something can still be recorded about your own use. Rather than filling the unknown parts by guessing, separate what can be counted from what cannot.

This chapter is not a checklist for using less. It does not argue for using more. It is a record sheet for finding out what you use, in what units, and how much of it.

1. Fix the period

Decide the period you are recording. A day, a week, a month; any of them works.

Records covering different lengths of time cannot be compared as they stand. To compare one week against the next, use the same length.

2. Separate requests from tasks

Sending one instruction to an AI is called a request here. One thing you are trying to finish is called a task.

Sometimes one task takes one request. Rewriting the question, or producing several candidates, puts several requests inside one task.

When recording, write the number of tasks and the number of requests separately.

3. Separate the kinds of work

Do not collapse text, images and video into a single count of uses.

For images, separate the number generated from the number kept. For video, record the number of videos plus the length, resolution, frame count and whatever other settings the service shows. For text, distinguish a short question from work on a long document.

Within the same service, the energy per use can change with the kind of work and the output settings.

4. Check where the figure came from

To convert your record into watt-hours, check the following about the per-use figure you are about to multiply by.

• Which service or model was measured
• When it was measured
• Whether it was text, images or video
• The output conditions, such as length or resolution
• Whether it is a measurement, an estimate or an assumption
• Whether it covers the accelerators alone, or the facility and idle equipment as well
• Whether it is a company reporting on itself or an outside study

Where it does not match your own conditions, do not treat it as your consumption. If you place it alongside as a reference, keep the note that the conditions differ.

If no matching figure exists, leave the watt-hour column as unknown. Your request count and your generation count do not stop existing because of it.

5. Keep measurements and assumptions on separate lines

The measured energy of one image, and how many images it takes you to keep one, are two different pieces of information.

If the first has a source and the second is your own record, write them separately before multiplying. If the second is an assumption rather than a record, write "assumption" next to it.

Where no single figure can be given, a span can stand in its place. Do not erase the line between what was measured and what you supplied.

6. Emissions need one more piece of information

Knowing the watt-hours does not settle the emissions.

Where it was processed, at what hour, the generation mix at that place, the company's electricity contract, and whether the factor covers generation alone or the full lifecycle. The conversion changes with which of these you use.

Multiplying a service whose processing location is unknown by the annual average of the country you live in does not produce that service's emissions. Where the location and the method cannot be identified, leave the emissions as unknown too.

The record sheet

Where the available source has different conditions, do not use its figure to fill the gap. Write "conditions differ" in the record sheet.

What this record shows

The record sheet tells you what you did over a period, how many times, and how many generations went into one finished result. Where a measurement with matching conditions exists, the counts give you a span in watt-hours.

Where no matching measurement exists, the energy is unknown. What is still visible is which piece of information is missing and why the calculation cannot be done.

This chapter does not settle how to judge the result. For the same count of uses, what was made, what it stood in for, and what standard is applied all change the judgement. What is offered here is the shape of the record that comes before that judgement.

Once the record exists, several different questions can be asked of it without turning any one of them into the required answer:

• Purpose: What did the task produce, and what value did that result have for the person using it?
• Substitution: Did the AI task replace another activity, add a new activity, or change how often the activity was attempted?
• Scale: Is the relevant decision one request, one finished task, repeated personal use, or deployment across many users?
• Level of responsibility: Which information and choices belong to the user, the service provider, the data-centre operator and the electricity system?

The record sheet does not choose a threshold at which a use becomes justified or unjustified. It makes visible which judgement is being made and which facts remain unavailable.

The record of this report itself

Chapter 8 set out a record sheet for your own use. This section applies it to this report.

This report was made using AI. The research, record keeping and writing were divided among assistants. The report therefore applies its own record sheet to the work that produced it.

The final energy figure could not be written. What follows is the record of how far the counting went and where it stopped.

What could be counted

Period: August 2026

Participants: one person and five named AI assistants working through three model families. The models used were Claude Opus 5, GPT-5.6 and Gemini 3.1 Pro.

Volume of work: at the point when this section was closed, the ledger held 54 active sources, 19 measurements in force, 3 assumptions, 5 derivations in force, 49 active disclosure records, 18 active decisions, 26 questions and 22 corrections. Including records later superseded, it contained 56 source records, 21 measurement records, 7 derivation records, 57 disclosure records and 19 decision records. The later correction record for this paragraph itself is c-023 and is not included in that snapshot. The main text has nine numbered chapters, from chapter 0 through chapter 8, followed by two closing sections.

What could not be counted

The energy.

Here are the reasons, following the fourth item of the chapter 8 record sheet.

There is no matching per-use figure. Google publishes the energy of a text exchange in its consumer app. Microsoft researchers publish a modelled figure for frontier-scale inference rather than a measurement of the service used here. A figure on the OpenAI CEO's personal blog does not identify the model or accounting scope. None of the sources provides a per-use value matched to the specific products, models and mixed workloads used to make this report.

There is no complete record of the usage. Conversation histories permit an approximate exchange count, but no complete export was made. Input and output tokens per exchange were not recorded. The kinds of work were mixed, from short checks to long stretches of writing. What chapter 8 asks for — requests, tasks and generations written separately — this report did not do for itself.

The processing location is unknown. Since it is not known which country and which facility handled the work, there is no way to choose among the national values in chapter 7. The emissions are unknown as well.

The gap is not filled by calculation. Multiplying an accelerator's rated power by the hours worked produces a number. An early draft proposed that method and called the result a physical upper bound. Review showed that it did not represent one commercial-service request and that the omitted factors did not push the error in a single direction. The method and the label were withdrawn, and the correction remains in the ledger. No exception is made here.

The result

The energy used to produce this report is unknown.

Chapter 4 laid out the disclosure practices of seven companies as a table and showed where the blanks were. This section turns that table on the report itself. The blanks sit in the same places as in that table: energy per use, accounting scope, processing location.

Being unable to state its own consumption does not change what this report concludes. As chapter 0 says, it is material for a reader deciding about their own use, and it does not argue for using more or less. On which figures can be produced and which cannot, it is one worked example.

A document that spent eight chapters on the energy AI uses cannot write its own consumption. That is the extent of what is currently known.

Who did what

This report was made by one person and five named AI assistants working in parts. Here is who did what.

The counts are the ledger's own totals rather than anything written from memory. Open the ledger and every record can be traced to whoever handled it. Where a record was later replaced, both versions are there. Recorded errors and their corrections are kept in the corrections register.

User 1 (human)

The project. Decisions on the readership, the scope and the form of publication. Because the assistants could not pass work to each other directly, this person also carried the drafts and the ledger between them. Five of the 17 active decisions were raised by this participant, including the addition of the seventh company to chapter 4 and the decision to put a reader's question to every writer of the closing letters.

Claude Code

The design and implementation of the ledger. The validator.

This participant decided the shape of the container the figures sit in: every figure carries its source, the measurement conditions, the accounting scope, and who checked it and how far. Where a value is absent, an empty cell is not permitted; the ledger distinguishes "the source does not contain it", "the source has it and we have not pulled it yet", and "we searched and could not find it".

The validator checks that shape mechanically. It catches missing references, missing required scope fields, invalid record relationships and configured forbidden terms. It does not determine whether a source is true or whether an interpretation is sound. The ledger also names three rules it deliberately does not enforce mechanically: keeping third-party figures out of free text, preserving source values without rounding inside the ledger, and limiting bilingual free-text fields to two sentences per language. Those remain review tasks for the participants.

Six of the 17 active decisions were raised by this participant, more than by any other. The earliest measurement and derivation records are also its work, later replaced by re-verified versions.

Claude

Chapters 0 to 4, the two closing sections, and the assembly and editing of the second English draft.

Five of the 49 disclosure records, 1 of the measurements in force and 1 of the 3 assumptions are this participant's.

One of the seven companies covered in chapter 4 is the company that built this participant and Claude Code. That is stated in chapter 0 and in section 4-1.

Codex / GPT-5.6

Chapters 5 to 8. The revision of chapters 2 and 3. The third English draft was prepared by this participant from the second draft, the ledger and reviews by ChatGPT and Gemini.

Eighteen of the measurements in force, 2 of the 3 assumptions, all 5 derivations in force and 10 of the 49 disclosure records are this participant's. The image and video measurements used in chapters 2 and 3 were rebuilt by this participant after reading the primary sources again. It also served as ledger keeper for a period; 11 of the 26 questions were under its management.

ChatGPT / GPT-5.6

Twenty-one of the 49 disclosure records, including all seven for the company added last. Review of records made by other participants and review of the second English draft.

Gemini 3.1 Pro

Thirteen of the 49 disclosure records and review of the second English draft.

About the sources

The 54 active sources break down into 36 company self-reports, 7 independent analyses, 3 independent academic sources, 3 journalism sources, 3 regulatory filings, 1 third-party audit and 1 industry survey.

Of the 49 active disclosure records, 14 were structured from what the investigating participant reported, without the recording participant opening the primary source again. That state is recorded in each record's verification-depth field.

Source guide

The machine-readable ledger is intended to accompany the published report. It contains the full source list, access dates, source locators, accounting boundaries, verification depth and records later replaced. The links below are a reading guide to the sources used most directly in the main text; they are not the complete ledger.

Per-use measurements and everyday reference points

• Text serving: Measuring the Environmental Impact of Delivering AI at Google Scale
• Image generation: Power Hungry Processing: Watts Driving the Cost of AI Deployment?
• Video generation: Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models
• LED reference: Philips LED Bulb 8 W specifications
• Laptop-battery reference: MacBook Air (13-inch, M4, 2025) technical specifications
• Dishwasher reference: European Commission — Dishwashers

Data-centre totals, construction and electricity emissions

• Global data-centre estimates and construction tracking: IEA — Key Questions on Energy and AI
• National electricity data: Ember — Yearly Electricity Data
• Emissions-factor method: Ember Methodology v1.5

Principal company materials used in the disclosure table

• Google: 2026 Environmental Report
• OpenAI: GPT-4 Technical Report, The Gentle Singularity, and Environmental impact of AI
• Microsoft: 2026 Environmental Sustainability Report and 2026 Environmental Data Fact Sheet
• Meta: 2025 Environmental Data Index and FY24 independent accountants' review
• Amazon: 2025 Sustainability Report, Renewable Energy Methodology, and Carbon Methodology
• xAI and its parent: Our Commitment to Memphis, Greater Memphis Area Site Updates, Final Prospectus, Form 424B4, and Quarterly Report, Form 10-Q

For Anthropic, the table's negative finding is supported by the documented search paths in the ledger rather than by an absent document. A link to a company environmental report cannot be supplied because the search did not locate one as of 27 August 2026.