Section: Technology & AI Author: V Reading Length: ~27 min Sources and further reading: 22 items Topics: AI, electricity, data centers, inference, training, PUE, water, ERÚ, IEA SEO / Working Title: How much electricity does artificial intelligence use?
How much electricity does artificial intelligence use? The question sounds simple until we notice that we are not comparing one account. The data center is not artificial intelligence. Training is not inference. A prompt watt-hour is not a grid terawatt-hour. And the Czech net 58 TWh is not a ChatGPT account in Brno. The article is therefore not looking for one percent to grab the headline. It looks for the denominator: which year, which boundary, measurement or scenario, and what doesn't yet belong.
1. Four numbers on one table are not one share of "AI"
In the editorial office, there are four issues on one table. The reporter wants to turn the global share of data centers into a sentence about artificial intelligence. Analyst stops him: IEA measures data centers in 2024, not pure AI share. The activist quotes a median of 0.24 Wh per Gemini Apps text prompt and multiplies it by an unknown number of queries. Energetik will show a Czech net of 58 TWh for 2024 and a global increase in data center consumption of approximately 70 TWh between 2024 and 2025. Managercloud puts Microsoft's 37,026,353 MWh of FY2025 electricity on the table and asks how much of that is AI [3][9][14][16].
At first glance, it is a dispute whether artificial intelligence is energetically small or large. In fact, each holds a different denominator. Reporter global account of all data centers. Analyst definition of IEA. Activist unit of inference for one service and one methodology. Energetician of the annual balance of the state. Business account manager including cloud for customers who don't need to use generative AI.
Therefore, they can all be partly right and the common headline still wrong. A number without year, border and scope is not information. It is the raw material for misunderstanding.
A data center is not the same as "AI" and a prompt watt-hour is not the same as a network terawatt-hour.
— Jiný Kontext
2. AI electricity has at least six layers
The first layer is chip and computation. It depends on the accelerator, usage, batching, context length, number of output tokens, and whether the system is running in easy or more demanding mode. The IEA states that a high-end GPU can draw up to 1000W and that the cluster used for the GPT-4 was estimated to be around 10MW for the GPU alone, or around 22MW with the rest of IT and cooling [2].
The second layer is training. It is disposable, high-end and concentrated. The third is inference, i.e. serving: repeated operation after each query, summary, translation, image generation or agent run. The fourth layer is the data center infrastructure. PUE tells how much electricity above the IT load is consumed by cooling and other operations. The fifth is water, where it differs on-site from water used in the energy sector. The sixth is the electricity system, including local networks and annualbalance.
Thus, the sentence "AI consumed electricity" is the beginning of the diagnosis, not the end of it. The failure may be in the model, in the calculation, in the cooling, in the network, in the emission factor, in the water scope, or in the fact that the 2030 scenario reads like a 2024 account.
This layering also changes the political question. In another way, the connection of a large campus is planned, in another way the carbon of one service is evaluated, and in another way it is decided whether the company should use a demanding agent workflow for a task that can be handled by a smaller model. The network doesn't see "AI" as an abstract capability. Sees power, consumption, time, location and reliability requirements. The user does not see the network again. He sees the answer on the screen. Between these two views lies the data center, its cooling,electricity contracts and operating rules.
That's why it's so easy to talk outside yourself. One person deals with the global climate, the second the local transformer, the third the API bill and the fourth the water at the cooling point. Everyone can use the word consumption. However, they are not talking about the same border. A good sentence about electricity AI must tell which layer it stands on. If it doesn't, it starts to look like a general explanation while only describing one link in the chain.
3. The best number is the one whose denominator you carry
The energy debate does not need the strongest sentence. It needs a number that knows what it measures. A cheap mistake is to round up the share of data centers so that the local pressure on the network disappears. It's an expensive mistake to issue a projection of 2030 accounting for 2024. It's even more expensive to multiply 0.24 Wh by an unknown number of tasks and pretend that the textual median of one service explains the global infrastructure.
| A kind of failure | What does he look like? | How to test | What will limit the damage |
|---|---|---|---|
| Data centers issued for AI | An AI share will be created from the account of all DCs | Verify if the source is talking about DC, AI-focused DC, or inference | Separate the computation boundary and the infrastructure boundary [1][3] |
| Script as an account | The year 2030 appears without a word of projection | Ask about the year and methodology | List the 2024/2025 measurements separately from the 2030 scenario |
| Wh times an unknown number | The prompt number is used to conclude about the network | Search count, job mix and percentile | Do not add the Gemini median to the global DC account [9] |
| Pouring water | The GPT-3's 500ml is up against the Gemini's 0.26ml | Compare model, year, on-site and total scope | Two accounts side by side, not one verdict [12][13] |
| The Czech Republic as an AI payer | 58 TWh reads like a single application account | Separate annual net country and world DC accretion | Use order comparisons with a caveat [16] |
A number is good if we can insert it into a sentence without hidden confusion. Year. Population. Unit. Limit. Measurement or scenario. What's in it and what's not in it.
The same rule applies to comparisons to help the reader. Households, states, or firms are useful orientation aids as long as they remain orientational. When the IEA compares a large data center to households, it shows the scale of power consumption and consumption. It does not say that the household has the same daily profile, the same location, the same tariff or the same importance to the network [2]. When we put the world increase in DC and the Czech net consumption side by side, we see an order of growth. We're not saying it's the same territory or the same decision-making.
So the best number is not necessarily the biggest or the smallest. It's a number that can be checked. If the resource says 415 TWh, year 2024 and all data centers, it can be worked with. If the resource states 0.24 Wh, the median of the text prompt and Google's methodology, it can also be worked with. An error occurs when one number begins to do the work of another.
4. The data center is not artificial intelligence
IEA in the report Energy and AI reports that data centers will consume approximately 415 TWh of electricity in 2024, about 1.5 percent of global electricity consumption. Since 2017, according to the report, their consumption has grown by approximately 12 percent per year. Of this 415 TWh, according to the IEA, 45 percent came from the US, 25 percent from China and 15 percent from Europe [1][2].
That's a solid point of reference. However, it is not a pure AI account. Data centers serve streaming, enterprise cloud, e-commerce, storage, databases, traditional servers and many other services. AI-focused centers are a subset that grows faster because they contain a large number of accelerators for training and running models. If 1.5 percent becomes the share of artificial intelligence in the world's electricity, the object of measurement changes.
This does not mean that data centers are negligible. Local pressure on the network can be high even when the global share seems small. The IEA repeatedly draws attention to the concentration of demand and connection restrictions. It just needs to be said: data centers in 2024, not artificial intelligence as a whole.
Additionally, a local issue can be the opposite of a global headline. A world share of around 1.5 percent in 2024 looks small in one sentence. In a particular region, several projects can change network planning, the need for backup resources, and the pace of construction. Energy is not decided according to the world average. It is decided according to the node, line, capacity reservation and the time when the consumption occurs.
Therefore, it is more accurate to say that AI is one of the drivers of the new demand for data centers. It is not accurate to say that every data center is AI. And it's not accurate to settle for a global percentage if the local connection hits capacity. One sentence must carry both sides: the global account of all DCs and local electrical restrictions are not the same.
5. 485 TWh and +17 percent is the data center account, not the ChatGPT account
In update Key Questions on Energy and AI The IEA reports that data center electricity consumption reached around 485 TWh in 2025, or almost 500 TWh, and was slightly above 1.5 percent of global electricity consumption. Year-on-year, according to the IEA, it grew by 17 percent. The difference against 2024 gives an increase of approximately 70 TWh [3][4][5].
That's a big move in energy. However, it is still all data centers. When the Czech net consumption of 58 TWh in 2024 is placed next to it, we get an order of magnitude comparison: the world's annual increase in DC is greater than the Czech Republic's annual net consumption. It is not a claim that the Czech Republic is paying for AI, nor that every new TWh belongs to chatbots [16].
The IEA projects around 950 TWh for 2030, around 3 percent of world electricity demand. Older report Energy and AI worked in the Base Case with approximately 945 TWh. Both values belong to the future scenario, not to the accounting of 2024 or 2025 [1][3].
6. AI-focused centers are growing faster than all data centers
In a press release to 2025, the IEA states that the consumption of AI-focused data centers has grown by approximately 50 percent year-on-year. That's faster than total data center consumption, which it says is 17 percent, and faster than global electricity consumption, which the press release says is about 3 percent [5].
This is the reason to take AI seriously. Not because every data center account should be overwritten by AI, but because a subset of it is growing rapidly. In the Base Case for 2030, the IEA states that AI-focused centers can exceed three times the baseline level and reach around 465 TWh. The definition applies to centers with a large number of accelerators for training and running models [4].
A caveat applies here as well. 465 TWh in 2030 is not the measured share of artificial intelligence in all the world's electricity. It is a scenario for a defined category of data centers. It is suitable for the question of how to prepare the network. It is not suitable for a sentence without a year and without a definition.
7. Training is not inference
Training is model making. Inference is its use. The training can be energetically concentrated and media visible, because it will take place as a large campaign on many accelerators. The inference is less obvious because it spreads over billions of individual requests whose cost depends on the length of input, output, model, dosage, and mode.
In their work on AI energy, Luccioni, Jernite and Strubell report historical corporate estimates: as cited by Patterson et al. at Google in 2022, approximately 60 percent of ML power was in inference and 40 percent in training; according to a cited AWS material from 2019, 80 to 90 percent of the ML cloud was for inference [7][8][21].
These numbers must not be frozen as today's world ratio. The mix changes with reasoning modes, agents, video, longer context and new models. But they demonstrate an important principle: great training is not the whole bill. The day-to-day operation of the models can be decisive in the sum, even if one answer seems small.
Another difference is hidden in inference: the number of steps. A simple query can mean one model call. Agent work can mean a plan, research, document reading, several drafts, review, correction, and another pass. The user still sees one task. The data center sees more computation. Therefore, the average cost of a prompt cannot be directly transferred to jobs that contain many prompts inside.
This is where energy meets product design. When an application sets an expensive mode for each query, it can be wasteful. By turning it on only where the simple model is not enough, it can save without losing quality. So the difference between training and inference is not just a technical definition. It's a rule of thumb where to look for savings: once when building a model, second time every time you use it.
8. 42.4 GWh of one GPT-4 training is not a global share
The IEA, referring to Epoch AI, estimates the energy per GPT-4 training at approximately 42.4 GWh. The background talks about training lasting about 14 weeks, 25,000 GPUs and a load factor of 84 percent. Daily consumption of around 0.43 GWh is compared in the report to around 28,500 households in advanced economies on a given day [2][20].
It's a good number to get an idea of the scale. It is not a global share. It measures a single training session, not subsequent inference, not an entire data center, not all experiment iterations, and not every service in the market. The IEA also uses order-of-magnitude analogies: one large data center can have consumption comparable to around 100,000 homes, and the largest under construction up to around 2 million homes. Such analogies help size. They are not the accounting identity of the Czechshouseholds [2].
The training numbers therefore belong to sentences with an exact boundary. One training session. Estimate. Specific model. Specific assumptions. Once we make them into a sentence about the whole artificial intelligence, we change the tip to the year and the campaign to traffic.
9. 0.24 Wh prompt is not 415 TWh grid
In the study by Elsworth and co-authors, Google reports a value of 0.24 Wh, 0.03 g CO2e and 0.26 ml water for the median text-generation prompt in Gemini Apps. The methodology includes accelerator, guest CPU and DRAM, idle capacity and fleet PUE. Google reports a fleet-wide PUE of 1.09 and a power drop to a median prompt of 33 times in the 12 months to May 2025 [9][10][11].
This is an important correction of old impressions that every text response is necessarily of a large order. However, it is just as dangerous to use it in reverse as a universal excuse. It is the median of the text prompts of one service, one methodology and one fleet. It does not measure video, agents, long reasoning, other providers, other PUEs or the tail of very demanding tasks.
How to test the prompt against the network
Does the resource have the number of tasks and their mix? Is it the median, mean, or tail of the distribution? Is it text, image, video, code agent or long search? Does it belong in the PUE number? Is it one fleet with a PUE of 1.09 or a different data center? Without count and mix, Wh and TWh do not add up.
10. Even if the normal searches were simple AI text, it's still less than 4 TWh
IEA v Key Questions on Energy and AI hypothesizes: even if all common searches were a simple AI text query, consumption would be less than 4 TWh per year, less than 1 percent of current data center consumption [3][4].
This sentence is useful precisely because it is narrow. It talks about simple text queries. They're not talking about agents who make dozens of calls. It doesn't talk about video generation. It doesn't talk about long context, private document search, or training a new model. And most importantly, they are not talking about the entire data center bill.
It does not follow that the inference is always small. It follows that the common text prompt and the global data infrastructure are different entities. If someone has the number of tasks, their mix and measurements per task, they can estimate. If he only has the median of one service and a sense of the volume of the Internet, he does not have an account. It has a multiplier without one factor.
The practical impact is simple. A company deploying AI in document search, customer support, or code agent doesn't have to count on one universal prompt. It should measure its own typical task: how many calls are made, how long the input is, how much output is discarded, how often the work is repeated and when the more expensive mode is turned on. Only then can we talk about the consumption of the received result.
11. PUE is not model consumption
PUE, power usage effectiveness, is the ratio of the total electricity of the data center to the electricity of the IT load. When Google reports a fleet PUE of 1.09, this means that for every 1 kWh of IT load, there is an additional 0.09 kWh of infrastructure. This infrastructure includes cooling, power and other operational overhead [9][10].
So PUE is not a property of the model. It's not "transformer gluttony". It is a property of the data center and its operation. Different campus, different climate, different cooling and different capacity utilization may give a different result. In their 2019 work, Strubell, Ganesh and McCallum used a global average PUE of 1.58 from 2018. This is part of their estimate and cannot be transferred to Google 2025 or Czech traffic without modification [6].
At the same time, PUE shows why training, inference and cooling should not be mixed. If we are asking about the calculation of the model, we are interested in the IT load. If we ask about the network, we are interested in the overall data center. If we ask about carbon or water, the energy mix, place and time are added.
PUE is also not an excuse. A low PUE says the data center has little overhead over IT load. It doesn't say that the IT burden itself is small. A very efficient center with a huge number of accelerators can use a lot of electricity. Conversely, worse PUE in less traffic may be less significant locally. Therefore, PUE should not be used as a standalone virtue label. Belongs to absolute consumption, capacity utilization and source of electricity.
A good question is: how much electricity goes into IT, how much does the infrastructure add, and how often is the compute actually used? If accelerators are waiting for a peak, idle capacity may appear in the methodology as part of the serving. If running at high utilization, the cost of latency and cooling changes. PUE helps read traffic. It does not close the energy bill.
12. Electricity is not carbon and carbon is not water
One TWh of electricity does not have the same carbon everywhere. It depends on the production mix, collection time and location. The ERO for the Czech Republic reports a net electricity production of 69 TWh in 2024, 3.9 percent less year-on-year. The export minus import balance fell from 9.2 to 6.7 TWh. The installed capacity of photovoltaic power plants increased from 3.3 to 4 GWe, and production from PV plants increased by 24.2 percent year-on-year. Both core and steam decreased year-on-year [16][17].
These numbers do not tell the carbon footprint of one Czech prompt. They give context to the denominator. The same consumption in the network with a different mix has a different emission impact. The same calculation at a different time may have a different marginal resource. And the same data traffic can have a different water footprint depending on the cooling and water used in electricity generation.
Therefore, it is wrong to convert electricity, carbon and water into one sentence without scope. Electricity is a physical consumption account. Carbon depends on production. Water depends on cooling, location and methodology. When these layers are distinguished, the debate is less spectacular but more accurate.
In the Czech context, this reservation is particularly important. ERO gives annual data on the system, production, consumption and balance. It does not say what carbon footprint a specific AI service that a Czech user calls up via a foreign cloud has. Such an answer would need to know where the calculation took place, at what time, with what electricity source and what contractual tools the provider uses. The annual Czech mix is the context, not the GPS of one prompt.
Likewise, water is not a universal transmitter of guilt. Cooling in a dry region has a different meaning than sampling in a place with a different water balance. Water for electricity production is not the same as water evaporated on site. Both may be relevant. However, they may not be added without marking and compared with a number that measures only one of them.
13. On-site water is not a total water footprint and Li is not a rebuttal to Google
Li, Yang, Islam, and Ren estimate that GPT-3 training in the US at Microsoft data centers may have used 5.4 million liters of water in total, including 0.7 million liters on-site. For the GPT-3 inference, they report approximately 500 mL of water for 10 to 50 medium-length responses based on location and time. This is on-site plus off-site water, i.e. water associated with electricity production [12][13].
In contrast, Google reports 0.26 ml of water for the median Gemini Apps text prompt according to its methodology and its operation in 2025 [9][11]. These two figures are not a direct dispute about one thing. Li talks about GPT-3, another year, another scope and the inclusion of water in energy. Google talks about a median prompt in Gemini Apps, a different infrastructure and a different border.
It does not follow that one number is true and the other a lie. It follows that water without a scope is even more dangerous than electricity without a year. The 500ml range should not be changed to a constant for each current query. And 0.26 ml should not be changed to the ceiling for all tasks and providers.
14. 626,155 pounds of CO2e from 2019 is not the ChatGPT of 2026
Strubell, Ganesh, and McCallum estimated in 2019 that training a large Transformer with neural architecture search could achieve 626,155 pounds of CO2e. They used, among other things, a US grid factor of 0.954 lb/kWh and a PUE of 1.58. It was one research pipeline connected to the search for architecture according to the work of So and co-authors. BERT trained on the GPU without such a search was an order of magnitude lower in their table [6].
This number played an important role. It reminded that the research search for models has a physical account. However, it is not correct to make it the carbon footprint of ChatGPT in 2026. The hardware, data centers, PUE, energy mix, models, serving method and scope of use have changed. Moreover, it was training and architecture retrieval, not ordinary inference.
The old number therefore belongs to the history of measurement and to the argument that the energy of AI should be visible. It does not belong in the current global share headline. When line illustration becomes accounting, it ceases to help.
15. Microsoft's 37 TWh is the company's account, not the share of artificial intelligence in the world
Microsoft's Environmental Data Fact Sheet lists 37,026,353 MWh of electricity for FY2025, or about 37.0 TWh. It's a company-wide account. It includes global operations and cloud services for customers who can use AI, but also many services not directly related to generative AI [14].
This number is useful for the question of how large energy firms have become from cloud providers. It is not useful as a direct contribution of artificial intelligence to the world's electricity. If one adds it to the 415 TWh of data centers, one is easily counting part of the same world twice, or mixing up the corporate account with the infrastructure category.
The correct sentence is narrower: Microsoft as a company reported tens of TWh of electricity consumption in the given financial year. Of that, we don't know how much is AI, how much is other cloud services, and how much is the rest of the traffic. Without internal allocation and methodology, the number remains a corporate account, not a model account.
Business accounts are still valuable. They show that digital infrastructure is no longer a marginal item in energy. As a large provider grows, its decisions about location, power purchase, back-up and cooling affect public grids and investments. You just need to keep the two sentences separate. First: cloud businesses have large and growing energy bills. Second: the share of generative AI cannot be read from one company account without allocation.
This distinction protects both sides of the dispute. Criticism does not have to be exaggerated to be serious. The defense cannot hide behind the fact that the exact AI share is not public. Missing allocation is not null proof. It is a reason to formulate the conclusion more carefully.
16. Czech 58 TWh is the country's benchmark. There is no ChatGPT account in Brno
ERÚ states in a report dated February 18, 2025 that the net consumption of electricity in the Czech Republic in 2024 amounted to 58 TWh, 0.6 percent less year-on-year. He also mentions the outage of local consumption by large enterprises. In the same context, ERO reports net production of 69 TWh, 3.9 percent less year-on-year [16][17][18].
The comparison with the IEA is strong if it remains accurate. The global increase in data center consumption between 2024 and 2025, approximately 70 TWh, is greater than the annual Czech net consumption. This shows the order of magnitude of growth of global infrastructure. It does not say that the Czech Republic consumed less than AI, nor that the Czech data center runs at such a volume. The file does not have ERÚ figures on the consumption of specific Czech data centers in TWh.
Before the comparison with the Czech Republic
Are you comparing a country's annual net consumption to the annual growth of the world's data centers, or to the inference of a single application? Is the sentence production 69 TWh or consumption 58 TWh? Is it a Czech mix, global growth, or a corporate account? Without these three answers, the Czech number serves more as rhetoric than as a measurement.
17. The question is not how much "AI" will consume. It says which account, which year and which border
An AI's energy bill cannot be canceled by one prompt coming out small. You can't even be puffed up by renaming all data centers AI. We need fewer stickers and more denominators.
When someone says a percentage, they have to say the year. The year 2024 with 415 TWh of data centers is not the year 2030 with a scenario of around 950 TWh. Must say border. All data centers are not AI-focused centers, AI-focused centers are not one prompt, and one prompt is not a video agent. It must say scope. Is it IT load, PUE, carbon, on-site water, total water, or corporate electricity purchase? And they must say whether it is a measurement, an estimate, a company methodology or a scenario.
Three tests of the denominator
Year: 2024 measurement, 2025 update, or 2030 scenario? Boundaries: all data centers, AI-focused centers, training, inference, or one service? Scope: electricity, cooling, carbon, on-site water, total water, or company account?
Without a year, definition and boundaries, training, inference and cooling, the share must not become a headline.
— Jiný Kontext
So the right question isn't how much AI will consume. It reads: which account are you currently reading, data centers, AI-focused centers, one training, median prompt, water or company, in which year, and what else is not the share of artificial intelligence on the world wide web?
Related texts in this series
- ChatGPT vs. Gemini vs. Claude vs. Grok… — the ranking does not measure watt-hours.
- IEA Energy and AI: 415 TWh of data centers in 2024 is an infrastructure bill, not a single chat bill.