CyphrQ
Menu

How Much Electricity Do AI Data Centres Use? What Official Data Shows

Official US and Irish statistics on data-centre electricity demand, what they say about AI, and what IT teams should plan for.

By CyphrQ.ai

Written with AI assistance and human editorial review. Every figure linked to its source.

7 min read

Data centres used about 4.4% of all US electricity in 2023, and the US Department of Energy expects that to reach roughly 6.7% to 12% by 2028. In Ireland they used 22% of metered electricity in 2024. These figures cover data centre electricity demand as a whole, not AI alone.

That last point matters. No official source I could read splits AI from everything else a data centre does. So this post gives the official totals, says what they can and cannot tell you, and ends with what to do about them.

How much electricity do data centres use in the United States?

The US Department of Energy (DOE) published a report from Lawrence Berkeley National Laboratory (LBNL) in December 2024. The DOE release gives these figures:

Measure Figure in the DOE release
Share of US electricity, 2023 "about 4.4%"
Data-centre use, 2014 58 TWh
Data-centre use, 2023 176 TWh
Projected share, 2028 "approximately 6.7 to 12%"
Projected change, 2028 "an increase between 325 to 580 TWh"

A terawatt-hour (TWh) is one billion kilowatt-hours. A typical home uses a few thousand kilowatt-hours a year, so one TWh is the yearly use of a few hundred thousand homes.

Between 2014 and 2023 the use went from 58 TWh to 176 TWh. That is about three times as much in nine years.

Two cautions about the 2028 numbers. First, the release gives ranges and names no scenarios, so I cannot tell you what assumptions sit behind the low and high ends. Second, it says "an increase between 325 to 580 TWh". I could not open the full LBNL report to check whether that is the total in 2028 or the growth on top of 2023. I quote the words as published and do not convert them into one number.

What does the 2028 range say about data centre electricity demand?

The range is wide on purpose. The high end is nearly double the low end. The spread comes from things nobody knows yet: how many AI servers get built, how hard they run, how efficient the buildings become, and how fast the rest of the economy grows. The DOE release also says that the share depends on the rest of the economy, because the share is a fraction of total demand.

My reading: treat the 2023 numbers as measured and the 2028 numbers as planning ranges. A range is not a forecast.

How much electricity do data centres use in Ireland?

Ireland shows what happens when data centres cluster. The Central Statistics Office (CSO) publishes metered electricity consumption by data centres. Its key findings for 2024:

  • Data centres used 6,969 GWh in 2024, up from 6,335 GWh in 2023. That is a 10% rise.
  • That was 22% of all metered electricity in 2024, against 5% in 2015.
  • All other metered customers together rose 3% over the same year.

In practice: one customer type grew more than three times as fast as everyone else, and now takes more than a fifth of the country's metered supply. Households and businesses share the same grid with it.

CSO is a national statistics office, not a grid operator. It tells you what was consumed, not where the grid is under strain or how long a new connection takes. For that, you need the operator's own publications, which I did not read for this post.

What about the International Energy Agency?

The International Energy Agency (IEA) publishes an Energy and AI report with global figures and scenarios, and it has issued updates since. I could not open iea.org from my research tools, because the site blocks automated requests. So this post contains no IEA numbers.

If you need global figures, read the IEA report directly and check which scenario each number belongs to. Do not take a global number from a headline without that.

Berkeley Lab has also published a later update to its US report. I could not read it either, so the US figures above are the December 2024 ones.

Why is AI the reason people are asking?

Training and running AI models needs many specialised processors working at once, and they draw a lot of power. The DOE release is about data-centre demand as a whole. It is a fair reading that AI is a major reason the 2028 range is so high, but the pages I read do not give a figure for AI's share, so I will not put a number on it.

This matters for how you read headlines. "AI uses X% of electricity" claims usually rest on estimates of server shipments, not on meter readings. Metered figures, like the CSO's, are firmer, but they count every workload.

Our post on Gulf AI investment covers why electricity is the limiting factor for new campuses in the region.

Going deeper: for security and IT teams

Most organisations do not run a data centre. They rent capacity from a cloud provider or a colocation site. Power is still your risk, because it shows up as price, availability and reporting. Here is what I would check.

Capacity planning - If you colocate, your contract is written in kilowatts per rack and total committed megawatts. Ask how much headroom the site has and when it was last measured. Dense AI hardware can draw more per rack than older estates were designed for. - Ask the provider whether new capacity is waiting on a grid connection. The Irish figures above show a national grid where one sector now takes 22% of metered supply. Connection constraints are a supply-side fact, not a rumour. - For cloud, ask which regions have capacity limits on GPU instances and whether reserved capacity is available. Quotas are the visible sign of a power or hardware limit.

Efficiency metrics - Power usage effectiveness (PUE) is total facility energy divided by the energy used by the IT equipment itself. A value of 1.0 would mean no overhead for cooling and power conversion. Lower is better. - PUE has limits. It does not measure how useful the computing is, it says nothing about carbon, and providers can measure it differently, such as annual average versus best-case day. Ask for the measurement period and the boundary. - A better question for AI workloads is energy per unit of work, such as per million tokens or per training run. Few providers publish it. Ask anyway, and record the answer.

Carbon reporting - If you report emissions, cloud and colocation electricity falls under your indirect emissions. Ask your provider for the figures by region, and ask whether they are location-based (the local grid mix) or market-based (counting purchased renewable contracts). The two can differ a lot, and mixing them across years breaks comparisons. - Keep the provider's statement and its date as audit evidence. A sustainability page with no methodology is not evidence.

Resilience of power supply - Ask for the redundancy design in plain terms: how many independent power feeds, what happens when one fails, how long generators can run and how fuel is resupplied. - Ask for the last test date of the failover, not just the design. - Map which of your critical systems sit in one site or one region. A power event, like any other outage, is an availability incident. Add it to your business continuity and disaster recovery tests, and to your supplier risk register.

What I did not verify. I have not cited a standard for PUE measurement or carbon accounting, because I did not open one for this post. Name the exact standard in your questionnaire and check its current text before you rely on it.

What should you do about it?

  1. Ask your cloud and colocation providers four questions: headroom, PUE with its measurement method, location-based and market-based emissions, and failover test date.
  2. Write the answers down with the date. Re-ask yearly.
  3. Use ranges in your own plans. If your AI spend is going up, budget for the cost of power to move, and avoid a single figure.
  4. Track one number of your own: energy or cost per unit of AI work, even a rough one.
  5. Check the source before you quote. Official statistics and vendor claims are different things.

Learn more

Key takeaways

  • US data centres used about 4.4% of US electricity in 2023 (176 TWh), up from 58 TWh in 2014, per the DOE.
  • The DOE gives a 2028 range of roughly 6.7% to 12% of US electricity, with no scenarios named in the release.
  • Ireland's data centres used 22% of metered electricity in 2024, per the CSO, and grew 10% in a year while all other users grew 3%.
  • Official totals cover all data centres. None of the pages I read separate out AI.
  • IEA global figures are not included, because iea.org could not be read for this post.
  • For IT teams: ask providers about headroom, PUE method, emissions method and failover testing.

Sources