AI Data Centers and Power: Why Electricity Is the Limit
- 7 views
- 0 likes
- 0 comments
- All languages combined

Sign in and earn +30 points by reading this
Read to the end, wait a few seconds, then tap the claim button below.
In short
Data centres used about 485 TWh of electricity in 2025 and the IEA projects around 950 TWh by 2030, roughly 3% of global demand [1]. The bottleneck is local: grid connections, turbines and memory are short, and the PJM grid's 2028/2029 auction fell 6,831 MW short of its requirement [2][5].
Key takeaways
- Data centres used about 485 TWh in 2025; the IEA projects about 950 TWh in 2030, around 3% of global electricity [1].
- AI-focused data-centre demand is projected to triple by 2030, even though energy per AI task keeps falling sharply [1].
- The bottleneck is local: grid connections, permits, gas turbines, transformers and high-bandwidth memory [1][2].
- PJM's 2028/2029 capacity auction cleared at its $325 per MW-day cap and fell 6,831.3 MW short [5]; its market monitor blames forecast data-centre load [6].
- In Ireland, data centres used 23% of metered electricity in 2025 [7].
- Google's $17 billion Wyoming investment includes a 702 MW fuel-cell and carbon-capture power plan and a pledge to pay for all the power it uses [8].
On this page 12 sections
- How much electricity do data centres use?
- Why is AI changing the picture?
- Where exactly is the bottleneck?
- What happens when a grid runs short? The PJM example
- Ireland: what a data-centre-heavy grid looks like
- What did Google announce in Wyoming?
- Will AI data centres raise my electricity bill?
- What should you watch next?
- Common mistakes
- Frequently asked questions
- Glossary
- Sources
How much electricity do data centres use?
Data centres used about 485 terawatt-hours (TWh) of electricity worldwide in 2025, according to the International Energy Agency (IEA), and its central projection roughly doubles that to 950 TWh by 2030, around 3% of global electricity demand [1]. For scale, the IEA's 2025 report had put 2024 use at about 415 TWh, around 1.5% of global electricity [3]. The 2030 projection is almost unchanged: the 2025 report gave about 945 TWh [3].
Demand is not spread evenly. In the IEA's 2025 analysis the United States and China together accounted for nearly 80% of global data-centre growth to 2030, with US data-centre consumption projected to rise by about 240 TWh, or 130%, compared with 2024 [3]. A study by Lawrence Berkeley National Laboratory, funded by the US Department of Energy, found US data centres used about 4.4% of the country's electricity in 2023 and could use 6.7% to 12% by 2028, depending on how fast AI servers are deployed [4].
So, globally, data centres remain a small slice of electricity use. Locally, in a few grids and towns, they are already one of the largest single loads. That local concentration is where the bottleneck appears.
Why is AI changing the picture?
Two trends pull in opposite directions. First, each AI task is getting cheaper in energy. The IEA says energy use per AI task has fallen by at least an order of magnitude a year in recent years, and that a simple text query now typically uses less electricity than running a television for the same length of time [1]. If every conventional web search were a simple AI text query, the IEA estimates the total would be under 4 TWh a year [1].
Second, use and spending are growing far faster than efficiency. The capital spending of five large technology companies passed $400 billion in 2025 and is set to rise by a further 75% in 2026, according to the IEA [2]. Electricity demand from data centres rose 17% in 2025, and demand from AI-focused data centres rose faster still, against 3% growth in global electricity demand [2]. Energy-intensive uses such as AI agents are rising [2].
The result: AI-focused data-centre consumption is projected to triple between 2025 and 2030, much faster than data centres overall [1]. Efficiency gains are real, but they are not keeping total demand flat.
Where exactly is the bottleneck?
The constraint is less about the world running out of electricity and more about getting enough power to a specific site, quickly and reliably. The IEA lists several pinch points [1][2]:
- Grid connections and permits. The pipeline of projects is straining planning and regulatory systems and holding up grid connections [2].
- Equipment. Supply chains for gas turbines and transformers tightened over the past year [2].
- Chips and memory. A shortage of high-bandwidth memory, used in AI chips, is expected to last until at least the end of 2027 [1]. Related reading: the guides RAM Shortage and Graphics Card Prices in 2026.
- Load behaviour. AI data centres can see server load swing by more than 50% of rated capacity within a second, which is hard for a power plant or grid to follow [1].
Is building power on site the answer?
Some US developers are building gas generation next to data centres. IEA satellite tracking suggests only around one-fifth of those projects have started land clearing or construction [1]. To serve critical, swinging loads reliably, on-site gas plants need to be overbuilt by 30% to 70% relative to demand, and with turbines in short supply the IEA says it is not clear this is faster at scale [1]. It expects around 15 to 27 GW of on-site gas to power data centres by 2030, mostly in the US, and around 20 to 25 GW of battery storage installed in data centres globally [1].
What happens when a grid runs short? The PJM example
PJM runs the high-voltage grid and wholesale power market across a large part of the eastern United States, including Virginia. Each year it holds a capacity auction, paying power plants to be available three years ahead.
In its auction for the 2028/2029 delivery year, published on 14 July 2026, PJM procured less capacity than its reliability requirement: total committed capacity was 6,831.3 MW short [5]. Every megawatt cleared at the price cap of $325.00 per MW-day, which at the cleared volume comes to about $16.4 billion [5]. PJM said its reliability requirement rose from 152,400 MW to 156,013 MW, largely because of forecast growth from large loads [5].
How much of that is data centres?
PJM's Independent Market Monitor, Monitoring Analytics, went further in an analysis dated 30 October 2026. It counted 21,470 MW of existing and forecast data-centre load in the auction's demand forecast, of which 16,436 MW was forecast growth, and concluded there would have been no shortfall but for the forecast data-centre load [6]. It estimated that data-centre load raised the auction's cost by about $6.27 billion for 2028/2029, and by about $29.4 billion across the four auctions it analysed [6]. These are the monitor's estimates; PJM's own report describes large-load growth without putting a single price tag on it [5].
Ireland: what a data-centre-heavy grid looks like
Ireland shows what happens when data centres cluster in a small country. Its Central Statistics Office reported that data centres used 23% of all metered electricity in 2025, up from 5% in 2015 [7]. Their use rose 10% in a year, from 6,973 GWh in 2024 to 7,663 GWh in 2025, while consumption by all other users rose 2% [7]. For comparison, all homes in Ireland together used 28% [7].
In practice, that means data centres in Ireland now use roughly four-fifths as much metered electricity as all homes combined (23% against 28%). Questions about new grid connections, backup generation and who pays for network upgrades are therefore national policy issues there, not technical footnotes.
What did Google announce in Wyoming?
On 5 November 2026 Google announced a $17 billion investment in Wyoming that includes its first data centre in the state, in Laramie County; it held a community meeting in downtown Cheyenne [8]. Google does not itemise what the $17 billion covers [8]. The announcement is a good illustration of how power now shapes these projects. Google said it would pay for 100% of the power it uses, and described a 702 MW fuel-cell development with carbon capture and storage, contracted with the utility AEP and using Bloom Energy fuel cells running on natural gas [8]. Captured carbon is to be stored at Tallgrass Energy's sequestration hub in eastern Wyoming [8]. Google said the site would be air-cooled [8].
What is announced and what is reported should be kept apart. Google's post does not give the data centre's total capacity in gigawatts or its site area; figures circulating in trade press for those are not confirmed on Google's own page [8]. Fuel cells on natural gas with carbon capture are also a plan, not yet a working plant, and their real emissions will depend on how much carbon is actually captured.
Will AI data centres raise my electricity bill?
Not necessarily, but in some places they can. The IEA's conclusion is that higher electricity demand does not automatically mean higher prices if the right policies and investment are in place [2]. It also warns that data centres pose special affordability problems because their loads are large, concentrated and grow quickly, often forcing new generation and grid spending [2].
Where costs land depends on the rules. The PJM market monitor measures the effect as an increase in costs to PJM customers [6]. Commitments like Google's to pay for all of the power it uses in Wyoming are one response [8].
The IEA proposes three policy principles, including proactive management of data-centre project pipelines and of electricity investment [1]. Tools it names include smarter integration into grids and incentives for data centres to run more flexibly [2].
What should you watch next?
- The next PJM auction. The market monitor says the 2029/2030 Base Residual Auction is scheduled for 9 to 15 December 2026 [6]. Its results were not published as of 13 November 2026.
- Clean power deals. Technology companies signed around 40% of all corporate renewable power purchase agreements in 2025, and conditional offtake agreements between data-centre operators and small modular reactor projects grew from 25 GW at the end of 2024 to 45 GW by April 2026 [2]. Conditional offtake agreements are not built reactors.
- Memory supply. The high-bandwidth memory shortage the IEA expects to last through at least 2027 can slow data-centre build-outs as much as power does [1].
- Emissions. The IEA projects data-centre emissions to double to around 350 million tonnes by 2035, still about 2% of global power-sector emissions [1].
The main point is that AI's electricity problem is mostly a local one: specific grids, substations and towns, rather than the global total.
| Measure | Figure | Source |
|---|---|---|
| Global data-centre use, 2025 | About 485 TWh | IEA, April 2026 [1] |
| Global data-centre use, 2030 (central projection) | About 950 TWh, around 3% of global demand | IEA, April 2026 [1] |
| Global data-centre use, 2024 | About 415 TWh, around 1.5% | IEA, April 2025 [3] |
| Growth in data-centre demand, 2025 | 17% (global electricity demand: 3%) | IEA, April 2026 [2] |
| US share of electricity, 2023 | 4.4% | Berkeley Lab [4] |
| US share of electricity, 2028 (range) | 6.7% to 12% | Berkeley Lab [4] |
| Ireland share of metered electricity, 2025 | 23% | CSO Ireland [7] |
| Item | Figure | Who says so |
|---|---|---|
| Shortfall versus reliability requirement | 6,831.3 MW | PJM [5] |
| Clearing price | $325.00 per MW-day (the cap) | PJM [5] |
| Total auction cost | About $16.4 billion | PJM [5] |
| Forecast data-centre load in the auction | 21,470 MW, of which 16,436 MW new | Market monitor [6] |
| Cost increase attributed to data-centre load, 2028/2029 | About $6.27 billion | Market monitor estimate [6] |
Common mistakes
Saying AI already uses a large share of the world's electricity
Fix: Data centres used about 485 TWh in 2025 and are projected at around 3% of global demand by 2030 [1]. The pressure is local, not global.
Assuming each AI query is wasteful
Fix: The IEA says a simple text query typically uses less electricity than running a TV for the same time; growth comes from volume and heavier uses such as agents [1][2].
Treating the market monitor's cost figures as PJM's own
Fix: The $6.27 billion data-centre estimate comes from the Independent Market Monitor [6]; PJM's report gives total results and cites large-load growth [5].
Repeating trade-press capacity figures for Google's Wyoming data centre as confirmed
Fix: Google's own post gives a $17 billion Wyoming investment and a 702 MW fuel-cell plan but no total data-centre capacity [8].
Calling a conditional reactor deal a new power source
Fix: The 45 GW of SMR deals are conditional offtake agreements, not built reactors [2].
Frequently asked questions
How much electricity does AI use?
The IEA estimates all data centres used about 485 TWh in 2025 and projects around 950 TWh by 2030; AI-focused data centres are the fastest-growing part, tripling over that period [1].
Why are AI data centres hard to power?
They need large amounts of power at one site, quickly, while grid connections, turbines and transformers are in short supply, and their load can swing by more than 50% within a second [1][2].
Do data centres raise household electricity prices?
The IEA says rising demand does not necessarily raise prices with the right policies, but concentrated, fast-growing loads can strain affordability [2]. In PJM, the market monitor attributes billions of dollars of capacity costs to data-centre load [6].
Is Google's Wyoming data centre confirmed?
Yes. Google announced a $17 billion Wyoming investment, including its first data centre in the state, on 5 November 2026, with a 702 MW fuel-cell development and carbon capture contracted with AEP [8].
Which country has a very high data-centre share of electricity?
Among the cases with official data cited here, Ireland stands out: data centres used 23% of its metered electricity in 2025 [7].
Is AI getting more energy-efficient?
Yes. The IEA says energy per AI task has dropped by at least an order of magnitude a year in recent years, but rising use outweighs that gain [1].
Glossary
- Terawatt-hour (TWh)
- A unit of energy equal to one billion kilowatt-hours, used for national or global electricity use.
- Capacity auction
- A market in which a grid operator pays power plants and other resources to be available in a future year; PJM's is the Base Residual Auction [5].
- Independent Market Monitor
- An organisation that reviews a power market's results and behaviour; for PJM this is Monitoring Analytics [6].
- High-bandwidth memory (HBM)
- Stacked memory chips used alongside AI processors; the IEA expects a shortage to last until at least the end of 2027 [1].
- Small modular reactor (SMR)
- A smaller nuclear reactor design meant to be built in factory-made modules.
Sources
- Key Questions on Energy and AI: Executive summary International Energy Agency (IEA) · 2026-04-16
- Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions International Energy Agency (IEA) · 2026-04-16
- Energy and AI: Energy demand from AI International Energy Agency (IEA) · 2025-04-10
- Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers Lawrence Berkeley National Laboratory · 2025-01-17
- 2028/2029 RPM Base Residual Auction Report PJM Interconnection · 2026-07-14
- Analysis of the 2028/2029 RPM Base Residual Auction, Part A Monitoring Analytics (Independent Market Monitor for PJM) · 2026-10-30
- Data Centres Metered Electricity Consumption 2025: Key Findings Central Statistics Office Ireland · 2026-07-07
- Investing in Wyoming's infrastructure, energy and workforce Google · 2026-11-05
Comments
Top comments