Data Centre Electricity Demand Jumps 17% as Grid Bottlenecks Tighten

A record year that the grid could barely absorb

The International Energy Agency's latest numbers show that the AI boom is now a power story. The agency reports that global data centre electricity demand rose 17% in 2025. Overall electricity demand grew about 3% over the same period.11 Facilities built for AI grew much faster. Several outlets report that the IEA put their consumption growth at 50% for the year.12 The IEA's full special report, Key Questions on Energy and AI, says the 17% figure matches what the agency had projected a year earlier.17

The forecast is just as steep. The IEA expects data centre consumption to roughly double, from 485 TWh in 2025 to 950 TWh in 2030, or about 3% of global electricity use. Over the same period, consumption from AI-focused facilities is expected to triple to around 465 TWh, close to what conventional data centres use.12 Outside summaries add a comparison: 950 TWh is about what Japan uses in electricity today.15

Spending explains the speed. Amazon Web Services, Google, Meta, Microsoft and Equinix together spent more than $400 billion in capital expenditure in 2025. That figure is expected to rise another 75% in 2026.1218

Efficiency is improving fast, but demand is growing faster

The IEA's findings show a clear paradox. The energy used per AI task is falling quickly, by at least an order of magnitude a year recently, which the agency calls unprecedented in energy history.12 Total consumption is still rising because more people use AI and because heavier workloads, especially AI agents, are spreading.11

The report includes one telling calculation. Even huge volumes of text queries would account for only a small fraction of current data centre demand. So most of the capacity being planned must be meant for other work, such as model training, image and video generation, and autonomous agents.17 Put plainly, the energy problem comes less from chatbots and more from the move toward agents and media generation, which use far more compute. The IEA says efficiency gains on their own will not flatten the curve, and that reading looks right.

The bottleneck moves to the substation

The report's main warning is about physical limits. Supply chains for gas turbines, transformers, advanced chips and other IT components have all tightened over the past year. The growing pipeline of data centre projects is also overloading planning and regulatory systems and delaying grid connections.11 The full report adds more detail:

• Gas turbine orders rose 70% in 2025.
• Trade restrictions have hit inputs such as critical minerals for power electronics.
• High-bandwidth memory is in short supply.
• Community opposition has led to cancelled or delayed projects.17

The IEA says these bottlenecks make the more aggressive near-term growth scenarios less likely, even with record investment. It also sees possible upside in the medium term if the constraints ease.17

Industry reporting from the US backs this up. Several analyses say the main constraint has moved from GPUs to grid power. Chip packaging capacity has expanded several times since 2024, while interconnection queues, transformer factories and utility planning have not kept pace.4 One industry analysis says power availability is now the main factor in choosing data centre sites, ahead of latency.2

The wait times are long. JLL's 2026 outlook, as cited by one industry analysis, puts the average grid-connection wait in major US markets at more than four years, and about seven years in Northern Virginia. A data centre building itself can go up in 18 to 24 months.3 Equipment is a separate obstacle. Wood Mackenzie data cited in the same coverage puts average lead times at 128 weeks for large power transformers and 144 weeks for generator step-up units.38

The cost of delay is showing up in build schedules. One analysis reports that only 5 GW of the 16 GW of new data centre capacity planned for 2026 had started construction. It projects that 30% to 50% of the rest could slip into 2027 or later.8 A Bloom Energy survey found that developers expect to get power about two years before utilities say they can provide it. More than half of developers said power had become harder to secure over the previous year.9

Different numbers, same direction

The sources do not agree on how large the US queue is, and the differences matter. Lawrence Berkeley National Laboratory's 2025 report counted 2,061 GW of generation and storage seeking connection. That was a 10% drop from the year before, but it came from developers withdrawing projects, not from faster processing.4 Other analyses give figures of about 2,300 GW1 or 2,600 GW once load requests are counted.35 These are different measures of the same backlog, not conflicting facts.

Texas shows how fast the numbers move. ERCOT's large-load queue is reported at 226 GW in one count, up from 63 GW a year earlier8, about 410 GW by April 20264, and 474 GW by September, with roughly 90% of it from data centres.6 The IEA's own analysis of ERCOT says large-load requests grew more than 3.5 times between 2024 and 2026.17

The IEA also warns that much of this pipeline will never be built. Only an uncertain share of applications becomes real demand, which makes planning harder for grid operators.17 Older data supports this: only 13% of interconnection requests filed between 2000 and 2019 were operating by the end of 2024.8 Queue totals are better read as a measure of demand pressure than as a forecast of future load.

The scramble for solutions

The IEA describes AI as both an "energy taker" and an "energy maker."12 Technology companies signed about 40% of all corporate renewable power purchase agreements in 2025. They are also driving interest in nuclear and advanced geothermal.12 Conditional offtake agreements between data centre operators and small modular reactor (SMR) projects rose from 25 GW at the end of 2024 to 45 GW.11 The first of those reactors is not expected to run until around 203017, so they cannot help with the current shortage.

The near-term fix in the US is on-site gas. The IEA says developers held back by slow grid connections are pursuing many such projects.11 The agency's satellite tracking suggests about one-fifth of them have started land clearing or construction.17 The IEA also finds that reliable on-site gas supply means building 30% to 70% more generating capacity than the expected load. With turbines in short supply, it is unclear that this route is actually faster at scale.17 AI workloads also swing sharply and suddenly, which can exceed what on-site gas plants can handle. That is why the IEA calls on-site batteries critical11, and summaries cite a need for about 20–25 GW of battery storage for data centres by 2030.15

Grid operators are changing their rules too. ERCOT adopted a batch-based review process, called "Batch Zero", in June 2026 for large loads of 75 MW and above.36 PJM has proposed a new category, Non-Capacity-Backed Load, for large customers that accept being cut off before others during periods of grid stress.10 Some states now make data centres carry more of the risk. Ohio's AEP requires minimum billing and collateral. Virginia has created a large-user rate class with minimum demand charges. West Virginia has set up microgrid districts that pair data centres with their own generation.8

Affordability becomes the political flashpoint

The IEA says rising demand does not have to raise prices if policy and investment keep up. It also says data centres put particular pressure on affordability because their loads are large and concentrated and grow fast.11 Recent US events suggest that risk is already real. One analysis links data centre growth to a $9.3 billion increase in PJM capacity costs for the 2025–26 delivery year. It estimates the effect at about $16–18 a month on residential bills in states such as Ohio and Maryland.6 Industry outlets also report sharper political reactions:

• Maine reportedly banned new data centre construction in April 2026.3
• Texas Governor Greg Abbott reportedly paused new data centre grid connections in August 2026 pending an audit.6

Each of those reports comes from a single outlet and should be treated with some caution.

The takeaway

In our reading, the overall trend is clear. Demand is not the limiting factor. Physical infrastructure and public consent are. The IEA's central forecast barely changed from last year, while every link in the delivery chain, from transformers to permits to public support, got tighter.17 The agency's policy recommendations are proactive management of project pipelines, flexibility that speeds up connections, and fewer barriers to using AI inside the energy sector.12 Those are reasonable, but none of them work quickly. Over the next few years, the companies that win the AI race may be the ones that can secure electricity, rather than chips.