The numbers stopped being abstract somewhere around the time Dublin started running on servers. Roughly 79 percent of the electricity flowing into Ireland's capital now feeds data centers, not homes, not factories, not the trams that move people around the city. That is one place. The pattern is global, and it is accelerating.
A new International Energy Agency report, released this April, lays out what the AI build-out is doing to the world's electricity systems. Capital expenditure from five large technology companies climbed past $400 billion in 2025 and is on track to grow another 75 percent in 2026. Electricity demand from data centers as a whole jumped 17 percent in 2025. Demand from AI-focused facilities specifically grew faster than that, well ahead of the modest 3 percent rise in global electricity consumption. By 2030, total data center electricity use is projected to roughly double to about 945 terawatt-hours, more than Japan consumes in a year. Power use from AI-optimized centers alone is set to triple and, in some scenarios, more than quadruple.
That is the headline. The interesting part is the contradiction sitting underneath it.
Europe is in the middle of an aggressive decarbonization program. Brussels has committed to cutting emissions, electrifying transport and heating, expanding renewables, and at the same time tripling the EU's data center capacity over the next five to seven years through the Cloud and AI Development Act. The math does not work cleanly. Building a transmission line in an advanced economy takes between four and eight years. Grid connection queues run anywhere from three to ten years. A data center, by contrast, can go from announcement to operational in two or three months. The infrastructure that keeps a continent's lights on is being asked to run at a speed it was never designed for.
The grid cannot keep up, so the industry is building its own
About half of all large-scale data center projects planned for the United States in 2026 are now expected to be delayed or cancelled. Not for lack of money, not for lack of demand, but because the grid physically cannot accommodate them. Of roughly 140 announced projects representing about 12 gigawatts, only a third are actually under construction. The bottleneck is mundane and intractable: transformers, switchgear, and batteries. High-power transformers used to take 24 to 30 months to deliver. Today the wait can stretch to five years.
Faced with that, hyperscalers have stopped waiting. They are building generation themselves or signing contracts that pull entirely new generation onto the grid for their exclusive use.
The IEA report tracks a striking shift toward nuclear. The pipeline of conditional offtake agreements between data center operators and small modular reactor projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts now. Microsoft signed a 20-year power purchase agreement with Constellation to restart Three Mile Island Unit 1, which had been mothballed since 2019. Meta announced 6.6 gigawatts of nuclear deals in early 2026, including agreements with Oklo and TerraPower for new SMRs targeted for 2032 and beyond. Amazon committed to more than 5 gigawatts of new nuclear capacity through deals with X-energy, Energy Northwest, and Dominion. Google is working with Kairos Power on advanced reactors tied to the Tennessee Valley Authority's grid.
These are big numbers and serious commitments. They are also bets on technology that is mostly not yet operational. The first commercial SMRs powering data centers are not expected online until around 2030, and even that timeline assumes regulatory and supply chain conditions that have rarely held in nuclear's modern history. The most concrete near-term project, Microsoft's Three Mile Island restart, is not new generation at all. It is reviving an existing reactor.
In the meantime, gas is filling the gap. New IEA satellite tracking shows a substantial pipeline of on-site natural gas generation projects at data centers, predominantly in the United States. Most are in early stages. Chevron, GE Vernova, and Engine No. 1 announced plans to deliver up to four gigawatts of gas-fired power directly to data centers, enough to supply 3.5 million homes for a year. ExxonMobil is building its own gas-to-power offering. Microsoft's Grange Castle campus outside Dublin has on-site gas generators totaling 239 megawatts.
The technical problem with gas at AI sites is interesting. AI workloads do not draw power smoothly. They produce sharp, sustained spikes followed by relative quiet, and gas turbines are not particularly graceful at handling that profile. The IEA notes that meeting AI demand reliably from on-site gas is genuinely stretching what the technology can do. The fix that is emerging looks like this: pair the gas plant with substantial on-site battery storage, smooth the load, and the same batteries can in principle export services back to the grid during low-demand periods. With the right tariff structure, an AI campus could become a grid asset rather than a grid burden. The current structure rarely provides those incentives.
Efficiency is improving faster than ever, and it is not enough
Power per AI inference is dropping at a rate the energy sector has not seen in any other technology. DeepSeek's V3 model demonstrated that training efficiency improvements alone can dramatically cut the energy required for state-of-the-art models. Token costs have fallen 280-fold in two years. Google has deployed liquid cooling that reduces power overhead by 30 percent in its TPU v6 clusters. Average rack densities have climbed from about 8 kilowatts in 2021 to 17 kilowatts in 2024, and AI racks now routinely exceed 50 kilowatts, which is hard on the building but better in watts per useful operation than what came before.
None of it is offsetting the volume increase. More people are using AI. The applications themselves are getting more demanding, with agentic systems running long chains of inference rather than single-shot queries. The image and video models are heavier than the text ones. Search providers are folding generative AI into their default product experience, where a single query that used to draw 0.3 watt-hours now draws closer to 3 watt-hours. The IEA's modeling has electricity demand from accelerated servers, the kind that run AI, growing 30 percent per year through 2030. Conventional server demand grows 9 percent. Total data center demand grows roughly 15 percent annually, more than four times the rate of all other electricity sectors combined.
What this means in practice depends entirely on where you live. In the United States, data centers are projected to account for almost half of all electricity demand growth between now and 2030. By that year, the American economy will use more electricity processing data than it does manufacturing every energy-intensive product combined: aluminum, steel, cement, and chemicals. In China, data center electricity demand is on track to roughly double from 2020 levels to 400 terawatt-hours by 2030. China and the United States together will account for nearly 80 percent of global growth.
Europe sits in an awkward middle position. EU data centers consumed about 100 terawatt-hours in 2022, around 4 percent of total demand. By 2030, McKinsey projects that data centers could account for up to a quarter of net new electricity demand on the continent. Ireland is the canary, and the bird is not in good shape. Data centers there consume around 22 percent of national electricity. In Dublin, half. A grid connection originally allocated to housing in Castlebaggot, West Dublin, was reassigned to data centers. The Irish regulator has openly acknowledged that data center demand is now competing with the country's targets for 550,000 new homes by 2040, 680,000 heat pumps, and 945,000 electric vehicles by 2030. There is not enough grid for all of it.
In West London, a cluster of new data centers has consumed so much of the local electrical capacity that residential and commercial development has been delayed by as much as a decade.
The cleanest version of the corporate response involves long-term renewable power purchase agreements. The technology sector accounted for roughly 40 percent of all corporate renewable PPAs signed in 2025. That sounds impressive, and the absolute volumes are large. The complication, particularly visible in Ireland, is that most of these contracts settle financially rather than physically, and most match electricity on an annual basis rather than hour by hour. A data center can claim 100 percent renewable power on paper while pulling fossil-generated electrons from the grid every night. Of corporate PPAs signed in Ireland, only about 14 percent are confirmed to involve a direct physical connection between the generator and the off-taker. The carbon accounting and the physics are running on separate tracks.
Then there is the other side of the ledger, the one the IEA is keen to emphasize. AI applied to energy systems can save energy. Smart building controls that respond to weather, occupancy, and prices could globally save about 300 terawatt-hours of electricity per year, roughly the combined annual generation of Australia and New Zealand. AI-optimized industrial processes can cut energy costs in heavy industry by 3 to 10 percentage points. AI is making it easier to integrate variable renewables, model new battery chemistries, and design SMRs themselves. The argument from the energy modelers is that if AI delivers on these efficiency gains widely, the emissions saved across the broader economy could outweigh the emissions added by powering AI itself.
That argument depends entirely on adoption, and the IEA itself flags that the energy sector is not making much use of AI yet. The barriers are not exotic. They are the absence of digital skills inside utilities and the absence of clean, accessible operational data. Fix those and the technology can pay for its own footprint and then some. Leave them unfixed, and you get the worst version of the trade: more electricity consumed, more gas plants built to power it, with the productivity gains showing up only in the corporate accounts of the companies selling the AI.
What the political economy actually looks like
Northern Virginia, home to the world's largest concentration of data centers, has effectively halted new permits. State legislators in Virginia, Georgia, Indiana, and Washington have either passed or are debating laws requiring data center operators to fund grid improvements proportional to the load they impose. The 2026 US State of the Union mentioned data center energy affordability by name, the kind of detail that does not make the speech without the politics getting hot. Brussels' upcoming Cloud and AI Development Act is being drafted alongside, and partly in tension with, a separate data center energy efficiency package due in early 2026 and a roadmap for digitalization in the energy sector. The same European Commission is trying to push two different accelerator pedals on different problems, and they are connected by the same wires.
The IEA report is, in tone, more constructive than alarmed. Birol's framing is that AI is becoming both an energy taker and an energy maker and that the right combination of policy and infrastructure investment can prevent rising data center demand from translating into higher electricity prices for everyone else. There is a real argument for that view. There is also a counterfactual sitting outside Dublin where households are paying for grid investments that are mostly going to serve a handful of American technology companies while their housing waits.
A useful test of whether all of this is being managed well or badly: look at where the new gas plants are being built, who is paying for them, and whether the same regulators promising decarbonization in 2050 are signing off on twenty-year fossil fuel commitments today. The answer to that question, in too many places, is yes.
Oracle admitted to a $20 billion funding shortfall on AI data center construction earlier this month and laid off staff. The company is far from alone. Every hyperscaler is scrambling for power, transformers, and skilled labor, all at once. The unit economics of AI training are still being worked out, the regulatory environment is unsettled, and the underlying electricity demand is growing at 15 percent a year. Whatever happens next, it will not look like a smooth transition.