The AI Power Paradox: Each AI Task Uses Far Less Energy, Yet Data Center Power Is Set to Double
The IEA says the energy needed for one AI task has fallen at least tenfold a year. It also forecasts that data center electricity will nearly double from 485 TWh in 2025 to 950 TWh in 2030. Both are true, because volume and the amount of work behind each request are growing faster than efficiency.
2030 figures are IEA forecasts, not measurements. AI-focused data centers are part of the 950 TWh total, not in addition to it. Source: IEA, Key Questions on Energy and AI (April 2026), CC BY 4.0.
A server rack is about the size of a refrigerator. By 2027, the most advanced AI racks are expected to draw around 600 kilowatts. The International Energy Agency (IEA) compares that to the peak power demand of 65 households (IEA, April 2026).
At the same time, AI is becoming more efficient at a pace rarely seen in energy technology. The IEA says the energy needed for one AI task has fallen by at least an order of magnitude a year in recent years.
So every task needs less energy, and the machines doing those tasks need more power than ever. This is the AI power paradox, and it has the same answer as many NodeZero stories: look at volume, not just the unit.
The wrong way to count AI energy
Most discussions start with one question: how much electricity does one AI query use? On its own, the answer says very little.
The IEA estimates that if every conventional web search became a simple AI text query, it would use less than 4 TWh of electricity a year, under 1% of what data centers use today Estimated. Simple text is not where AI electricity is growing.
A better way to count separates three forces:
One short answer may be cheap. A reasoning model tests alternatives before replying. An agent plans, searches, calls tools, checks the result and tries again. A video model generates millions of pixels across hundreds of frames. To the user, each looks like one request. Inside the data center, they are very different amounts of work.
Efficiency really is improving
Better chips do more arithmetic in parallel and use lower precision where it is good enough. Memory sits closer to processors. Software keeps more of the hardware busy, and models are compressed and routed so that each request does not need every part of the model. Together, the IEA says, these advances have cut the energy per AI task by at least an order of magnitude annually in recent years Reported.
Manufacturers make similar claims for new hardware. NVIDIA says its Vera Rubin NVL72 system delivers up to 10× more inference throughput per watt at one-tenth the cost per token, compared with its GB200 NVL72 system, in its own test of one reasoning model (NVIDIA, March 2026; test basis) Vendor claim. That is the manufacturer’s figure for a specific workload, not an independent measurement, but it points the same way.
Efficiency matters: without it, today’s AI services would need far more chips, money and electricity. But efficiency also lowers cost, and lower cost makes new products economical, attracts more users and makes heavier tasks affordable. Efficiency does not end the story. It changes its scale.
One prompt is becoming a workload
The second change is what counts as one task. The IEA estimates that reasoning, agentic work and video generation can use hundreds or thousands of times more energy per query than simple text generation Estimated.
That does not mean every agent request uses exactly a thousand times more than a chatbot answer. There is no standard AI query, and providers publish too little task-level energy data for that kind of precision. But the direction is clear: longer contexts, deeper reasoning, generated media and multi-step jobs.
An efficiency gain makes every step cheaper. If the product responds by taking fifty steps instead of one, total work can still rise. A more fuel-efficient car does not guarantee lower national fuel use; that also depends on how many cars travel, how far and on what journeys.
Volume is moving faster than efficiency
Now add volume. According to the IEA:
- Its satellite tracking shows that dedicated “AI factories” have more than tripled in capacity in 18 months Observed.
- Major model providers reported a threefold increase in active users and a fivefold increase in revenue over the past year. The IEA does not name the providers Reported.
- In 2025, electricity consumption across all data centers grew 17%, and at AI-focused data centers it grew 50% Observed.
Put those numbers into the identity. If the energy needed per task fell at least tenfold while the electricity used by AI-focused data centers still rose 50%, then the total amount of AI work, volume multiplied by intensity, must have grown roughly 15-fold or more (10 × 1.5) Calculated by NodeZero. This is a rough illustration rather than a measurement: the tenfold figure describes recent years in general, and facility electricity includes cooling and other overheads. But it shows the scale of the gap.
The IEA’s central forecast has global data center electricity rising from 485 TWh in 2025 to 950 TWh in 2030, an increase of 96%, with electricity used by AI-focused data centers more than tripling to around 465 TWh Forecast. Two points matter: this is a forecast, not a meter reading from 2030, and the 950 TWh covers all data center workloads, not only AI.
The rack changes the building
This growth changes more than the electricity bill. It changes the data center itself. The IEA says the power density of AI servers increased 11 times between 2020 and 2025, and expects a further fourfold increase by 2027 Reported.
The 13 kW (Ampere, 2020) and 130 kW (Blackwell) racks are IEA examples; 600 kW is the announced Rubin rack design that the IEA expects around 2027. The '65 households' comparison is the IEA's; it implies about 9 kW of peak demand per household (NodeZero arithmetic). Source: IEA, Key Questions on Energy and AI (April 2026), CC BY 4.0.
Packing more computing into each rack shortens network distances and improves performance, but it concentrates electricity and heat in a much smaller space. Air cooling gives way to liquid cooling. Power supplies grow. Batteries and backup systems must respond to denser loads. Transformers, switchgear and cables must carry more power to the same floor area.
The load can behave differently, too. Thousands of chips working in synchronized cycles can create rapid, repeated swings in power demand. In May 2026, the US Department of Energy (DOE) published guidance on monitoring these oscillations from large data centers, noting that standard grid sensors can “miss or misrepresent high-frequency oscillations” (DOE, 28 May 2026) Reported.
The bottleneck has moved outward: first the chip, then memory and networking, and now the building, the substation and the grid beyond the fence.
$720 billion meets the power system
This is why the investment numbers have become so large. The IEA says capital spending by the largest technology companies exceeded $400 billion in 2025, and that leading cloud providers have announced a 75% rise for 2026, to about $715 billion, mostly for new data center capacity Reported.
Company guidance shows the scale. The four largest US cloud and platform companies alone expect to spend $720–740 billion in 2026 Calculated by NodeZero:
Ranges are drawn at their midpoints. Definitions differ: Microsoft's and Meta's figures include finance leases; Amazon's is cash capital expenditure. Microsoft's is for calendar 2026. These are company-wide totals, not AI-only spending. The combined range is a NodeZero calculation. Sources: Amazon Q3 2026 results (SEC filing); Alphabet Q3 2026 earnings call; Microsoft FY2027 Q1 earnings call; Meta Q3 2026 results (SEC filing).
These are not pure AI totals; they include buildings, servers, networks and other infrastructure. But the physical demand behind them is clear. Microsoft, for example, lowered its calendar-2026 outlook from about $175 billion to about $170 billion in October, citing delays in capacity from third-party providers. That is a reminder that the limit is increasingly physical delivery, not willingness to spend. You can follow these companies’ quarterly spending on the AI Ecosystem Dashboard.
A data center can be financed, designed and filled with servers faster than a region can add transmission lines, substations and power plants. The technology industry works in product cycles measured in months; electricity infrastructure works in construction cycles measured in years.
The transformer problem
A transformer changes voltage as electricity moves from transmission lines into substations, buildings and racks. Without the right one, a finished data center cannot plug in.
According to the DOE, demand for distribution transformers has risen 41% since 2019. Typical delivery times went from three to six months in 2019 to one to two years or longer in 2024, and lead times for large transformers have grown to as much as four years (DOE webinar, 5 March 2026) Reported.
Ranges from DOE: 3–6 months for distribution transformers in 2019; 1–2 years 'or even longer' in 2024; large power transformers up to about 4 years. Bars show the top of each range. Source: US DOE, distribution transformer webinar (5 March 2026).
AI did not create this shortage on its own. The grid is aging, manufacturing capacity is limited, and electrification, renewable generation and new factories all need the same equipment. Data centers arrive on top of that pressure, often as unusually large projects that want power unusually quickly.
Supply is also concentrated abroad. The DOE says the United States imports about 80% of its power transformers and 50% of its distribution transformers (DOE, 26 October 2026) Reported. Estimates vary: a Congressional Research Service report puts distribution transformer imports at 25–30% (CRS R48933, April 2026). The DOE also counts almost 40,000 different distribution transformer configurations in use across the country, which makes standardized production harder.
A model can be updated overnight; the equipment that connects its data center can take years.
Globally small, locally enormous
Is AI consuming the world’s electricity? No. Under the IEA’s central forecast, all data centers together account for around 3% of global electricity demand in 2030 Forecast. The real issue is concentration.
Ireland shows what that looks like. Its Central Statistics Office (CSO) measured data center consumption at 7,663 GWh in 2025, up 10% in a year, while use by all other customers rose 2% (CSO, 7 July 2026) Observed.
Source: Central Statistics Office (Ireland), Data Centres Metered Electricity Consumption 2025, CC BY 4.0.
The United States is earlier on that curve. Lawrence Berkeley National Laboratory (LBNL) estimates that data centers used 192 TWh, or 4.7% of US electricity, in 2024. In its 2030 reference case, that reaches 649 TWh, or 11.8%, with scenarios from 9.5% to 15.3% (LBNL, June 2026) Forecast.
2030 is a projection. The reference case is 11.8%; LBNL's scenarios range from 9.5% to 15.3%. Source: LBNL, United States Data Center Energy Usage Report: 2025 Update (June 2026).
The width of that range shows how much depends on chip shipments, how busy the hardware is, inference demand, efficiency, and whether announced facilities are actually built. AI’s electricity challenge is not spread evenly across the planet. It shows up first in particular substations, towns and power markets.
Build, wait or bend
Developers have three broad options: wait for the grid, bring their own power, or make their demand more flexible.
- Wait. In many regions, a new grid connection can take years.
- Build on site. Some US projects plan gas-fired generation next to the data center. Reliable on-site power needs spare capacity: the IEA says it can mean building 30% to 70% more generation than the site’s demand Estimated. Turbines, fuel connections and permits have their own queues.
- Bend. A data center can accept a staged or “non-firm” connection, reducing its demand when the grid is stressed. Batteries can smooth swings, and work that is not urgent can sometimes move to another hour or place. The IEA estimates that 750–900 GW of projects waiting in connection queues worldwide could be enabled through more flexible, non-firm connection agreements (IEA, Electricity 2026) Modelled. That covers all kinds of projects, not just data centers, and it is modelled potential, not capacity available today.
Flexibility is not free. An AI cluster is expensive to leave idle, and cloud operators do not always control when their customers’ workloads run. In practice the answer is a combination: build where power exists, make connection queues credible, add equipment and storage, and shift the work that can safely move.
Who pays?
The final question is not only how much electricity data centers need, but who pays for the infrastructure built to serve them. Where there is spare generation and grid capacity, a large, steady customer can spread fixed costs across more sales and may lower average costs. In a tight system, a new campus may need new generation, transmission and substations, and costs can rise, especially if a project is delayed or cancelled.
In June 2026, the US Federal Energy Regulatory Commission (FERC) ordered all six regional grid operators under its jurisdiction to justify or reform their rules for connecting data centers and other large energy users (FERC, 18 June 2026) Reported. Speaking at the same meeting, a FERC commissioner described the principle that large loads should pay their fair share of the costs of serving them, “regardless of whether the large load comes online as planned.” Grid operators’ responses are due in November 2026.
There is no single answer to what AI will do to electricity bills. It depends on location, spare capacity, regulation, financing and whether projects are real.
What the numbers do not tell us
- Task-level energy. AI providers publish little standardized data on energy per task. The IEA’s per-task and per-query figures are estimates, and the “15-fold” figure above is our rough illustration, not a measurement.
- Forecast risk. The 2030 figures from the IEA and LBNL are projections. Announced projects are not completed projects, and small changes in how busy the hardware is can move the results a lot.
- Vendor claims. NVIDIA’s efficiency figures come from its own tests on specific workloads.
- Different definitions. Company capital spending figures use different accounting definitions and include more than AI.
- Conflicting estimates. The DOE and the Congressional Research Service give different shares of US distribution transformers that are imported.
- This is not investment advice. Figures are from the public sources cited; forecasts are estimates, not certainties. For information only, not financial or investment advice.
What to watch next
- The IEA’s next data center update. Whether 2026 growth beats the 17% (all data centers) and 50% (AI-focused) seen in 2025.
- 2027 capital spending guidance, due with fourth-quarter results in late January and early February 2027.
- FERC’s large-load proceedings. The grid operators’ responses, due in November 2026, and any rules on who pays for new connections.
- Transformer lead times. Whether new factories shorten the wait in 2027.
Key terms
- Terawatt-hour (TWh): one billion kilowatt-hours, a measure of electricity used over time. One gigawatt-hour (GWh) is a thousandth of a TWh.
- Kilowatt (kW) / gigawatt (GW): a measure of power at a moment in time. A rack’s power rating is in kW; grid connection capacity is in GW.
- Inference: running a trained AI model to answer a request, as opposed to training it.
- Rack: a cabinet holding servers, network gear and power equipment.
- Distribution transformer: a transformer that lowers voltage for delivery to homes and businesses. Power transformers are larger units used at substations and power plants.
- Non-firm connection: a grid connection that lets the operator limit the customer’s demand at times of stress, in exchange for a faster or cheaper connection.
Notes on the data
- IEA figures are from Key Questions on Energy and AI (World Energy Outlook Special Report, April 2026), except the non-firm connection estimate, which is from Electricity 2026 (February 2026). Both are licensed CC BY 4.0. Credit: IEA (2026), Key Questions on Energy and AI; IEA (2026), Electricity 2026; License: CC BY 4.0. Charts on this page are drawn by NodeZero from IEA figures and are not IEA charts.
- The 600 kW rack figure and the 65-household comparison are the IEA’s. The implied 9 kW per household is NodeZero arithmetic (600 ÷ 65).
- The “roughly 15-fold” growth in AI work is NodeZero arithmetic (10 × 1.5) using the identity above. It is an illustration, not an estimate of actual task volumes.
- The IEA attributes the threefold users and fivefold revenue figures to major model providers without naming them.
- The $720–740 billion range sums the four companies’ latest guidance: Amazon about $220 billion; Alphabet $195–205 billion; Microsoft about $170 billion (calendar 2026); Meta $135–145 billion. The IEA’s $715 billion covers a different group of “leading hyperscalers and neo-clouds”.
- Irish figures come from the Central Statistics Office (Ireland), licensed CC BY 4.0. LBNL figures come from its 2025 update, published in June 2026, which replaced its December 2024 report.
- The DOE 41% and lead-time figures were given by DOE staff in a public webinar (transcript published 5 March 2026).
Sources
International Energy Agency: Key Questions on Energy and AI (April 2026) · Electricity 2026: grids (February 2026)
US government: DOE – Innovative solutions for grid equipment shortages (26 Oct 2026) · DOE – distribution transformer webinar (5 Mar 2026) · DOE – monitoring oscillations from large data centers (28 May 2026) · FERC – large load integration orders (18 Jun 2026) · CRS R48933 (23 Apr 2026) · LBNL – US Data Center Energy Usage Report, 2025 Update (June 2026)
Ireland: CSO – Data Centres Metered Electricity Consumption 2025 (7 Jul 2026)
Companies: NVIDIA – Vera Rubin platform (Mar 2026) · Amazon Q3 2026 (SEC) · Alphabet Q3 2026 call · Microsoft FY2027 Q1 call · Meta Q3 2026 (SEC)
Every figure in this article links to its source and was checked against that source before publishing. How we check figures