AI Data Center Energy & Water Use: The 2026 Statistics

Artificial intelligence runs on electricity and cooling water, and in 2026 both moved from a footnote to a headline. The best-available estimates put global data-center electricity at roughly 1,000 terawatt-hours (TWh) this year — comparable to the annual consumption of Japan — with AI the fastest-growing driver. This report compiles the key numbers on ai data center energy consumption, water use, per-query estimates, and where the United States is headed. Every figure here is an approximate value drawn from public sources, not a precise measurement, so treat each as a range and verify before citing. Last updated: July 2026

How big is data-center energy use in 2026?

The global total: roughly 1,000 TWh

Compiled projections from public energy outlooks and industry trackers broadly point to a global data-center electricity figure of around 1,000 TWh for 2026, with a plausible range of about 900 to 1,100 TWh depending on methodology and what counts as a "data center." To put that in scale, 1,000 TWh is on the order of the annual electricity consumption of Japan, and more than the total electricity use of most individual countries. The precise number is unsettled — different sources draw the boundary differently — so read it as an estimate, not a measured total.

Share of world electricity

That total represents an estimated 4 to 5 percent of global electricity demand. A decade ago the figure was closer to 1 to 2 percent. The share roughly doubled as cloud computing scaled and then rose again as AI training and inference workloads spread. Data centers are still a minority of world power use — heavy industry, buildings, and transport dwarf them — but they are among the fastest-rising line items on the grid.

How much of this is AI versus regular cloud?

The AI slice is an estimated 15–20% and climbing

Most data-center electricity still powers conventional workloads: storage, streaming, web hosting, enterprise software, and networking. AI-specific compute is estimated at roughly 15 to 20 percent of data-center power as of 2026 — on the order of 150 to 250 TWh — but it is the fastest-growing segment and the main reason new capacity is being built. The split is inherently fuzzy because AI and non-AI workloads often share the same servers and campuses, so these shares are approximate.

Training versus inference

Training a frontier model is enormously energy-intensive but happens in bursts. Inference — answering the billions of everyday queries — is now widely believed to consume more total energy than training over a model's life, simply because it runs constantly at massive scale. As AI features get embedded into search, office software, and phones, inference is where the sustained load lives.

The core numbers: global and US data-center electricity

The table below compiles best-available estimates as of 2026. Figures are approximate and rounded; historical rows are estimates and future rows are projections, so treat the year-by-year US path as a central scenario rather than a locked forecast. Verify any specific value against its original source before citing.

The rough doubling of the global total between 2020 and 2026 is the single clearest signal in the data: demand that grew steadily for years bent sharply upward as AI capacity came online. The exact slope is uncertain, but the direction is not.

The US trajectory: heading toward roughly 400 TWh

From ~200 TWh to ~400 TWh by 2029

The United States is the epicenter of the buildout. US data-center electricity is estimated at roughly 200 TWh in 2025 and, on current projections, could reach around 400 TWh by 2029 — potentially 9 to 12 percent of total US electricity by the end of the decade. That would roughly double US data-center demand in about four years, an unusually steep climb for any single sector of the grid. As a projection, it carries wide error bars.

Why the forecast is uncertain

The ~400 TWh figure is a central estimate, and the error bars are real. Faster chip and cooling efficiency gains would pull it down; a slower efficiency curve, or AI adoption outrunning forecasts, would push it up. Grid constraints, permitting delays, and power availability may also cap how fast capacity actually comes online, regardless of demand. Read 400 TWh as "somewhere in the high-300s to mid-400s," not a precise landing spot.

Per-query energy: what one AI answer costs

Roughly 0.3 to 3 watt-hours per response

A single text answer from a large AI model is commonly estimated at about 0.3 to 3 watt-hours (Wh). For comparison, a traditional web search is often put near 0.3 Wh. So a complex AI response can use several times the energy of a search — but a single query is still tiny in absolute terms, roughly on par with running a microwave for a few seconds. The impact comes from scale: multiply a small number by billions of daily queries and it becomes grid-relevant. These are estimates, not published measurements.

Images and video cost much more

Generating an image, and especially video, can use far more energy per output than a text reply — plausibly tens of times more, though estimates vary widely by model and resolution. These per-task figures are order-of-magnitude approximations. Most providers do not publish exact per-query energy, so any single number should be treated as an estimate rather than a measured fact.

Water use: the overlooked number

Cooling is thirsty

Data centers use water two ways: directly, through evaporative cooling systems, and indirectly, through the water that thermal power plants consume to generate the electricity. A large AI-focused data center may consume on the order of hundreds of millions to a few billion liters of water per year, depending on climate and cooling design. Facilities in hot, dry regions tend to use more water for cooling, which is why siting decisions matter so much. The wide range reflects genuine uncertainty, not a single agreed figure.

Per-query water estimates

On a per-interaction basis, a common back-of-envelope estimate is that a handful of AI exchanges may indirectly use roughly half a liter of water once evaporative cooling and power-generation water are counted. Like the energy figures, this is a wide-range estimate that depends on the local grid mix and cooling method; some efficient, cool-climate sites use dramatically less. It is best cited as an illustration of scale, not a fixed rate.

The regional picture: who hosts the compute

Data-center capacity is concentrated in a handful of markets, and the estimated shares below shift as new campuses open. The United States leads by a wide margin, hosting an estimated 40 percent or more of global capacity — Northern Virginia is the single biggest hub. China follows at roughly 15 to 20 percent, driven by state-backed buildout and large domestic cloud and AI demand. Europe (EU plus the UK) sits near 15 percent, led by Ireland, the Netherlands, Germany, and the Nordics, with several grids near their limits. The rest of Asia-Pacific — Japan, Singapore, India, and Australia — accounts for roughly 10 percent (Singapore has imposed capacity caps), and the rest of the world, including expanding Gulf states and Latin America, makes up the remaining 10 percent or so from a smaller base. These are approximate shares as of 2026.

The concentration explains why local grid stress shows up first in a few places — Northern Virginia, Ireland, and parts of the Nordics — rather than evenly worldwide.

The outlook: efficiency versus demand

The race that decides the numbers

Two forces pull in opposite directions. On one side, hardware keeps getting more efficient — newer AI chips deliver more computation per watt, and better cooling lowers overhead. On the other, demand keeps rising as AI spreads into more products. For most of the 2010s, efficiency gains kept total data-center energy roughly flat even as workloads grew. Since around 2023, demand has been winning, which is why the totals are climbing again.

What to watch

The figures that will move the 2027–2030 numbers most are: the pace of chip efficiency, how much AI inference actually scales into everyday software, how quickly grids can add clean power, and whether water-lean cooling designs become standard. Small shifts in any of these swing the global total by hundreds of TWh. For deeper comparisons, see our top 10 lists and the full data library at Countly.

Frequently asked questions

How much electricity do data centers use in 2026?

Global data-center electricity is estimated at roughly 1,000 terawatt-hours (TWh) in 2026 — close to the annual electricity consumption of Japan and around 4 to 5 percent of world electricity. This is an approximate, best-available figure compiled from public agency projections and industry trackers as of 2026; different methodologies bracket it somewhere between about 900 and 1,100 TWh, so treat it as a range rather than a precise measurement and verify the latest number before citing it.

How much of that is AI specifically?

AI-specific workloads are widely estimated at roughly 15 to 20 percent of data-center power as of 2026, or on the order of 150 to 250 TWh, with the share rising. Most data-center electricity still goes to conventional cloud, storage, and networking, but AI training and inference are the fastest-growing slice and the main driver of new capacity. The split is inherently fuzzy because AI and non-AI workloads often share the same hardware, so these percentages are estimates.

How much energy does one AI query use?

A single text response from a large AI model is commonly estimated at roughly 0.3 to 3 watt-hours, versus roughly 0.3 watt-hours for a traditional web search. Image or video generation can be far higher. These are order-of-magnitude estimates as of 2026 — actual figures vary widely by model size, hardware, and how the request is served, and most companies do not publish exact per-query numbers, so any single value should be treated as illustrative.

How much water do AI data centers use?

Estimates suggest a large AI data center can consume on the order of hundreds of millions to a few billion liters of water per year for cooling, and that a handful of AI interactions may indirectly use roughly half a liter of water once evaporative cooling and power-plant water are counted. These are wide-range estimates as of 2026 that depend heavily on climate, cooling design, and the local electricity mix, so they are best cited as illustrations of scale rather than fixed rates.

Where is US data-center electricity headed by 2029?

Public projections point to US data-center electricity rising from roughly 200 TWh in 2025 toward somewhere around 400 TWh by 2029 — potentially close to 9 to 12 percent of total US electricity by decade's end. As of 2026 the exact path depends on chip efficiency, grid constraints, and how quickly AI demand materializes, so treat the ~400 TWh figure as a central estimate with wide error bars, not a certainty, and verify against the latest source.

Are these figures reliable enough to cite?

They are directionally reliable but imprecise. Every number here is a compiled estimate drawn from public sources such as energy-agency outlooks, company filings, and reputable trackers as of 2026, and different methodologies disagree. Always verify the latest figure against its original source before citing it in research or reporting, and present these numbers as approximate ranges rather than exact values.

Methodology and disclaimer: The figures in this report are best-available estimates compiled from public sources — including energy-agency outlooks, government statistics, company filings and earnings disclosures, and reputable industry trackers — as of July 2026. No single figure should be read as authoritative from one specific source; data centers rarely publish exact energy and water metrics, and independent estimates use different methods and boundaries, so numbers are presented as approximate values or ranges rather than precise measurements. Estimates may be revised as new data is released. Please verify any specific figure against its original source before citing it in research, reporting, or decision-making.