Your chat window feels weightless. Type a question, get an answer, close the tab. No smoke. No bill that says “inference.” Meanwhile, somewhere else, a refrigerator-sized rack is pulling household-scale power and dumping heat into a cooling loop that may be evaporating municipal water.
That gap — weightless software, heavy infrastructure — is the real story of AI in 2025–2026. Not a morality lecture about turning the chatbot off. A clearer look at what the meters actually show: electricity demand that is climbing fast, water that communities notice before global carbon charts do, and local grids that cannot expand at chat-app speed.
What the electricity numbers actually say
The International Energy Agency’s updated Key Questions on Energy and AI (2026) is the cleanest public scoreboard we have. Global data-centre electricity use grew about 17% in 2025, to roughly 485 TWh. AI-focused facilities rose faster — about 50% in the same year. The IEA’s central path still sees total data-centre electricity roughly doubling to about 950 TWh by 2030 (around 3% of world electricity), with AI-focused load tripling over that stretch.
Those are not science-fiction ceilings. They sit next to a money signal that is hard to ignore: capital spending by five large tech companies topped $400 billion in 2025 and is expected to jump roughly another 75% in 2026, according to the same IEA summary. Hyperscale balance sheets are no longer enough on their own; capital markets are increasingly part of the buildout story — which means sentiment and financing conditions now shape how fast the energy demand shows up.
A United Nations University Institute for Water, Environment and Health (UNU-INWEH) report published in June 2026 lands in a similar ballpark from a different angle: about 448 TWh in 2025, rising toward roughly 945 TWh by 2030 — country-scale consumption if you treated data centres as a nation (UNU-INWEH / Environmental Cost of AI’s Energy Use).
Put simply:
- 2025 baseline: data centres already sit in the mid-400s TWh globally.
- 2030 path: roughly double overall; AI-heavy sites grow faster still.
- Share of world power: still a few percent globally — but highly concentrated in a handful of hubs where locals feel it first.
Efficiency is real. So is the rebound.
Here is the part that confuses almost everyone: energy per simple AI task has been falling fast. The IEA says efficiency gains per task have been dropping by at least an order of magnitude annually in recent years. Google’s 2026 environmental reporting estimates a median Gemini text prompt at about 0.24 Wh of energy, 0.03 gCO₂e, and roughly 0.26 mL of water — on the order of a few drops (Google Sustainability / 2026 Environmental Report). That is genuinely small for a short text question.
Then product design pulls the other way. Reasoning traces, agent loops, image generation, and especially video can burn hundreds or thousands of times more energy than a plain text reply. UNU-INWEH notes that inference — running models for everyday prompts after training — already accounts for an estimated 80–90% of total AI energy use, and that a single short AI video can rival hundreds of thousands of trivial classification queries. ChatGPT alone has been estimated at on the order of 2.5 billion prompts per day in that report’s framing.
Economists call the pattern a rebound (Jevons) effect: when each unit gets cheaper, people (and product defaults) use more units. Efficiency does not automatically shrink the total. It often funds growth.
The meter people forget: water
Electricity headlines travel well. Water is quieter until a drought year or a utility hearing makes it loud.
A U.S. Congressional Research Service FAQ cites IEA-linked illustration numbers: a 100 MW U.S. data centre may average on the order of 2 million liters per day across cooling strategies, with roughly 725,000 liters per day consumed on site (CRS, Data Centers and Their Energy Consumption). That is a facility-scale number, not a per-prompt drop — and it is why siting in a stressed watershed is not the same decision as siting on a hydro-rich grid.
UNU-INWEH’s 2030 projection pairs that electricity path with an associated water footprint on the order of 9.3 trillion liters — framed as matching the basic annual domestic water needs of about 1.3 billion people in Sub-Saharan Africa. The same report’s punchline on measurement is worth keeping: “low-carbon” is not automatically “low-water” or “low-land.” Switching generation mixes can cut carbon while raising water or land intensity. Single-metric green labels can hide trade-offs.
Operators are not asleep. Microsoft reported cutting average water-use effectiveness from about 2.3 L/kWh in early generations to about 0.27 L/kWh in 2025, and says roughly 90% of its owned 2025 fleet uses low- or zero-water cooling approaches. In 2024 it introduced an AI-oriented design that uses closed-loop, direct-to-chip cooling with zero water evaporation during operations (Microsoft Cloud Ops blog, June 2026). Intensity can fall while absolute withdrawals still rise if capacity grows faster than efficiency — a tension that has already shown up in journalism around revised corporate water forecasts.
Google reported that data-centre water withdrawals in 2025 exceeded 13.5 billion gallons across its fleet, with individual campuses in the billions of gallons (Council Bluffs among the largest). At the same time, its stewardship portfolio claimed replenishment of roughly 78% of 2025 freshwater consumption. Both can be true: absolute use is large, and replenishment projects are part of the response.
Why locals feel it before global averages do
Global percentages sound modest. County maps do not.
- Ireland: UNU-INWEH highlights data centres taking about 21% of national metered electricity in 2023 — more than all urban households — with new Dublin-area approvals paused into the late 2020s. That is a planning collision, not a vibe.
- Drought-linked siting fights: the same report flags compute expansion against water stress in places such as Querétaro and earlier Uruguay drought politics around water-intensive projects.
- Grid connection queues: IEA satellite tracking shows “AI factories” more than tripling capacity in about 18 months, while grid equipment, transformers, high-bandwidth memory, and permitting move on multi-year clocks.
- Onsite gas as a workaround: constrained by slow grid ties, some U.S. developers are pursuing onsite natural-gas generation. IEA analysis suggests reliable onsite gas for variable AI loads may require overbuilding generation by roughly 30–70%, with battery storage becoming critical for second-scale load swings.
Power density makes the mismatch visceral. The IEA notes that by around 2027, a single advanced AI rack — about the size of a large fridge — could have peak demand on the order of 65 households, with heat rejection likened to dozens of residential boilers per rack and load swings of more than 50% of rated capacity within a second. That is why batteries and flexible power designs are suddenly “AI infrastructure,” not just EV accessories.
What this means if you are not a hyperscaler
Most readers are not permitting a campus. You still live downstream of product defaults and utility politics. A few grounded takeaways:
- Prefer the lightest tool that works. Short text beats agent loops; still images beat video; smaller models beat frontier models when the task is boring. Defaults are footprint decisions.
- Treat “AI for everything” as a cost, not a free upgrade. Background agents that keep thinking when you walk away are not free on the grid even if they are free on your credit card this month.
- Read local coverage, not only global charts. Rate cases, interconnection pauses, and watershed fights are where the story becomes concrete.
- Watch disclosure, not slogans. Carbon-only dashboards miss water and land. Intensity metrics can improve while totals rise. Ask for both.
- Remember the flip side. The IEA also flags AI as a tool for grid monitoring, failure prediction, and industrial energy savings — if skills, data access, and cybersecurity barriers get out of the way. The energy sector is still under-using those upsides.
What the numbers still do not settle
Honest gaps remain. Public, standardized per-query energy and water disclosures are still uneven across providers. Facility-level water reporting varies by jurisdiction. How much of the 2030 path is delayed by chip, turbine, and transformer bottlenecks — versus how much simply shifts to onsite generation — is still being resolved in real time. Emissions from data centres roughly double in IEA projections toward the mid-2030s and still look like a small slice of global power-sector emissions — which can coexist with very real local price and watershed stress.
So the calm version is this: chat feels free because the meters are somewhere else. Electricity demand is rising on a measurable path. Water and land are the sibling meters that carbon-only storytelling skips. Efficiency is impressive and insufficient by itself. And the places that host the racks — not the places that write the prompts — are where the politics land first.
As an Amazon Associate I earn from qualifying purchases. If you buy through links on this page, I may earn a commission at no extra cost to you.
Further reading
How Infrastructure Works — Deb Chachra’s clear tour of the shared systems — power, water, networks — that make cloud and AI campuses possible. Useful background when the article’s meters start to feel abstract.
How the World Really Works — Vaclav Smil on energy, food, materials, and the physical constraints behind digital-looking economies — a sober companion to AI’s watt and water story.