How Much Energy Does AI Actually Use? The Real 2026 Numbers
AI's electricity use is one of the most argued-about numbers in tech. Here's what Gartner, the IEA, and Google's own disclosures actually say — and why the debate is more nuanced than either side admits.
Every AI energy conversation online tends to collapse into one of two extremes: "AI is destroying the planet" or "one AI query barely uses more power than a lightbulb, relax." Both framings are technically supported by real numbers — which is exactly the problem. Here's what the data actually shows.
The headline numbers for 2026
Global data center electricity use is projected to hit roughly 565 terawatt-hours in 2026, up about 26% from 2025 — and AI-optimized servers are almost the entire reason for that jump, growing at nearly 84% annually while conventional server power draw stays essentially flat. By 2027, AI-specific power draw is expected to overtake conventional data center consumption entirely — a genuine structural crossover, not just a rounding shift.
For scale: that 565 TWh figure is still under 2% of global electricity consumption today. The International Energy Agency projects total data center electricity use roughly doubling to around 945-950 TWh by 2030, reaching about 3% of global demand — significant, but nowhere near the scale some viral claims suggest.
Why both the alarmist and dismissive takes miss something
The dismissive take usually points to a real, verified number: Google has disclosed that a median Gemini text prompt uses about 0.24 watt-hours of electricity — genuinely tiny, comparable to running a microwave for about one second. If you stop the analysis at "per query," AI looks negligible.
The alarmist take points to the aggregate: total AI energy demand is compounding fast enough that per-query efficiency gains aren't keeping up. The IEA's own analysis found that power consumption per AI task has been falling by roughly an order of magnitude per year in some cases — and yet total consumption keeps rising anyway, because the number of users and the complexity of what they're asking for (longer conversations, AI agents doing multi-step tasks) are growing even faster than the efficiency gains.
Both of these are true at the same time. That's not a contradiction — it's the same pattern seen with almost every efficiency gain in computing history: cheaper-per-unit almost always leads to more total usage, not less overall consumption.
Where this actually shows up locally
The abstract global numbers matter less than what's happening in specific places already hosting heavy data center infrastructure. Data centers already account for roughly a fifth of electricity demand in Ireland, and a similar concentration exists in parts of Virginia — one of the world's densest data center regions. Local grid operators in these areas are now planning infrastructure expansions specifically to keep pace, and the honest constraint on AI's growth in 2026, according to multiple industry analyses, isn't chip supply anymore — it's the availability of electricity itself.
The part that gets skipped: water
Electricity gets most of the attention, but data center cooling also uses real amounts of water — Google's own figures put water use at roughly a quarter of a milliliter per typical text prompt, small per-query but, like electricity, compounding at scale. Newer liquid-cooling approaches reduce direct water use substantially (estimates suggest 70-90% less than older air-cooling designs), though they still carry an indirect water footprint through the electricity they consume.
What this actually means if you're using AI tools
For an individual user, the per-query energy footprint of a typical chatbot interaction is genuinely small — smaller than most other common digital habits like streaming video. The environmental conversation that actually matters is happening at the infrastructure level: how fast new data centers get built, what energy sources power them, and how efficiently hyperscalers manage the compute behind the scenes — decisions made by a small number of very large companies, not by individual prompt volume.
If you want a single number to remember: AI-specific electricity demand is growing at multiple times the rate of general data center growth, the constraint has shifted from chips to power availability, and both of those trends are expected to continue through at least 2030 based on every major forecast currently published.