A live, source-checked counter of the electricity, water, and carbon the world's AI data centers consume — plus a free calculator for your own AI energy footprint, per prompt.
175 TWh
AI servers, 2026 (Gartner)
565 TWh
all data centers, 2026
945+ TWh
projected by 2030 (IEA)
0.24 Wh
per Gemini text prompt
—
current global AI power draw
—
MW
—
L/sec
—
kg CO₂/sec
Since you opened this page i
0
Consumed by AI servers since Jan 1, 2026
0
ProjectAbout AI Energy Counter
AI Energy Counter is an independent, informational project built to make dense industry research about AI's electricity, water, and carbon footprint readable at a glance. Every figure on this page traces back to a named public source — Gartner, the International Energy Agency, or Google's own published technical papers — instead of an unsourced guess.
The site isn't affiliated with any AI company, cloud provider, or research institution mentioned here. Its goal is to translate slow-moving annual industry reports into something you can watch tick by in real time, understand the assumptions behind, and use to .
Methodology — how this is calculated
This is a transparent estimate, not a measured live feed — no organization publishes a true second-by-second global AI power reading. The numbers extrapolate published annual figures at a constant rate.
Energy baseline:Gartner's June 2026 forecast puts AI-optimized server electricity use at ~175 TWh for 2026, a run-rate of roughly 5,550 kWh/second, applied continuously from Jan 1, 2026.
Water conversion: assumes 1.9 L per kWh, a commonly cited industry mid-range cooling intensity figure. Real-world water use per kWh varies by cooling method and climate.
Carbon conversion: assumes 0.45 kg CO₂ per kWh, a global average grid emissions factor. Actual intensity depends heavily on each facility's local grid.
Personal calculator: baselined on Google's published Gemini Apps technical paper (Aug 2025) — 0.24 Wh, 0.26 mL, 0.03 g CO₂e per median text prompt — used as an illustrative, not universal, baseline. Image and video generation use substantially more.
Country ticker: the top-10 list at the very top of the page is a synthesized estimate, not a single official ranking — no agency publishes country-level electricity use broken out for AI specifically. It's built from data center electricity-share reporting including the IEA's Energy and AI report, used as a proxy for AI's geographic footprint since AI-only figures aren't separately published per country.
Treat every number here as an order-of-magnitude estimate for an informational tool, not a citable scientific measurement.
AI-optimized servers are on pace to use ~175 TWh of electricity in 2026 — about 5,550 kWh every second.
A single AI text prompt uses roughly as much energy as one Google search (~0.24–0.3 Wh).
That same prompt carries about 0.03g of CO₂ and 0.26mL of water footprint, per Google's own published figures.
Global data center electricity use could roughly double by 2030, per the IEA.
Global dataHow Much Electricity Does AI Use in 2026?
AI has become the fastest-growing source of electricity demand in the data center industry. According to Gartner's June 2026 forecast, AI-optimized servers alone are on track to consume about 175 TWh of electricity this year — up 84% from roughly 95 TWh in 2025 — and now make up around 31% of all data center power use worldwide. Total worldwide data center electricity consumption, AI and non-AI combined, is forecast to hit 565 TWh in 2026, a 26% jump from 2025, with peak power demand climbing to 132 gigawatts.
The International Energy Agency (IEA) puts 2024's baseline at about 415 TWh, roughly 1.5% of global electricity use, and projects that figure could roughly double to 945 TWh by 2030 in its base case. In the United States, which accounts for close to 45% of global data center electricity use, the Department of Energy's Lawrence Berkeley National Laboratory projects consumption could climb from 176 TWh in 2023 to as much as 580 TWh by 2028.
A growing share of that power may never touch the public grid at all. Meta, Microsoft, Google, and Amazon have each announced dedicated natural gas power plants built specifically to bypass grid interconnection queues — a trend a September 2026 BloombergNEF forecast suggests could push US data centers to consume more natural gas than Germany and Japan combined by 2035, roughly 18 billion cubic feet per day. This tracker's electricity, water, and carbon figures are all grid-average estimates and don't separately capture this on-site "behind-the-meter" generation — see the for what that term means and why it matters.
Year
Global data center electricity
Source
2024
~415 TWh
IEA
2025
~447–485 TWh
Gartner / IEA
2026
~565 TWh
Gartner
2030 (forecast)
~945–1,200+ TWh
IEA / Gartner
Continue reading: · ·
CoolingHow Much Water Does AI Use?
Densely packed AI chips generate enormous heat, and many hyperscale data centers rely on evaporative cooling to keep them running — water that evaporates or needs treatment rather than simply being reused. Google reports its median Gemini text prompt uses about 0.26 milliliters of water. Scaled across billions of daily queries, and combined with training runs that can consume millions of liters at a single facility, water has become one of the more contested resource questions in AI infrastructure siting, particularly in drought-prone regions.
See it live: ·
EmissionsWhat's the Carbon Footprint of One AI Prompt?
Google's figure for a median Gemini text prompt is about 0.03 grams of CO₂ equivalent — a number that reflects heavy efficiency engineering and access to lower-carbon electricity. That figure isn't universal: a prompt processed on a coal-heavy grid, run through a larger or slower model, or involving image and video generation can carry a meaningfully larger carbon cost. This counter uses a global average grid intensity of 0.45 kg CO₂ per kWh to estimate aggregate emissions across all AI infrastructure worldwide, not just one optimized provider.
·
What's Your Personal AI Footprint?
Drag to match how many AI prompts you send a day — chats, image generations, searches — and see the estimated yearly cost.
20 prompts / day
Electricity / year
—
Water / year
—
CO₂ / year
—
🧾
Want a full breakdown by activity type?
This calculator blends text, image, and video into one estimate. For a detailed receipt that separates them — including AI video's outsized energy cost — try Prompt Footprint, our companion calculator.
By countryAI Energy Use by Country: Who's Consuming the Most?
No agency publishes a country-by-country breakdown of AI-specific electricity use — that data doesn't exist as a single clean source. The table below is a synthesized estimate built from data center electricity-share reporting (IEA, the Energy Institute's 2026 Statistical Review, IDCA, and Our World in Data), used as a proxy for AI's geographic footprint. The same figures also drive the scrolling ticker at the top of this page.
Rank
Country
Est. share of global AI data center electricity
#1
United States
~45%
#2
China
~20%
#3
Germany
~4%
#4
Japan
~3.5%
#5
Ireland
~3%
#6
United Kingdom
~3%
#7
Singapore
~2.5%
#8
Netherlands
~2%
#9
France
~2%
#10
Canada
~1.5%
See the behind this estimate ·
Per querySearch vs. Text vs. Image: Comparing AI's Energy Cost
It's a closer call than most people assume — for text, at least. Google's own August 2025 technical paper reports a median Gemini Apps text prompt uses about 0.24 watt-hours of electricity — in the same ballpark as the company's long-cited estimate of roughly 0.3 watt-hours for a typical web search. Image generation is where the gap actually opens up: a 2023 Hugging Face and Carnegie Mellon study found a single AI-generated image uses around 2.9 watt-hours — roughly ten to twelve times a text prompt or search.
Google Search
~0.30 Wh
Gemini text prompt
~0.24 Wh
AI image generation
~2.9 Wh
Video generation is off this chart entirely — one widely-cited estimate puts a single 5-second AI video clip at ~940 Wh (MIT Technology Review's analysis of a Hugging Face CogVideoX test), roughly the same energy as 7-8 hours of Netflix streaming at ~120 Wh/hour (TRG Datacenters). Other independent tests found a wide range (roughly 7-214 Wh), so treat this as a high-end estimate, not a fixed number. See the full breakdown on Prompt Footprint, our companion calculator.
Advertisement — in-content unit
NewsroomLatest AI Energy & Data Center News
A running list of notable stories on AI's power demand, grid strain, nuclear deals, and hardware efficiency — curated with links to the original reporting. Updated periodically as the story develops.
Want the numbers behind the headlines? ·
Advertisement — mid-page unit
AnswersFrequently Asked Questions
Based on Gartner's June 2026 forecast, AI-optimized servers are on pace to consume roughly 175 terawatt-hours (TWh) of electricity in 2026, up from about 95 TWh in 2025 — a run-rate of around 5,550 kilowatt-hours every second. That's part of a wider global data center electricity total projected at 565 TWh for 2026.
Google's own August 2025 technical paper found a median Gemini Apps text prompt uses about 0.26 milliliters of water, mostly for data center cooling. Water use varies a lot by data center location, cooling technology, and local climate, and is typically much higher for image or video generation.
Google reported a median Gemini text prompt emits about 0.03 grams of CO2 equivalent. That figure reflects an efficiency-optimized system on a partly clean grid; a prompt run on a coal-heavy grid, or a longer reasoning or image-generation task, can emit meaningfully more.
Roughly comparable, according to the companies' own published figures. Google's often-cited estimate for a search is about 0.3 watt-hours, close to the 0.24 watt-hours Google reports for a median Gemini text prompt. Longer AI conversations, reasoning models, and multi-step AI agents can use substantially more than a single search.
The IEA's base case projects global data center electricity use roughly doubling from about 415 TWh in 2024 to 945 TWh by 2030. Gartner's more recent forecast puts total data center consumption above 1,200 TWh by 2030, with AI-optimized servers accounting for close to half of that.
No — and no organization currently publishes a true real-time global feed of AI power draw. This counter takes published annual estimates (like Gartner's 175 TWh 2026 figure) and extrapolates them at a constant per-second rate, which is clearly disclosed in the methodology section.
At the 175 TWh/year run-rate for AI-optimized servers (Gartner, 2026) and this page's 1.9 L/kWh water-use-effectiveness assumption, that's roughly 911 million liters of water per day — equivalent to about 15 million average American showers a day (59.8 L per shower, per the Residential End Uses of Water Study). Treat this as an order-of-magnitude comparison, not a precise measurement — actual water use varies hugely by data center, cooling method, and climate.
Data centers generate enormous heat from densely packed AI chips, and many cooling systems use evaporative water cooling to keep servers from overheating. That water either evaporates or requires treatment before being returned, which is why it counts as consumption rather than simple usage.
Estimates vary by model size, but Google's own figures put a typical Gemini text response at around 0.24 watt-hours of electricity — in the same range publicly discussed for ChatGPT-style queries. Longer conversations, reasoning models, and multi-step AI agents can use several times more per response.
A 2023 Hugging Face and Carnegie Mellon University study found generating a single image with a diffusion-style model used around 2.9 watt-hours of electricity on average — roughly ten times more than a typical text prompt, and comparable to partially charging a smartphone.
AI-specific figures aren't tracked separately from broader data center electricity use, but the IEA estimated all data centers consumed about 1.5% of global electricity in 2024. AI workloads are the fastest-growing share of that total and are expected to make up the majority of data center power within the next several years.
Google's 2024 environmental report disclosed that its global data centers consumed about 8.1 billion gallons (roughly 31 billion liters) of water that year, up sharply from 6.1 billion gallons in 2021 as AI workloads scaled. That figure covers all of Google's data centers, not AI exclusively.
It depends on scale and how the electricity is generated. A single AI query has a small footprint, but billions of daily queries plus the electricity and water used to build and cool AI infrastructure add up to a meaningful and fast-growing environmental impact — one several major research bodies, including the IEA, are actively tracking.
The same 2023 Hugging Face and Carnegie Mellon study estimated roughly 1.1 grams of CO2 equivalent per generated image on average for diffusion-style models — notably higher than the 0.03 grams Google reports for a median optimized text prompt, though both vary widely by model and hardware.
Significant, and often overlooked. On the materials side, research on data center construction finds concrete and steel typically account for 70-85% of a facility's construction-related emissions, with a lifecycle study by operator Data4 finding construction (not operation) made up 39% of one facility's total carbon footprint. On the labor side, a 2023 TIME investigation found OpenAI had outsourced labeling of toxic and disturbing content to Kenyan workers paid roughly $1.32-$2/hour, some of whom reported lasting psychological harm. This page focuses on energy, water, and carbon, but these supply-chain and labor costs are a real part of AI's footprint too.
They're now in a similar order of magnitude. The Cambridge Centre for Alternative Finance estimates Bitcoin mining consumes roughly 130-150 TWh per year, close to Gartner's 175 TWh estimate for AI-optimized servers in 2026 — though AI's electricity use is growing far faster and is projected to pull well ahead by 2030.
It depends on scale. Training a single large model can consume many thousands of megawatt-hours in one continuous run, but once deployed, the cumulative energy spent answering millions of daily queries — inference — typically overtakes the one-time training cost over a popular model's lifetime.
The United States, by a wide margin — estimated at around 45% of global AI data center electricity use, based on synthesized data center electricity-share reporting from the IEA, the Energy Institute, and IDCA. China follows at an estimated 20%, with no other single country above 5%.
Using this tracker's current estimate of roughly 20 gigawatts of continuous global AI power draw, and a typical nuclear reactor's output of about 1 gigawatt, that's the rough equivalent of 20 reactors running continuously today — a figure Gartner and IEA forecasts suggest could triple or more by 2030.
A high-end AI training chip like Nvidia's H100 draws up to roughly 700 watts under full load. Large training clusters run thousands of these chips continuously for weeks or months at a time, which is why a single model's training run can consume as much electricity as a small town.
Yes. This page's calculator blends all AI activity into one estimate. Prompt Footprint, a companion calculator, separates text prompts, image generation, and video generation individually — useful since video generation can use two to three orders of magnitude more energy than a single text prompt.
Yes. Meta, Microsoft, Google, and Amazon have all announced dedicated on-site natural gas power plants to bypass grid interconnection delays. A September 2026 BloombergNEF forecast projects US data centers could consume about 18 billion cubic feet of natural gas per day by 2035 — more than Germany and Japan combined — with roughly 3 billion cubic feet of that coming specifically from these on-site, behind-the-meter plants.
DefinitionsAI Energy Glossary: Key Terms Explained
A quick reference for the units and acronyms used throughout this page.
TWh (Terawatt-hour)
A unit of energy equal to one trillion watt-hours — the standard measure for national or global electricity consumption. One TWh could power roughly 90,000 average U.S. homes for a year.
GW (Gigawatt)
A unit of power (the rate of energy use) equal to one billion watts. This tracker's gauge shows global AI's current draw in GW — a snapshot of power, not cumulative energy.
PUE (Power Usage Effectiveness)
A ratio measuring a data center's total energy use against the energy actually used by its computing equipment. A PUE of 1.0 would mean zero energy overhead from cooling and infrastructure; real-world facilities typically run 1.1–1.5.
WUE (Water Usage Effectiveness)
The data center equivalent of PUE for water — liters of water consumed per kilowatt-hour of IT energy delivered. This tracker uses an assumed WUE of 1.9 L/kWh, a commonly cited industry mid-range figure.
Inference vs. Training
Training is the one-time (but extremely energy-intensive) process of building an AI model. Inference is every subsequent use of that model to answer a query — cheap per request, but multiplied across billions of daily uses.
Hyperscaler
A company operating data centers at massive scale — typically Amazon, Google, Microsoft, and Meta — with enough infrastructure to run cloud and AI services for millions of customers simultaneously.
kWh (Kilowatt-hour)
A unit of energy equal to one thousand watt-hours — the unit this page's live counters actually tick up in every second. For scale: running a standard microwave for about an hour uses roughly 1 kWh.
Grid Interconnection Queue
The waitlist a new power source or major consumer (like a data center) must join to connect to the electrical grid — currently stretching years in high-demand U.S. markets, a key bottleneck covered in the .
Behind-the-Meter Generation
Power generated on-site — typically by natural gas turbines — and consumed directly by a data center without ever passing through the public grid. It's the workaround companies are increasingly using to skip the grid interconnection queue above, at the cost of significant additional natural gas consumption and emissions not counted in grid-average electricity figures.
Global buildoutThe 20 Largest AI Data Centers, Ranked
Ranked by current, disclosed IT power capacity — not planned end-state — sourced from Epoch AI's AI Data Centers Hub (CC-BY 4.0), an independent database built from satellite imagery and permitting records rather than press releases alone. Colossus 2 (xAI, Memphis) leads at 946 MW. Coverage skews heavily toward the United States — that reflects real concentration more than a blind spot, per Epoch, though international tracking is still expanding. Several under-construction sites (like Microsoft's Fairwater Wisconsin, planned for ~2,263 MW) will eventually dwarf everything here — this ranks what's operating today.
This site uses essential cookies to function, and — only if you agree — cookies from Google AdSense to show ads based on your visits here and elsewhere. Visitors in the EU/UK and California have specific rights over this. See for details.