Live estimate, not a meter reading
How much electricity and water AI data centers are using right now
There is no global meter for AI data centers. This counter takes the International Energy Agency’s annual figures and spreads them across the current second. The IEA itself treats today’s totals as estimates.
Since January 1 this year
Calendar year in UTC
Share of the UTC calendar year: 63.7%
Electricity
123.1 billion
kWh
6,123 kWh per second
Annual rate in this calendar year: 193.1 TWh
Water
43 billion
liters
2,142 liters per second
Annual rate in this calendar year: 67.5 billion L
Facilities with large numbers of accelerator chips for training and inference. That is the IEA’s definition. Water consumed in cooling, mostly through evaporation, and not returned to the same loop.
For scale
What these volumes mean
- Households, using the IEA rule of thumb that 100 MW equals 100,000 homes
- 22,042,043
- Olympic swimming pools in the year-to-date water total
- 17,219
- Times the electricity use of all German data centers in 2025 (Bitkom: 21.3 TWh)
- 9.1×
IEA curve
Electricity and water through 2030, year by year
The annual anchors come from the IEA reports. Between them sits the CAGR, not a meter reading. That is why 2026, and the live counter, already sit well above 2025.
Electricity for all data centers and AI-focused sites
The area between anchors is interpolated. 2024, 2025, and 2030 are IEA values. The dashed line marks the current calendar year.
2024
415 TWh
AI-focused: 103.3 TWh
2025
485 TWh
AI-focused: 155 TWh
2026
554.8 TWh
AI-focused: 193.1 TWh
2030
950 TWh
AI-focused: 465 TWh
IEA Key Questions on Energy and AI (2026). In-between years interpolated.
From 415 TWh in 2024 to 950 TWh in 2030
AI share of all data-center electricity
The share is about 32% in 2025 and just under 49% in the IEA 2030 base case.
Electricity in TWh, water in billion liters
Two units, one timeline from 2024 to 2030. The water line reaches 1,200 billion liters in 2030.
gradually.aiWater in 2023 and 2030, same mix
The IEA publishes the split only for 2023. 2030 uses that same mix, as the counter does. That is an assumption, not a new IEA figure.
2025 versus 2030
Watt-hours per text query
Median figures for short text prompts. Training is not included.
gradually.aiWatt-hours per agent task, GPU only
IEA Figure 2.1. This is GPU electricity, not a full facility load.
Germany 2025 against the world totals
Bitkom counts every German data center. The world totals on the right are IEA figures, excluding crypto mining.
gradually.aiYour use
What a single query costs
The world counters above are annual totals. The figures here are marginal costs per text query, as providers and researchers published them in 2025. Training is not included. The IEA agent figures are GPU electricity only, not a full facility load.
- ChatGPT, OpenAI (0.34 Wh)
- 6.8 Wh
- Gemini Apps, Google (0.24 Wh)
- 4.8 Wh
- ChatGPT water (0.000085 gallons)
- 6.4 ml
- Gemini Apps water (0.26 ml)
- 5.2 ml
- IEA agent without reasoning (1.14 Wh)
- 22.8 Wh
- IEA agent with reasoning (50 Wh)
- 1,000 Wh
today
IEA anchors
The annual figures behind the counter
| Year | All data centers | AI-focused | Total water consumption |
|---|---|---|---|
| 2024 | 415 TWhpublished | 103.3 TWhderived from +50% in 2025 | 624.4 billion Linterpolated |
| 2025 | 485 TWhpublished | 155 TWhOWID reading of IEA figure | 696.2 billion Linterpolated |
| 2026 | 554.8 TWhinterpolated | 193.1 TWhinterpolated | 776.3 billion Linterpolated |
| 2030 | 950 TWhpublished | 465 TWhpublished | 1,200 billion Lpublished |
Methodology
How the counter calculates
- 01
IEA annual anchors
Electricity comes from “Key Questions on Energy and AI” (April 16, 2026): 415 TWh in 2024, 485 TWh in 2025, and 950 TWh in 2030 for all data centers. Our World in Data reads IEA Figure 1.5 as 155 TWh for AI-focused sites in 2025. 155 plus 330 TWh of non-AI load equals the published 485 TWh. Those AI sites reach about 465 TWh by 2030. Our World in Data’s chart caption still prints 945 TWh for 2030, the total from the April 2025 IEA report. The 2026 update uses 950 TWh. Water comes from “Energy and AI” (April 10, 2025): 560 billion liters consumed in 2023 and 1,200 billion liters in 2030.
- 02
Growth between anchors
Between two IEA years the annual total grows exponentially at the CAGR. That is why 2026 AI electricity sits near 193 TWh rather than a linear midpoint. After 2030 the counter holds the last IEA value. It does not invent its own long-range forecast.
- 03
Even spread across the calendar year
The IEA publishes annual totals, not hourly loads. Inside a UTC calendar year the counter spreads that total evenly across every second, including leap years. There is no seasonality, no outage, and no peak-shaving.
- 04
Water follows the electricity share
The IEA does not split water into AI and the rest. The counter applies the current AI share of data-center electricity to the global water total. The 2023 mix stays fixed: 25% direct cooling, about 67% power plants, about 8% hardware. Hardware stays out of the live ticker because it is not an on-site operating flow.
Formula
annualTotal(year) × (seconds since January 1 UTC / seconds in the year)
What this is not
Not a live meter, not a satellite wattmeter, not a real-time operator feed. The IEA figures are modeled too. Other institutes such as S&P or the Energy Institute sit higher, in part because they include crypto mining. Newer AI halls often evaporate less water. The direct water line may therefore overstate the future.
Sources
Primary sources, not press copy
- IEA, Key Questions on Energy and AI
- IEA, Energy and AI
- Our World in Data, How much energy do data centers and artificial intelligence use?
- Google Cloud, Measuring the environmental impact of AI inference
- Sam Altman, The Gentle Singularity
- Epoch AI, How much energy does ChatGPT use?
- Bitkom/Borderstep, Rechenzentren in Deutschland
Frequently asked questions
- Is this the real live consumption?
- No. There is no public real-time measurement of all AI data centers. The counter interpolates IEA annual figures onto the current UTC second and says so on the page.
- What counts as an AI data center?
- The IEA means sites with large numbers of specialized accelerators for training and running AI models. Classic cloud, mail, and streaming halls sit in the larger “all data centers” total. The line is not always clean in practice, because the same campuses often mix loads.
- Why is power-plant water larger than cooling?
- The IEA estimates about 140 billion liters of direct cooling in 2023 and 373 billion liters tied to electricity generation. On-site evaporative cooling is visible. The water behind coal, gas, or some nuclear plants is not, but it is larger in total.
- How much does a ChatGPT query use?
- OpenAI put it at about 0.34 watt-hours and 0.000085 gallons of water in 2025. Epoch AI independently estimates a typical GPT-4o query at around 0.3 watt-hours. Google reports 0.24 watt-hours and 0.26 milliliters for a median Gemini Apps text prompt. IEA Figure 2.1 puts an agent without reasoning at 1.14 watt-hours of GPU electricity, and 50 watt-hours with reasoning.
- Why do other headlines differ so much?
- All data centers get mixed up with AI data centers, US figures get scaled to the world, or crypto mining gets included. Some older ChatGPT figures of 3 watt-hours now look about ten times too high.
- Is Germany in the counter?
- Not as its own live series. Bitkom/Borderstep put all German data centers at 21.3 billion kWh in 2025. That is a comparison only, not a second ticker.