By the WahLiao desk · Last verified 1 October 2026
Super Intelligence, the kind of AI behind ChatGPT, Claude and Gemini, runs in data centres that use a lot of electricity in total but very little per question. The International Energy Agency (IEA) puts the world’s data centres at about 415 terawatt-hours (TWh) in 2024, around 1.5% of global electricity, and expects that to more than double to about 945 TWh by 2030, with AI the biggest driver. Per question, Google says a median Gemini text prompt uses 0.24 watt-hours and 0.26 millilitres of water; OpenAI’s chief executive has put an average ChatGPT query at about 0.34 watt-hours. In Singapore, data centres use about 7% of the country’s electricity, which is why the government paused new ones from 2019 to 2022 and now lets them grow only in small, green-conditioned batches.
The single most useful rule: your own chatbot use is a rounding error next to your air-conditioner. The real levers are how data centres are built, cooled and powered, and those are set by policy and operators, not by you skipping a question.
AI energy and data centres: quick facts
| World data centres, 2024 | About 415 TWh, around 1.5% of global electricity (IEA) |
| World data centres, 2030 | Projected about 945 TWh, more than double, AI the main driver (IEA base case) |
| One Gemini text prompt | 0.24 Wh and 0.26 ml of water, median, per Google (August 2025) |
| One ChatGPT query | About 0.34 Wh and about one fifteenth of a teaspoon of water, per Sam Altman (June 2025) |
| Singapore data centres | Over 1.4 gigawatts of capacity, more than 70 facilities (IMDA, 2024) |
| Share of Singapore’s power | About 7%, which an IMDA article says could reach 12% by 2030 |
| New capacity | 80 MW pilot (2023), 200 MW second call (awarded August 2026), under a roadmap for at least 300 MW more |
The WahLiao Verdict
| Is AI boiling the planet? | No. It is a fast-growing slice of a still-small share of world electricity. |
| Is it nothing? | Also no. For a small island with no spare land or local energy, 7% is a lot. |
| Your chats | Negligible next to air-con, a water heater or a car. |
| Per-query figures | Useful, but company-published and narrowly defined. Read the fine print. |
| Where it is decided | Capacity caps, efficiency rules, green power and imports. That is policy. |
How much electricity does AI use, really?
It helps to separate two numbers. The first is the total: every data centre on earth, running search, video, banking, cloud storage and now Super Intelligence. The IEA’s Energy and AI report estimates that total at about 415 TWh in 2024, growing roughly 12% a year since 2017, more than four times faster than electricity demand overall. The United States accounts for about 45% of it, China about 25% and Europe about 15%. The IEA’s base case has data centres reaching about 945 TWh by 2030 and around 1,200 TWh by 2035, but its scenarios for 2035 range widely, from about 700 to 1,700 TWh. That spread is the honest answer: nobody knows exactly how fast demand will grow.
The second number is per question. Here the only figures come from the companies themselves. In August 2025 Google published a method for measuring Gemini, putting the median text prompt at 0.24 watt-hours and 0.26 millilitres of water, roughly five drops, and said the carbon footprint per prompt had fallen sharply over the previous year. In June 2025 Sam Altman, OpenAI’s chief executive, wrote that the average ChatGPT query uses about 0.34 watt-hours, “about what an oven would use in a little over one second”.
The fine print on per-query numbers
Treat those figures as a floor, not the full bill. Google’s estimate covers text prompts only; it leaves out image, video and audio generation, and it excludes the energy used to train models, networking and your own device. OpenAI’s figure came in a blog post without a published method. A long “deep research” task, a generated video or an agent that runs for minutes will use many times more than a quick text answer, and training a frontier model is a large one-off cost spread across all its users. Both are the companies’ own figures. For why training is costly, see How Super Intelligence works.
Singapore’s data-centre story: pause, pilot, roadmap
Singapore is one of Asia’s biggest data-centre hubs, with over 1.4 gigawatts of capacity across more than 70 facilities, according to IMDA. Because data centres take up land and power on a small island with little energy of its own, the government paused new data-centre growth in 2019. The pause was lifted in 2022, replaced by a selective process.
The first step was a pilot call for applications. In July 2023 the Economic Development Board (EDB) and IMDA picked four proposals, from AirTrunk-ByteDance, Equinix, GDS and Microsoft, for about 80 MW of new capacity, chosen for energy efficiency, connectivity and economic value. In May 2024 IMDA published the Green Data Centre Roadmap, which aims to provide at least 300 MW of additional capacity in the near term, with more possible for operators using green energy, and pushes all data centres towards a Power Usage Effectiveness (PUE) of 1.3 or lower. PUE is total facility power divided by the power used by the computers themselves; 1.0 would mean zero overhead.
The second call, launched on 1 December 2025, raised the bar: at least 50% of the capacity powered by green energy such as biomethane, low-carbon ammonia or hydrogen, and a PUE of 1.25 at full load. On 21 August 2026 EDB announced the results: 200 MW split equally between Digital Realty, Equinix, Keppel Data Centres and ST Telemedia Global Data Centres, 50 MW each, chosen from over 20 proposals, all in a low-carbon data-centre park on Jurong Island and all using liquid cooling. For how this fits Singapore’s wider plans, see Singapore’s own Super Intelligence story and the National AI Strategy.
The tropical cooling problem
Singapore is hot and humid all year, which makes cooling expensive. An IMDA article notes that cooling can take 30% to 50% of a data centre’s energy here, compared with 15% to 30% in temperate climates. Two fixes are being pushed. One is running data halls warmer: IMDA says operators can save 2% to 5% of energy for every 1°C rise in operating temperature. The other is liquid cooling, which pipes coolant close to the chips instead of chilling whole rooms; EDB says it can cut energy use at the data-centre level by over 30% compared with traditional air cooling.
Green power, imports and climate targets
Most of Singapore’s electricity is generated from imported natural gas. To decarbonise, the Energy Market Authority (EMA) raised its target for low-carbon electricity imports from 4 GW to 6 GW by 2035, with the first projects aiming to start around 2028. More clean supply makes room for more data centres, but it also has to serve homes, transport and industry.
Singapore has pledged to reach net zero emissions by 2050, with a 2035 target of 45 to 50 million tonnes of CO2-equivalent, down from a 2030 target of around 60 million tonnes. Asked in Parliament about the 300 MW expansion, the government replied in January 2025 that the growth was in line with its projections to peak emissions in 2028. In other words, data-centre growth is being budgeted into the climate plan, not left to the market.
What it means for your household
Put the numbers side by side. A 4-room HDB flat used an average of 380.7 kWh of electricity a month in 2024, according to the Ministry of Sustainability and the Environment. At Google’s median figure, that is the same as roughly 1.6 million Gemini text prompts. Even a heavy user asking a hundred questions a day would use about 24 watt-hours at that rate, about what a 1 kW air-conditioner uses in under two minutes. NEA notes that air-conditioners, fridges, lighting, water heaters and washing machines make up about 80% of a typical household’s electricity use.
On your bill, SP Group reviews tariffs every quarter under EMA guidelines, and the biggest moving part is the cost of imported natural gas. No official source we found links data-centre growth directly to household tariffs, so we would not claim that it raises them. If you want to trim your footprint, the air-con thermostat does far more than cutting back on chatbots. If you use Super Intelligence tools, using them well, with fewer wasted retries, saves your time more than it saves the planet.
AI energy and data centres: FAQ
How much electricity does one ChatGPT question use?
OpenAI’s chief executive has said an average ChatGPT query uses about 0.34 watt-hours. Google puts a median Gemini text prompt at 0.24 watt-hours. Both are company figures for text; images, video and long agent tasks use more.
How much water does AI use per prompt?
Google says a median Gemini text prompt uses 0.26 millilitres, about five drops; Sam Altman put a ChatGPT query at about one fifteenth of a teaspoon. Totals across billions of prompts are larger, and water use varies with the site and cooling method.
What percentage of Singapore’s electricity do data centres use?
About 7%, according to IMDA, which has cited a projection of 12% by 2030.
Why did Singapore stop building data centres?
It paused new data centres from 2019 to 2022 because of their demand for land and electricity. Growth has since resumed in controlled batches: 80 MW in a 2023 pilot and 200 MW in a second call awarded in August 2026, with strict efficiency and green-energy conditions.
Should I stop using AI to save energy?
There is no strong case for it. A day of normal chatbot use is tiny next to your air-conditioner; setting it a degree or two warmer saves far more.
Read next
This page belongs to Super Intelligence. Next, read Super Intelligence to 2030.
Sources checked 1 October 2026: IEA, Energy and AI, executive summary; The New Stack, Google’s median Gemini prompt figures and caveats; Sam Altman, The Gentle Singularity; IMDA, Green Data Centre Roadmap factsheet; IMDA, data-centre cooling and electricity share; EDB, pilot data centre call for application results (2023); EDB, second data centre call results (2026); EDB, DC-CFA2 fact sheet; King & Wood Mallesons, DC-CFA2 launch summary; MTI, written reply on data-centre capacity and emissions; Argus, EMA raises low-carbon import target to 6 GW; EDB, Singapore’s 2035 emissions target; MSE, household electricity consumption by dwelling type; NEA, household electricity consumption profile; SP Group, how tariffs are reviewed. General information. Last updated 1 October 2026.
