AI Energy and Water Use, and Singapore's Data Centres

By the WahLiao desk · Last verified 3 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, as far as the published figures show, 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. That’s why the government paused new ones from 2019 to 2022, and now lets them grow only in small, green-conditioned batches.

If you remember one thing, make it this: your own chatbot use is a rounding error next to your air-con. The bigger levers are how data centres are built, cooled and powered, and those are set by policy and operators, not by you skipping a question. Here’s what we found, and if we’ve got something wrong, tell us and we’ll fix it.

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?	Not on the IEA’s figures. It’s 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	Tiny next to air-con, a water heater or a car.
Per-query figures	Useful, but published by the companies and narrowly defined. Worth reading the fine print.
Where it’s decided	Capacity caps, efficiency rules, green power and imports. That’s policy.

How much electricity does AI use, really?

It helps to keep two numbers apart. 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%. In the IEA’s base case, data centres reach 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 tells you something useful: nobody knows exactly how fast demand will grow, and these are projections rather than certainties.

The second number is per question, and here the only figures we could find come from the companies themselves. In August 2025 Google published a method for measuring Gemini. It put 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

It’s best to 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, and we haven’t seen them independently checked. 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. Data centres take up land and power, and we’re a small island with little energy of our own, so the government paused new data-centre growth in 2019. The pause was lifted in 2022 and 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. It 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. In plain terms, 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 four are to be sited at a sustainable data-centre park on Jurong Island, and all will use 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

As anyone who has waited at a bus stop at noon knows, Singapore is hot and humid all year, and that 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 cooling 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. Data-centre growth has to fit inside those targets, which is why new capacity is released in capped, green-conditioned batches rather than left to the market.

What it means for your household

Let’s 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’s 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. The Big Buy explains the tariff, and when switching retailer is worth it, in your electricity bill. No official source we found links data-centre growth directly to household tariffs, so we won’t claim that it raises them. If you’d like to trim your footprint, the air-con thermostat does far more than cutting back on chatbots. And 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 how it’s cooled.

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 how much land and electricity they need. 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?

On the figures we’ve seen, there’s no strong case for it. A day of normal chatbot use is tiny next to your air-con; setting it a degree or two warmer saves far more.