One AI answer in a reference 20 MW Malaysian AI hall uses about 0.17 Wh of facility energy, about 0.27 mL of water and about 0.13 g CO₂e on the Peninsular grid. These are illustrative values from Croutons Research’s benchmark of an open-weight model, scaled to MITI’s design PUE of 1.4 and WUE of 2.2 (MITI); they are not measurements of any Malaysian facility. The method is open.
The grid changes the carbon most: the same answer is about 0.034 g on Sarawak’s grid and 0.092 g on Sabah’s (Energy Commission). For comparison, Google reports 0.24 Wh, 0.03 g CO₂e and 0.26 mL for its median Gemini text prompt, on its own fleet and a lower-carbon grid (AI Business).
- About 0.17 Wh
- Facility energy per answer (illustrative)
- About 0.27 mL
- Water per answer at WUE 2.2 (illustrative)
- About 0.13 g CO₂e
- Per answer, Peninsular grid at 0.740 (Energy Commission)
- About 5.9m
- Answers per MWh of facility energy (illustrative)
How much electricity does one AI query use?
About 0.17 Wh of facility energy per answer in our reference hall, or about 5.9 million answers per MWh (1,000,000 Wh ÷ 0.17). That includes cooling and other overhead at a PUE of 1.4; the IT share is 0.17 ÷ 1.4 = about 0.12 Wh. The figure moves a lot with load: at low load, energy per answer was about 3.3 times higher than at high load.
How much water does AI use per prompt in Malaysia?
Water follows the site’s WUE, which is set per kWh of IT energy:
| Cooling assumption | WUE (L/kWh IT) | Water per answer | Calculation |
|---|---|---|---|
| MITI recommended design value | 2.2 | About 0.27 mL | 0.1214 Wh × 2.2 mL/Wh |
| Johor allowable (unit not published) | 1.8 | About 0.22 mL | 0.1214 Wh × 1.8 mL/Wh |
| Dry-cooled hall | Close to 0 | Close to 0 for cooling | No evaporative top-up |
WUE sources: MITI guideline; w.media on Johor. 2.2 L/kWh is the same as 2.2 mL/Wh. This counts on-site cooling water only, not water used to generate the electricity.
How much carbon does one AI query emit on Malaysia’s grids?
| Grid | Emission factor (kg CO₂e/kWh) | Per answer | Per 1,000 answers |
|---|---|---|---|
| Peninsular Malaysia | 0.740 | About 0.13 g | About 126 g |
| Sabah | 0.539 | About 0.092 g | About 92 g |
| Sarawak | 0.199 | About 0.034 g | About 34 g |
Calculation: 0.17 Wh × factor. For example 0.17 × 0.740 = 0.126 g. Factors: Energy Commission, 2024 provisional. Location-based; a renewable-energy contract would change the market-based figure, not this one.
Why do published per-query figures differ so much?
- Model size and answer length. A larger model or a longer answer uses more energy.
- Load. Energy per answer at low load was about 3.3 times that at high load in our benchmark.
- Workload mix. Long summaries were 7% of requests but 37% of energy.
- The site. Its PUE, its WUE and its grid. Google’s Gemini figures come from its own fleet and a lower-carbon grid (AI Business); Malaysian rules assume a WUE of 1.8–2.2.
None of these differences shows up in PUE. The method paper sets out how to measure them per facility.
Reference values from Croutons Research’s controlled benchmark of an open-weight model, scaled to a 20 MW AI hall at PUE 1.4 and WUE 2.2 on the Peninsular grid (section 6 of the method paper). They illustrate the scale; they are not measurements of any Malaysian facility.
What is not public
We could not confirm these from a published source as of 10 November 2026. Do not assume them:
- Measured per-answer energy, water or carbon for any Malaysian facility
- Measured WUE for named Malaysian sites
Croutons Research publishes open data and methods on AI data centres in Malaysia. See the open method for measuring AI work per megawatt.
Sources
- w.media, how Johor deals with thirsty data centres
- Energy Commission, grid emission factors 2022–2024 (provisional)
- MITI, Guidelines for Sustainable Development of Data Centres (PDF)
- AI Business, Google discloses the environmental impact of Gemini prompts
Last checked 10 November 2026