A billion AI queries a day: how much electricity?
Serving a billion queries a day to a large AI model takes about 0.73 GWh of electricity, or 730,000 kWh. That is the estimate in a Microsoft study (Oviedo et al., arXiv preprint 2509.20241, version of 9 June 2026, published in the journal Joule). If one query in ten makes the model reason at length, the total rises to 1.7 GWh. With cautious efficiency gains, it drops back to 0.33 GWh without reasoning and 0.8 GWh with it. For comparison, the study puts a billion web searches at about 0.3 GWh.
A billion AI queries a day use 0.73 GWh according to the Microsoft study: 2.4 times a billion web searches, and roughly what a 30 MW data centre draws in a day. Reasoning matters more than volume: 10% long queries push the total to 1.7 GWh, 2.3 times more. Cautious efficiency gains bring it back to 0.8 GWh.
Figures updated on November 18, 2026
| Scenario | Total per day | Per query (site's arithmetic) | 30 MW centres (site's arithmetic) |
|---|---|---|---|
| Ordinary queries, no efficiency gains | 0.73 GWh | 0.73 Wh | 1.3 |
| Ordinary queries, cautious gains | 0.33 GWh | 0.33 Wh | 0.6 |
| One query in ten with reasoning, no gains | 1.70 GWh | 1.70 Wh | 3.0 |
| One query in ten with reasoning, cautious gains | 0.80 GWh | 0.80 Wh | 1.4 |
| Reference: a billion web searches (Google, 2009) | 0.30 GWh | 0.30 Wh | 0.5 |
Source: Oviedo et al. (Microsoft), arXiv 2509.20241, version 2 of 9 June 2026, Figure 3 and the section "Energy use of a billion queries per day", read on 18 November 2026. Models above 200 billion parameters on H100 cards. The totals are the study's. The "per query" and "30 MW centres" columns are computed by this site: a 30 MW centre at 80% load uses 0.576 GWh a day. These are simulated estimates, not measurements.
How much electricity does a billion AI queries a day take?
About 0.73 GWh a day, according to the Microsoft study. The figure applies to models above 200 billion parameters, served on Nvidia H100 cards, with answers of 300 tokens at the median (a token is a chunk of a word). The study adds a factor of 1.31 for uneven demand, networking and standby machines. Per query, that is 0.73 Wh on average: a division done by this site. With efficiency gains the authors call conservative (hardware, serving software, models), the total falls to 0.33 GWh, less than half.
Why isn't the total simply a billion times 0.31 Wh?
Because 0.31 Wh is a median, not a mean. Half the queries cost less, half cost more, and the longest cost a lot more. So the study draws a billion queries from its distribution and adds up their energy. Naive multiplication would give 0.31 GWh. The published total is 0.73 GWh. Take out the 1.31 factor and about 0.56 Wh per query remains on average. That is this site's inference: the study does not publish the mean. The lesson holds for any large-scale estimate: you multiply a mean, not a median.
What does reasoning change at a billion queries a day?
It more than doubles the bill. The study assumes one query in ten is long, with 5,000 output tokens at the median: what happens when the model "thinks" before answering. That mix takes the total from 0.73 to 1.7 GWh a day, 2.3 times more, for 10% of queries. According to the authors, those long queries then account for about half the tokens served. With the cautious efficiency gains, the same mix comes back down to 0.8 GWh. This site's page "The reasoning tipping point" covers the cost of a single query.
Is a billion AI queries a day a lot of electricity?
It is what one large data centre uses, not a country. The study's yardstick is a 30 MW centre fitted with H100 cards and running at 80% around the clock: it uses just under 0.6 GWh a day. The baseline scenario needs 0.73, about 1.3 such centres. The mix with reasoning needs 1.7, about three centres. Both ratios are computed by this site. The authors push the exercise to ten billion queries a day, the volume of web search: around 1.2 TWh a year, close to 1% of US data centre consumption. In their view, the strain on power grids comes mostly from training models and from new centres clustering in the same places, not from answering queries.
What doesn't the billion-query figure tell you?
It only counts the electricity used to answer in text. Model training, cooling water, CO₂ and building the hardware are left out. So are images, video and agents that chain tools together. It is a simulation built from throughput and power figures, not a meter reading. It describes very large models on H100 cards: a small model or a newer chip changes the result. Finally, every author works for Microsoft, which sells this service. Their per-query median lines up with figures published by Google, Epoch AI and OpenAI, but it is not an independent measurement.
Sources
- Oviedo et al. : « Energy use of AI inference, efficiency pathways, and test-time scaling », Joule (avril 2026, DOI 10.1016/j.joule.2026.102430 ; préprint arXiv 2509.20241) : médiane de 0,31 Wh par requête, intervalle interquartile 0,16 à 0,60 Wh
- OpenAI / TechCrunch : ChatGPT ~2,5 milliards de prompts/jour (2026) ; part de l'IA générative ~80 % (Demandsage / Statista). Calcul : ChatGPT ~900 Md/an ÷ 0,8 ≈ 1 100 Md pour toute l'IA générative.
- Epoch AI : énergie par requête ChatGPT
Frequently asked
How much electricity do a billion ChatGPT prompts a day use?+
The Microsoft study does not model ChatGPT specifically. It picks a billion queries a day because that is the scale of the big assistants: OpenAI reported 2.5 billion ChatGPT queries a day in July 2025. For a billion queries to models above 200 billion parameters, it estimates 0.73 GWh a day, and 1.7 GWh if one in ten makes the model reason. Applying those numbers to ChatGPT would assume its models and hardware look like the study's, which OpenAI does not disclose.
Why does this site's calculator use 0.3 Wh per prompt rather than 0.73 Wh?+
Because it describes an ordinary query, not a whole fleet. The 0.3 Wh matches figures published for a text query (Epoch AI, Google, OpenAI) and the Microsoft study's 0.31 Wh median. The baseline's 0.73 Wh per query also counts the rarer, costlier long queries and the 1.31 factor for networking and standby machines. To estimate a large volume, the mean is what matters. The "Method" page sets out the calculator's assumption.
Do a billion AI queries use more than a billion Google searches?+
Yes, about 2.4 times more in the baseline scenario: 0.73 GWh against about 0.3 GWh. But the search figure is old. It comes from a 2009 Google disclosure (0.3 Wh per search), which the authors call dated and probably too high today. With the cautious efficiency gains, AI drops to 0.33 GWh, level with that reference point. This site's page "ChatGPT vs Google search" compares the two query by query.
Are the Wh per query and equivalent centres in the study?+
No. The study publishes the four daily totals (0.73, 0.33, 1.7 and 0.8 GWh), the 30 MW centre yardstick and the web search reference. The Wh per query, the number of equivalent centres, the ratios and the 0.56 Wh mean are this site's arithmetic on those totals. The 1.2 TWh a year for ten billion queries is in the study, without a stated scenario: this site checked that it matches the scenario with efficiency gains (0.33 GWh × 10 × 365 days).
Knowing what a billion queries cost does not tell you which ones are worth it in your business. The GHIS studio helps choose the AI uses that actually pay off, before rolling them out.