About
An informative site to put AI's consumption back into perspective, with the numbers to back it up.
Our goal
HowManyPrompts is purely informative: it puts the real consumption of a generative-AI prompt back into perspective. We set it against everyday actions, and even a simple Google search, to give concrete orders of magnitude rather than gut feelings.
The goal is neither to demonize AI nor to oversell it. Large language models (LLMs) are tools that save time, already woven into hundreds of millions of people's daily lives; like any technology, they use resources. Understanding that footprint, without doom or denial, is what lets you use them clear-headed.
Where the numbers come from
All the data comes from public, verified sources: public agencies (ADEME, US EPA / DOE / EIA, EEA), research institutes and scientific papers (Water Footprint Network, Our World in Data after Poore & Nemecek, arXiv), the International Energy Agency (IEA), and official figures from AI companies (Google, OpenAI, Epoch AI).
When sources diverge, we keep a credible median and show the range: honest orders of magnitude, never false precision. Every value is dated and linked to its source.
Putting a prompt in perspective
An average text prompt uses roughly 0.3 Wh, 0.3 mL of water and 0.2 g of CO₂, about the energy of a Google search, and a few drops of water. On its own, it's tiny.
- 1 AI prompt≈ Google search
- Kettle (1 cup)22 Wh · ~73×
- Vacuum 10 min90 Wh · ~300×
- Oven for 1 h2,100 Wh · ~7,000×
But the unit isn't the whole story: multiplied by billions of requests a day, the consumption becomes significant at a global scale. Both are true at once, and that's exactly what this site helps you picture.
The debate, honestly: two readings
On AI's impact, two camps coexist, and each has solid arguments. We lay them side by side.
- The unit impact is small: ~0.3 Wh, comparable to a Google search (Google, OpenAI, Epoch AI).
- Efficiency is improving: newer models use less per request (Google: 0.24 Wh).
- Concrete benefits: productivity, medical-diagnosis support, climate research, accessibility.
- Compared with streaming, meat, cars or flights, AI stays modest per person.
- The scale effect is real: ~2.5 billion prompts/day; data centers already account for ~1 to 1.5% of global electricity and are growing fast (IEA).
- Reasoning models, and especially video, cost far more (up to ×1,000).
- Water and local water stress around data centers (UC Riverside, Water Footprint Network).
- Jobs, bias, privacy, concentration of power: up to ~40% of worked hours could be affected (WEF, ILO, Stanford HAI).
AI and jobs: a transformation, not a novelty
As with every major revolution, industry then the web, some jobs change or disappear while others emerge. Generative AI follows the same path. Studies (WEF “Future of Jobs”, OECD, PwC) point mostly to a transformation of tasks rather than outright replacement: the tool augments the expert more than it replaces them, wherever judgment, responsibility and human context stay decisive.
Our stance
Neither starry-eyed technophiles nor technophobes. We believe that staying aware of how AI evolves, how it's used and what it costs is the best way to benefit from it, without naivety or fear. This site gives the numbers; everyone draws their own conclusions.
Transparency
Building this site (AI-assisted research and design) cost a modest order of magnitude, the equivalent of a few hours of TV. We show it too, for consistency: we don't ask of others a transparency we don't apply to ourselves.