AI has become a natural part of everyday life at Grace, and the technology is now used across many parts of our work. At the same time there's an important discussion going on about the energy and infrastructure needed to run AI. That got us thinking about our own usage: how much energy does the AI we use day to day actually require?
It quickly became clear that giving an exact answer is difficult, since AI is built into so many different tools and services. Was it even possible to calculate? Mathematically, of course it was — but the calculation shouldn't be read as a measurement of our actual energy use. It's meant to give us a sense of the order of magnitude and a starting point for understanding our AI use better. We chose to begin with the three AI services we use most: ChatGPT, Claude and Google Gemini.
How would calculated it
We based this on 36 employees and estimate that each person runs around 70 prompts per working day (including work with various agents).
With 220 working days a year, that gives:
36 people × 70 prompts × 220 working days = 554,400 prompts per year
We estimate that these break down roughly as follows:
60% ChatGPT: 332,640 prompts
20% Claude: 110,880 prompts
20% Gemini: 110,880 prompts
The breakdown is our own estimate, not log data from the services. But not all prompts require the same amount of energy. A short text query takes relatively little computation, while longer analyses and reasoning tasks, for instance, can take considerably more.
For simpler queries we've used the following published benchmark figures:
ChatGPT: approximately 0.34 Wh per prompt
Gemini: approximately 0.24 Wh per prompt
Claude: approximately 0.31 Wh per prompt*
*There is no comparable published figure from Anthropic for Claude. We therefore use 0.31 Wh as a general benchmark for modern AI models.
At the same time, our usage consists of much more than short queries. We use AI for research, analysis, longer texts, code, agents and other tasks that require more processing. In our main scenario we therefore assume that 50% of our prompts are simpler and 50% more advanced. For the more advanced ones we use 3.91 Wh per prompt as a standard figure. That value comes from research on more reasoning-heavy AI queries and shouldn't be taken as an exact figure for our prompts.
With these assumptions, the estimated electricity use is:
Simpler prompts: approximately 85 kWh per year
More advanced prompts: approximately 1,084 kWh per year
Total: approximately 1,170 kWh per year
Because the uncertainty is high, we round to approximately 1,200 kWh per year
It's important to stress that this is a scenario, not a measured value. Above all, the result is heavily influenced by our assumption that half of our usage is more advanced. The calculation covers only the operational electricity needed to run the AI services. It doesn't cover the full AI life cycle, such as training the models or manufacturing servers and other hardware.
What did we learn?
The effects of the AI build-out are becoming increasingly clear, and the growing AI infrastructure is driving up demand for electricity. That makes the question relevant to us as users of these services too. We don't own the data centres or train the models, but our usage is part of the demand. So the calculation also makes us think about how we use AI. Can we write better prompts from the start? Give relevant context up front? Avoid unnecessary back-and-forth? And get better at judging when AI actually adds value?
Prompting better isn't only about getting better results. It can also be a way of using the technology and the resources more efficiently. The goal isn't to use as little AI as possible, but to use it more thoughtfully and where it does real good. The most important question raised by this experiment is how large companies are supposed to report the climate impact of generative AI when the computation happens at external suppliers and visibility into actual energy use is limited?
We're happy to share more, and to talk about how we can create digital solutions together.
