What AI costs: computing power, energy, responsibility.

Every request to a language model needs power and computing time. Anyone using AI can factor in efficiency.

The invisible consumption

Training large models and operating data centres require a lot of energy. The International Energy Agency estimated the electricity consumption of data centres, AI and crypto at around 460 terawatt-hours in 2022, just under two percent of global electricity use, and expected a significant increase by 2026.

Efficiency as a competitive factor

Not every task needs the largest model. Smaller, specialised models, reusing results and batching requests reduce cost and consumption. Major providers report progress on energy and water efficiency in their sustainability reports.

What companies can control

• Choose the right model size for the task.
• Cache results instead of recalculating them.
• Bundle usage and avoid duplicate requests.
• Prefer providers with transparent efficiency and origin information.

Green AI as a principle

Research on "Green AI" calls for measuring efficiency on a par with accuracy. For companies this means: a solution is only good once it solves its task with reasonable effort.

The best request is the one that was not needed. The second best is the one answered efficiently.

Conclusion

Using AI has a price, in energy and in money. Those who know it can decide deliberately where AI creates value and where it does not.