Was ein KI-Prompt uns wirklich kostet | Studio Q

Summary

The video examines how much water AI uses, why data centers consume it, and how reliably that impact can be measured.

Why data centers use water

AI chips generate heat while processing prompts. Some data centers use outside air for cooling, but in hot weather water may be sprayed into the air to help cool equipment. The video says that up to 90% of this water can evaporate, making it temporarily unavailable to local users.

Local water impacts

The report focuses on a planned Amazon data center in Baruth/Mark, Brandenburg, a region already facing water concerns. It says beverage companies Red Bull and Rauch account for 92% of the town’s released groundwater, leaving 8% for local residents.

Authorities say groundwater supplies are sufficient, but Amazon did not provide figures for the planned facility’s water use. Amazon said its German data centers use air cooling for 95% of the year.

Gaps in public data

Operators are not required to publish planned water use before construction. Germany’s register collects consumption figures retrospectively, after a data center has operated for a year. The report also cites the EU Commission’s estimate that about a quarter of German data centers are unregistered.

As a result, the current and future impact of data centers is difficult to assess publicly.

Estimated water use per prompt

Prompt-level figures are model estimates, not direct measurements. Researchers estimate water use based on factors such as the AI model, processor, response-generation time, and water associated with producing the electricity. The study compared models using prompts of different lengths.

For a short prompt of about 75 words, the report gives these estimates:

• GPT-4.1: 4 ml
• Claude model named in the subtitles as “2.3.7”: 5 ml
• DeepSeek V3: 19.3 ml

The report compares this range with the water use associated with roughly 10–40 Google searches. Longer prompts can substantially increase the estimate: tasks involving thousands of words may reach a small-cup scale, while generating an anime-style video of a dog may require several liters.

Scale and context

Small per-prompt estimates can add up as AI services and data centers expand. For context, the video cites annual water use of about 25 million liters for an average German data center, compared with 45 billion liters for a lignite power plant near Jänschwalde.

These comparisons do not eliminate the need to consider local water scarcity.

Video type and sources

The video is an explanatory report and analysis, not a product review or how-to guide.

The report is narrated by an unnamed presenter or reporter. Other sources include Amazon, local and regional authorities, researchers whose study is used to estimate prompt-level water consumption, and the EU Commission’s estimate on data-center registration.