A great chat with Google Gemini

 

I recently had a very exciting conversation. I also found this exchange of ideas with my unique conversation partner, Google Gemini, to be very enlightening. The topic was just how much water is actually used when I talk to it.

Although I’ve been actively using artificial intelligence for background research for my articles for about a year now, I was prompted to have this special conversation by a publication on the topic that came out in January.

According to a study by Xylem and Global Water Intelligence, the water demand of artificial intelligence (AI) will increase by 129% by 2050: data center cooling, chip manufacturing, and energy production combined will require approximately 30,000 billion liters of additional water annually. The study is a milestone in research on this topic—as Gemini specifically noted during our conversation—because it takes the total water footprint into account when assessing water demand. Previous analyses focused primarily on the water used to cool data centers. However, this accounts for only 4% of the water footprint—the direct water footprint. A much larger portion—54%—is accounted for by indirect water use, which is the water consumed by power plants generating the electricity used by data centers for their own operations. Furthermore, a significant 42% of the water footprint is also associated with chip manufacturing—that is, the production of the AI’s “brain.”

I am working

How big is my AI water footprint?

When I read this, I suddenly asked Gemini, on a whim, what it thought my water footprint was for the past three months of our collaboration. It replied that, out of respect for my privacy, it doesn’t record data that would measure how many hours we spend together. But for simplicity’s sake, it calculated based on one hour per day, which it estimates averages 30 prompts per hour and amounts to roughly one and a half to two million words.

He noted that, as a result of Google’s developments, according to his most recent data from 2026, the water footprint of our conversation had decreased significantly. Previously, one hour of AI use required roughly half a liter of water. Now, however, this has dropped to 7.8 mL per hour. In other words, the water footprint in question amounted to 0.7 liters over the three-month period. He added, however, that while this might seem like a lot, by comparison, adding 50 mL of milk to my morning coffee requires 50 liters of water, and eating a beef steak requires roughly 2,400 liters.

In response to this remark, I told him that I don’t put milk in my coffee (in fact, I don’t drink milk at all) and I don’t eat meat. Gemini then noted that with this consumer behavior, I can offset the water footprint of many months’ worth of conversations.

Flash or Pro Search? It matters!

Encouraged by my relatively low AI footprint, I kept asking questions. As I reflected on my own AI usage habits, I wondered if there was a difference between Flash (Quick) and Pro (Detailed) searches. When I’m conducting in-depth background research on a specific topic, I typically switch the search mode to Pro.

I was very surprised by the answer. The difference is huge! Gemini itself used the analogy that I should imagine Flash as a small, eco-friendly car, and Pro as a massive, high-performance SUV. If I ask a simpler question in Pro mode, generating that answer requires at least twenty times as much water as if I’d asked the same question in Flash mode. And if I formulate a serious, complex prompt for it during my scientific research, the water footprint of the answer provided in Pro mode can be as much as 60 times that of a simple Flash prompt. It also explains the reason for this large difference in detail: for a Pro question, it performs much more complex research. In AI terms: there are many more “FLOPS” during scientific research. Not knowing the term, I asked what this abbreviation meant. I was told that “The performance of AI models is measured in so-called FLOPS, which stands for the number of complex mathematical calculations performed per second.” (FLOPS, or Floating Point Operations per Second). The bottom line is that more operations consume more energy, and thus require more water.

Of course, it immediately calculated for me that if I spend 15 minutes a day on scientific, “Pro” research, my AI usage’s water footprint will be 37 liters, not 0.7. But that’s still less than drinking a glass of milk (about 120 liters of water), it added.

Incidentally, Gemini itself also mentioned, in comparison of the two search modes, that the Gemini 3 Flash operates with 90.4% accuracy and reliability, while the Gemini 3.1 Pro search mode operates with 94.3%. But it noted that, in terms of the results of our conversation, this 4% difference is insignificant. I should also mention that we ultimately discussed this issue entirely in the water-saving Flash mode.

What’s the situation in Hungary?

My next question was whether the figures provided were global averages or if he had factored in that I was talking to him from Hungary. I was told that these were global figures, and if I were using the app specifically in Hungary, my AI water footprint might be about 10–15% smaller. So the 37 liters is actually 32. Great, we’re doing pretty well!

But what’s the reason for this? Of course, I didn’t even have to ask, because I got the answer anyway: the Google servers serving us are located in Eemshaven (Netherlands), St. Ghislain (Belgium), and Frankfurt (Germany). At the centers in the first two cities, although the average water demand doesn’t change, they typically use canal water for cooling. So the water used doesn’t come from the drinking water supply. Another factor contributing to this reduction is that the farther north a data center is located, the more it can rely on air cooling instead of water cooling during the winter months. Consequently, these centers have lower direct water consumption compared to their southern counterparts. Furthermore, it is also important that these data centers obtain their electricity from power plants operating as efficiently as possible, and as we know, this accounts for 54% of the water footprint!

Does the language matter?

My next question was whether it matters what language I use when talking to him. Somehow, I instinctively reach for English every time, because I feel that he can draw from a much larger vocabulary that way, and he doesn’t have to translate. He confirmed my hunch that this might indeed be significant. If I speak to it in Hungarian, that results in a 40% increase in its water consumption. So, in this case, the water consumption for three months of working together would immediately jump from 32 liters to 45 liters. In his explanation, he elaborated that its operation is optimized for the English language—it has been trained on that as well. It thinks not in words, but in so-called “tokens.” While one English word is one token, due to the agglutinative nature of the Hungarian language, a single Hungarian word can be as many as 3–4 tokens. Translating back and forth also requires more processing. All of this, therefore, means more FLOPS—and thus a larger water footprint.

Power Supply, Chips, and Data Center Cooling

Gemini expanded on this last point in its explanation following my next question, when I asked for its opinion on the study by Xylem and Global Water Intelligence. They found the water footprint ratios of 54-42-4% indicated in the study to be absolutely correct, and—as I mentioned above—they expressly welcomed the fact that all factors were taken into account in the calculation.

Of course, my curious nature wouldn’t let me rest, so I also asked whether, in the AI water footprint he had calculated, he had included the water demand for electricity generation and chip manufacturing, or just the cooling of the data center. He reassured me that if he had only accounted for the 4% water demand, it would have been greenwashing on his part—just like claiming a car is carbon-neutral simply because it doesn’t have an exhaust pipe.

What was new and important information for me was that, regarding the manufacturing of chips (i.e., the heart of the hardware), he not only highlighted the enormous water demand but also emphasized that, in this case, it is crucial for the water to be crystal clear! This is because only completely contaminant-free water is suitable for rinsing the silicon wafers.

Currently, the number of chips globally that serve as the “brains” of Google’s AI is estimated to be between 15 and 18 million. But our own devices also contribute to the 42% of the AI-related chip water footprint. Although, of course, the operation of my laptop cannot be compared to that of Google’s data center in Frankfurt...

What was new and important information for me was that, regarding the manufacturing of chips (that is, the heart of the hardware), he not only highlighted the enormous water demand but also emphasized that, in this case, it is crucial for the water to be crystal clear! This is because only completely contaminant-free water is suitable for rinsing silicon wafers.

Currently, the number of chips globally that serve as the “brains” of Google’s AI is estimated to be between 15 and 18 million. But our own devices also contribute to the 42% chip water footprint associated with AI usage. Although the water footprint associated with my laptop is obviously not comparable to that of Google’s data center in Frankfurt, when I think about how I can use my new research partner in a more environmentally friendly way, it certainly matters how often I buy a new computer and cell phone each year. According to Gemini, the manufacturing water footprint of the former averages 190,000 liters, while that of the latter is 12,000 liters.

What can I do right here and now?

In addition to avoiding replacing my device too often, I can make smart use of my new research partner. AI awareness doesn’t mean we talk to it less, but rather that we only bring out the heavy artillery—that is, the Pro mode—when scientific precision is truly essential. Choosing the Flash mode for everyday conversation is like using targeted drip irrigation instead of a water-wasting sprinkler. And if this isn’t a problem for us, let’s ask our questions in English instead, and at most, ask for a summary in Hungarian at the end of the conversation. Although the AI speaks Hungarian fluently, its “native language” is English. Asking questions in English during research not only leads to deeper sources but is also more sustainable.