Skip to content
AI Universe
Français

Step 10 of 10

The way back

So, was your first guess right?

Make a guess

Google and Mistral both published the water used by one answer from their assistant. Their numbers are very far apart. Why?

  • One of them has to be wrong.
  • They don't count the same things.
  • You can't measure the water of one answer: those numbers are pulled out of thin air.

Show the answer

The answer: They don't count the same things.

Google counts the water evaporated to cool its data centers. Mistral also counts the water used to make the servers and all along the supply chain. That difference in scope explains a large part of the gap; the size of the model, the length of the answer and the location of the data centers matter too.

What happens

  1. Your question has made the round trip: cut into tokens, sent by radio and fiber, read by a model deep inside a data center.

  2. No answer was waiting for it there. The model worked it out token by token, picking a plausible continuation each time. Only US adults who say AI guesses the next words from patterns it learned, the accurate descriptionSame survey and same question as the database belief; 6% say a human writes the answer behind the scenes, 21% say it follows a scriptSource: Searchlight Institute, Americans Have Mixed Views of AI and an Appetite for Regulation (December 16, 2025)Read at the source, from the organization that produced it. of US adults surveyed describe AI that way.

  3. The trip has a cost, and its number depends on what you count. For the same prompt, Google gets The same prompt, counted the narrow way common estimates do (chips only, most efficient data centers)Active accelerators only, on a sample of prompts that fell in the 10% most efficient data centers; Google presents it as less complete than its own measurementSource: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734) (August 21, 2025)Read at the source, from the organization that produced it. counting only the chips in its most efficient data centers, and Energy of a median text prompt sent to GeminiAI chips, server processor and memory, idle standby machines, cooling and electrical losses of the data center; excludes training, data storage, the network and your phoneSource: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734) (August 21, 2025)Read at the source, from the organization that produced it. with the whole data center.

  4. Mistral, which also counts making the servers, reports Emissions of a 400-token answer from Mistral Large 2 in Le ChatMarginal impact of inference, including server manufacturing and upstream emissions; excludes training (reported separately) and users' devicesSource: Mistral AI, Our contribution to a global environmental standard for AI (July 22, 2025)Read at the source, from the organization that produced it. and Water for a 400-token answer from Mistral Large 2 in Le ChatMarginal impact of inference over the life cycle, including server manufacturing and the upstream chain; excludes training (reported separately) and users' devicesSource: Mistral AI, Our contribution to a global environmental standard for AI (July 22, 2025)Read at the source, from the organization that produced it. for one answer from its assistant, Le Chat.

  5. Google, counting only evaporated water, gets Water used for a median text prompt sent to GeminiWater evaporated to cool Google's data centers; excludes the water of power plants and of chip manufacturingSource: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734) (August 21, 2025)Read at the source, from the organization that produced it.. The gap comes first from scope, and scope is what the ticket below lets you change.

Your turn

Go back to your guess from the start and see whether it holds up. Then read the ticket for your trip: move the data center farther away, widen the scope, and watch the numbers change.

And now, how does the AI come up with its answer?

See the answer

The accurate description: "It guesses the next words from patterns it learned".

The model writes its answer token by token, working out each time the most plausible next one from the patterns it learned in training. It doesn't look the answer up in a database or follow a script, and nobody writes in its place. An assistant connected to a search tool also reads excerpts found on the web, then answers the same way.

What American adults answered

  • It looks up the exact answer in a database45%
  • It follows a script of answers written in advance21%
  • It guesses the next words from patterns it learned28%
  • A human writes the answer behind the scenes6%

Tavern Research survey for Searchlight Institute, online panels, August 1 to 6, 2025, 2,301 US adults, margin of error 3 points; a question with four answers to pick from.

Source: Searchlight Institute, Americans Have Mixed Views of AI and an Appetite for Regulation, published December 16, 2025, read at the source.

Why is the sky blue?

Tokens

61 tokens

Your question is 6 tokens. With the site's instructions and the conversation tags, the model reads 36, then writes 25 to answer.

Computed by the open model Qwen3-4B, revision 1cfa9a7.

Distance

500 km

We don't know where your question would be computed: this distance is a guess, and you set it.

Guess: 500 km to the data center. With JavaScript, you can change it.

Travel time

at least 21 to 36 ms round trip

15 to 30 ms round trip for the 4G network, then 6 ms for 500 km of fiber. Fiber doesn't run in a straight line, so this is a minimum.

The time the model takes to compute its answer isn't counted: we have no published figure to measure it.

Source: Arcep, La qualité des services mobiles, published September 2026, read at the source.

Energy and water

What we count

The chips alone

Energy

0.14 Wh

For a median text prompt sent to Gemini, measured by Google. That isn't the model on this site, which we have no measurement for.

Without the server processor and memory (0.06 Wh), idle machines (0.02 Wh) or the data center (0.02 Wh); the four parts add up to 0.24 Wh.

Counted the narrow way common estimates do: 0.1 Wh.

Source: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734), published August 21, 2025, read at the source.

Water

No published water figure for this scope.

The server

Energy

No published figure for the server alone. Google gives the parts of its measurement:

Without the server processor and memory (0.06 Wh), idle machines (0.02 Wh) or the data center (0.02 Wh); the four parts add up to 0.24 Wh.

Source: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734), published August 21, 2025, read at the source.

Water

No published water figure for this scope.

The whole data center

Energy

0.24 Wh

For a median text prompt sent to Gemini, measured by Google. That isn't the model on this site, which we have no measurement for.

AI chips, server processor and memory, idle standby machines, cooling and electrical losses of the data center; excludes training, data storage, the network and your phone.

Source: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734), published August 21, 2025, read at the source.

Water

0.26 mL

Water evaporated to cool Google's data centers; excludes the water of power plants and of chip manufacturing.

Source: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734), published August 21, 2025, read at the source.

The life cycle, manufacturing included

Energy

No figure in watt-hours for this scope: Mistral publishes emissions and water.

Emissions

around 0.07 g CO2e

Mistral publishes 1.14 g CO2e for a 400-token answer from its Mistral Large 2 model. Scaled to the 25 tokens of this answer, assuming the impact follows its length: an order of magnitude, not a measurement.

Marginal impact of inference, including server manufacturing and upstream emissions; excludes training (reported separately) and users' devices.

Source: Mistral AI, Our contribution to a global environmental standard for AI, published July 22, 2025, read at the source.

Water

around 3 mL

Mistral publishes 45 mL for a 400-token answer from its Mistral Large 2 model. Scaled to the 25 tokens of this answer, assuming the impact follows its length: an order of magnitude, not a measurement.

Marginal impact of inference over the life cycle, including server manufacturing and the upstream chain; excludes training (reported separately) and users' devices.

Source: Mistral AI, Our contribution to a global environmental standard for AI, published July 22, 2025, read at the source.

Three questions to finish

Question 1

Why can an AI be wrong in a perfectly confident tone?

  • It writes the most plausible continuation, which isn't always true.
  • It lies on purpose when it doesn't know.
  • Its database contains mistakes.

Show the answer

The answer: It writes the most plausible continuation, which isn't always true.

The model has no database to search. It computes, token by token, whatever looks most like a good answer, and a plausible invention comes out with the same confidence as a fact.

Question 2

When you type your question, what does the model receive?

  • The letters, one at a time.
  • Whole words, as in a dictionary.
  • A string of tokens: numbered pieces of text.

Show the answer

The answer: A string of tokens: numbered pieces of text.

Your question is first cut into tokens, and the model only gets their numbers. The same sentence often takes more of them in French than in English.

Question 3

A company announces how much energy one question to its assistant uses. What should you look at first?

  • Whether the number is small.
  • What it counts: the chips alone, the whole data center, manufacturing.
  • Nothing: a number the company published is good enough.

Show the answer

The answer: What it counts: the chips alone, the whole data center, manufacturing.

The same calculation, counted two ways, can give very different numbers. Before comparing, look at the scope, then at who did the measuring.

Takeaway

The model works out its answer token by token, from patterns it learned. It doesn't look anything up, and it can be wrong with total confidence. An answer also has a cost, and the number depends on what you count.

Go further

What the ticket leaves out

None of the numbers on the ticket count training the model, making your phone, or the network's energy for your question alone. For the network, the sources read for this site give no reliable measurement at the scale of one question. For training, the companies that put a number on it publish it separately, as Mistral does.

Less per answer, more answers

Comparing two measurements taken a year apart, Google found that a text prompt to Gemini used How many times less energy the median text prompt sent to Gemini used, May 2024 to May 2025Same measurement as the 0.24 Wh, made by Google and not checked by a third party; the drop comes from the models (23 times) and better use of the machines (1.4 times)Source: Google, Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734) (August 21, 2025)Read at the source, from the organization that produced it. less energy. Most of the gain came from the models themselves, the rest from making better use of the machines.

Several techniques bring down the cost of an answer. The inference server batches questions from several people on the same chip: the authors of vLLM, one of these programs, get Throughput gain of an inference server that batches requests more efficiently, at the same response delayMeasured by the vLLM authors on open models, against the systems of 2023; the gain grows with longer texts and larger modelsSource: Kwon et al., Efficient Memory Management for Large Language Model Serving with PagedAttention (arXiv 2309.06180) (September 12, 2023)Read at the source, from the organization that produced it. that way at the same delay. Quantization stores the model's numbers in fewer bits, which takes less memory and less computing. With speculative decoding, a small model guesses several tokens ahead and the big one checks them in a single pass, without changing the answer.

Meanwhile, Google's products were handling more and more tokens each month: Growth in the number of tokens processed each month across Google's products and APIs, in the year to May 2025From 9.7 trillion to over 480 trillion tokens a month; all of Google's products, not just the Gemini app; announced by Google, with no breakdown by modelSource: Google, Google I/O 2025: Sundar Pichai's opening keynote (May 20, 2025)Read at the source, from the organization that produced it. as many at the end of that year as at the start. The two numbers don't cover the same requests. Taken together, they show that each answer costs less and that there are many more of them. The total for data centers keeps rising (see "The data center").

Trips to come

This one followed a text question. Other trips are planned along the same lines: the chip factory, the model's schooling, agents and their tools, AI in society, AI beyond text.

Sources