"Saying thank you to ChatGPT costs tens of millions of dollars"

In mid-April 2025, an internet user asks on X how much money OpenAI has lost in electricity because of users who say "please" and "thank you" to its models. Sam Altman, the head of OpenAI, replies: "tens of millions of dollars well spent - you never know". The sentence is picked up by the French press from 18 April, then by TechCrunch on 20 April. The question was about "please" and "thank you": a "please" slipped into a question adds only a few tokens to read, whereas a thank-you sent on its own triggers an extra reply.

The reply comes with no calculation, neither the number of thank-yous received, nor the electricity of each, nor the price of a kilowatt-hour. TechCrunch notes that the tone of the reply suggests Altman did not make a precise calculation. The figure is therefore an order of magnitude thrown out in one sentence, which nothing makes it possible to verify.

It touched on a widespread habit. According to a survey by the publisher Future, which publishes TechRadar, picked up by GovTech in February 2025, 67% of Americans who use AI say they are polite to it; the survey does not say how many end their conversations with a thank-you.

In June 2025, Julien Delavande of Hugging Face publishes a measurement of the electricity of a thank-you, taken up in January 2026 in a study co-signed with Régis Pierrard and Sasha Luccioni. They took 10,000 conversations from the UltraChat corpus, dialogues between a user and an assistant, ending in "thank you", and had an open model, Llama 3.1 with 8 billion parameters, reply on an H100 GPU, one request at a time.

They measure on average 0.245 Wh per thank-you and its reply: 0.202 Wh for the GPU, 0.024 Wh for the processor and 0.019 Wh for the memory. That is, they say, a 5 W LED bulb left on for nearly 3 minutes. The measurement counts neither the cooling nor the rest of the data centre; it is made under conditions less efficient than a production service, where the servers handle many requests at once.

The AI Impact Database, which starts from production measurements and published estimates for the major assistants, estimates a thank-you and its short reply at 0.044 Wh, against 0.3 Wh for a quick question and its half-page reply. Our analysis of the actions explains the gap between the 2 values: the length of the conversation reread and that of the reply.

The study breaks the energy down into 2 stages. The model first reads the whole conversation, thank-you included, then writes its reply, one token after another. The authors publish a formula for each: on the GPU, reading costs about 0.00006 Wh per token and writing about 0.002 Wh per token produced, which is 35 times more.

The word "thank you" adds 1 or 2 tokens to read: according to these formulas, about 0.0001 Wh, by our calculation. What costs is what the thank-you triggers: the rereading of the whole conversation, which grows with its length, and above all the model's reply. A reply of 50 tokens costs about 0.1 Wh on its own under the conditions of the study. The authors also note that large models write longer replies, and that model size is the factor that weighs most.

In July 2025, ChatGPT claimed 2.6 billion requests per day, a figure that ADEME repeats in its opinion of July 2026. OpenAI does not publish the share of thank-yous. To bound the question, we assumed that each of these messages was followed by a thank-you, which greatly overstates the result.

With the Database value, 0.044 Wh, these thank-yous would use about 115 MWh per day, against about 780 MWh for the 2.6 billion questions at 0.3 Wh: of the order of 15% of the electricity of the questions themselves. With the 0.34 Wh per request announced by Sam Altman, ADEME counts 884 MWh per day for these requests, the equivalent of the electricity consumption of 145,000 people in France. With the same ratio, 115 MWh per day corresponds to the electricity consumption of about 19,000 people in France. These calculations are our own, on a deliberately high assumption.

With the Database value, even under this high assumption, thank-yous would make up a minority share of the electricity of the questions asked to ChatGPT. The laboratory measurement, made under conditions less efficient than in production, does not lend itself to this calculation. Converted into dollars, the result would depend on the price of electricity paid by OpenAI, which is not public: we do not convert it.