A lighthearted remark by Sam Altman — that polite ChatGPT users are “burning millions of OpenAI dollars” — has gone viral, raising eyebrows over whether manners come with a financial and environmental price tag. Altman’s comment, made in jest on X, referred to how user prompts filled with niceties like “please” and “thank you” add more tokens to ChatGPT queries. These tokens, in turn, consume compute resources and energy.
AI models such as GPT-3.5 Turbo and GPT-4 process billions of prompts daily, each prompt consuming electricity to parse, analyze, and generate outputs. The underlying concern is not the pleasantries themselves, but the cumulative effect: are we overloading data centers in our pursuit of civility?
By the Numbers: Millions Spent or Cents Per Exchange?
Estimates suggest that each token processed by GPT-3.5 Turbo costs OpenAI around $0.0015 per 1,000 input tokens and $0.002 for 1,000 output tokens. On average, polite phrases add 2–4 tokens per prompt — equating to approximately $0.000002 per exchange. Multiplied across 1 billion daily queries, that still only translates to about $400 a day or roughly $146,000 a year — far from Altman’s quip of “tens of millions.”
Furthermore, data from the U.S. Energy Information Administration and Electric Power Research Institute places OpenAI’s estimated electricity costs at $140 million annually — but this includes all prompts, not just the polite ones. Thus, while there is a cost to processing niceties, it’s more symbolic than significant.
AI, Energy, and Etiquette: What This Debate Reveals
At its heart, the debate underscores a larger truth about AI’s hidden infrastructure: vast energy consumption, environmental impact, and operational complexity. OpenAI’s infrastructure burns through 1,058,500,000 kWh of electricity annually to sustain ChatGPT’s responses, translating into over $139 million in energy bills in the U.S. alone.
Altman’s remark — “money well spent” — is as much a cultural commentary as it is an operational insight. In a world where digital interfaces are becoming more human-like, the expectation of kindness isn’t just quaint — it’s training data. Politeness might cost a few fractions of a cent, but it shapes how machines learn to communicate and respond.