ChatGPT’s ability to respond in a human-like way often leads people to use polite phrases such as “please” and “thank you” when chatting with the AI. While it might seem harmless, all these extra messages add up, quietly increasing the cost of running the service for OpenAI. Earlier this week, a user on X jokingly wondered how much OpenAI spends on electricity just to process polite words, and CEO Sam Altman replied in a lighthearted way that it was “tens of millions of dollars well spent.”
Jokes aside, running large-scale AI models is extremely expensive. Every time someone asks ChatGPT to help with a cover letter or solve a maths problem, it triggers a powerful system of computers behind the scenes. These computers use specialised chips called GPUs, which consume a huge amount of electricity and drive up operational costs.
Recent estimates from tech analyst Dylan Patel suggest that keeping ChatGPT running costs about $700,000 per day. This estimate was based on an older version of the model, GPT-3, so it is likely that the new and more advanced GPT-4o is even more expensive to operate. When new features such as image generation are added and go viral, these costs can climb even higher.
The main reason for these high expenses is the need for immense computing power. Each prompt—whether it is a complex request or a simple thank you—requires the system to process data using energy-intensive hardware. In fact, experts now believe that the cost of running these AI models day-to-day, called inference costs, is even greater than the initial price of training the models.
As AI becomes more popular and widely used, questions are being raised about how these operational costs will be managed and whether users will end up paying more. For now, every polite phrase sent to ChatGPT is a small part of a much bigger bill.
