The Staggering Energy Bill Behind Every ChatGPT Response
Most people do not think twice when they type a question into ChatGPT. The response arrives in seconds, and that is that. But behind every one of those exchanges is a surprising amount of electricity, and as the platform approaches 900 million weekly active users, the numbers are getting very hard to ignore.
A closer look at the numbers reveals just how much power OpenAI’s ChatGPT uses to answer questions, and what that energy actually costs. The findings are jaw-dropping. Based on an estimated 3.2 billion user queries per day, ChatGPT consumes roughly 60.7 gigawatt-hours of electricity every single day. Over the course of a year, that adds up to more than 22 terawatt-hours, at an annual electricity cost of approximately $3.019 billion.
Why AI Searches Cost So Much More Than Google Searches
When you submit a query to ChatGPT, the process is called inference. The model passes your question through billions of parameters, identifies patterns, and generates a response. It sounds simple, but it is anything but efficient by traditional computing standards.
Researchers at the University of Rhode Island found that the current GPT-5 model, released in August 2025, consumes between 2 and 45 watt-hours per medium-length prompt, with an average of around 18.9 watt-hours. For comparison, a typical Google search uses about 0.3 watt-hours. That means a single ChatGPT response can consume over fifty times more energy than a conventional web search. This energy consumption is roughly six and a half times greater than when we reported on this in our 2024 article: With AI, How Are We Going to Keep the Lights On? As users increasingly turn to AI chatbots for information rather than traditional search engines, that gap has real consequences.
Putting the Numbers in Context
Abstract energy figures can be hard to put into context, so consider a few comparisons. The average U.S. household uses about 10,500 kilowatt-hours of electricity per year. ChatGPT’s annual energy consumption to handle user queries alone is enough to power more than 2.11 million American homes for an entire year. Every day, ChatGPT uses as much electricity as the entire island of Puerto Rico or country of Croatia.
Look at it another way. There are now more than 7 million electric vehicles on U.S. roads. Fully charging all of them once requires around 512.6 million kilowatt-hours. ChatGPT’s yearly energy consumption could accomplish that roughly 36 times over. Or consider smartphones: the electricity ChatGPT uses in a year could fully charge 4.23 billion iPhones every single day for an entire year.
The comparison to video streaming is perhaps the most sobering. One hour of video streaming in Europe uses about 188 watt-hours. ChatGPT, running for just one hour, consumes approximately 2.5 billion watt-hours. To match that, you would need to stream video continuously for more than 1,535 years.
A Growing Platform with Growing Costs
ChatGPT’s energy footprint is not static. The platform has grown at a remarkable pace. In December 2024, OpenAI reported 300 million weekly active users. By February 2026, that figure had reached 900 million, announced alongside a $110 billion private funding round. Monthly active users are now estimated at somewhere between 1.2 and 1.5 billion.
GPT-5 is also more powerful than its predecessors, and significantly more demanding. The model can handle complex reasoning tasks, multimodal inputs, and PhD-level problem solving. These capabilities come with a cost: GPT-5 may consume up to eight times more power than GPT-4. As OpenAI continues to expand features like audio, video, and image generation, that figure is likely to climb further.
Training Costs Are a Separate Challenge Entirely
The energy required to run ChatGPT is separate from the energy required to train it. Training large-language models is extraordinarily intensive. GPT-3, with 175 billion parameters, required roughly 1.3 million kilowatt-hours over 34 days. GPT-4, with over one trillion parameters, consumed around 62 million kilowatt-hours across 100 days of training. That is 48 times more than its predecessor.
For GPT-5, OpenAI has not published training data, making direct comparisons impossible. What researchers generally agree on is that the process involved months of work, multiple retraining phases, and enormous distributed GPU clusters. The architecture itself also evolved significantly. GPT-5 builds on the o1 model, which was designed to think before answering by generating internal reasoning steps before producing a response. This chain-of-thought approach produces stronger results on complex tasks but at a considerably higher energy cost.
The Business Case Still Holds, For Now
Here is an important piece of context. OpenAI currently has around 50 million paying subscribers across its various plans, ranging from the $8 per month Go tier to the $200 per month Pro plan with several other options in between. Based on rough estimates of how those subscribers are distributed across tiers, the company may be generating approximately $1.3 billion in subscription revenue every month, or close to $15.8 billion per year.
The $3.019 billion annual electricity cost, while substantial, represents less than three months of that subscription revenue. OpenAI almost certainly pays below average commercial electricity rates and likely draws on its own renewable energy sources. The actual cost to the company is probably lower than these public estimates suggest.
Still, as the platform grows and models become more capable and more compute-intensive, the energy question is not going away. The industry is in a race to build more efficient systems, but efficiency gains have so far been outpaced by growth in usage. ChatGPT now commands a 60.4% share of AI search traffic globally, and rivals including Google Gemini, Perplexity AI, Microsoft Copilot, Grok, and Claude are all closing in. More competition may mean more efficiency pressure. But it will also mean more inference, more queries, and more electricity.
The next time a response appears in your chat window, it arrived on the back of a lot more than some clever prompting.
Nicholas Ames
With more than 30 years of experience across public accounting, audit innovation, and technology transformation, Nicholas brings deep expertise in applying advanced data solutions to drive performance and change. Most recently, he held key roles at PwC, where he contributed to the firm’s data lake solution, next-generation audit tools, and the rollout of AI agents to enhance service delivery and operational intelligence.