Everyone who is reading this right now has used Artificial Intelligence (AI) at one point or another. Whether you have used OpenAI’s Chat GPT to write you an essay, crafting a bullet pointed list with DeepSeek AI, or simply reviewing the Google AI summaries with Gemini, AI is almost inevitable at this point.
The convenience is incredible. You not only have access to the entire world’s knowledge, but you can have someone else do the hard work for you and give you everything you could possibly need on the subject completely for free.
Not quite. While AI does make gathering information incredibly convenient, there are a vast number of drawbacks that come with this easy access. Not only is it extremely expensive to maintain, it is also being developed at an alarmingly fast rate until eventually we won’t be able to control it anymore, it takes up a significant amount of rare resources just to keep it running, it is overtaking creative artists with its ability to create art and music at a much faster rate than human artists, and it is beginning to take physical forms in humanoid robots that will eventually outpace humans in every area.
The cost of AI can be seen most directly in its physical cost. With the goal of building up to ten data centers to operate the AI from, the US Government spent an estimated $500 billion dollars on the Stargate Initiative. Apple and Google have also announced multi- billion dollar data center operations, and President Donald J Trump recently signed off on the Gemini Project, aimed at advancing AI using government information and statistics with the help of 17 new data centers.
Even with this money spent, there is still more needed to keep these data centers running. The energy demand for these expanding data centers is on track to quadruple in the next five years, as well as consume more energy for processing data than all energy-intensive goods combined, with the market for AI hardware projected to reach $1 trillion dollars by 2027.
AI’s Development is also an issue. It is happening so rapidly that the CEO of Perplexity, which is an AI LLM that is based on answering questions, admitted that he cannot predict where AI will be in even the next year, only allowing Perplexity to plan for a few months in advance. Even with the power and practice of AI changing so quickly, Perplexity’s CEO said they are still planning to add visual components and go beyond the current limits of search engines.
Not even mother nature is safe, because the resources used in data centers creates a substance known as electric waste, which often contains mercury and lead, very dangerous chemicals. Water, used to cool the servers, is another issue, with an estimated minimum of 4.2 billion cubic meters of water used for global AI data centers in 2027.
How much energy and resources are used to answer my questions, though? Surely five or six Chat GPT questions per day can’t amount to that much.
Well, according to the International Energy Association (IEA), a single Chat-GPT entry is the equivalent of ten separate Google searches. Tack that on to the fact that Google now has Gemini’s automatic AI summaries that don’t have an off switch, and suddenly a single Google search is the same energy cost of eleven separate searches for the same amount of information.
In science terms, a single Chat-GPT prompt uses up about three watt-hours (Wh). A watt-hour is a term used for measuring the amount of energy used as it relates to time, with a single watt being the equivalent of one joule of energy per second.
On its own, that number is insignificant. To put that in perspective, a standard phone battery has roughly ten watt-hours. So if a person uses Chat-GPT for four separate prompts, that is the equivalent of charging a phone battery from zero to one hundred percent.
Recently, OpenAI released a number for roughly how many prompts Chat-GPT receives per day. 2.5 billion. So if a single entry takes up 3 Wh, and there are 2.5 billion entries per day, that means that in a single day, Chat-GPT alone uses up a staggering 7.5 billion Wh of energy. This number, in relation to international usage of electricity, is sur- prisingly insignificant. However, when you multiply this number to account for the entire year, you get the unimaginable figure of two trillion seven hundred thirty seven billion five hundred million.
Once again, however, this number doesn’t really even begin to touch the amount of energy used by countries annually, as each country uses the metric of terawatt-hours (TWh) instead of Wh. a single TWh is the equivalent of one trillion Wh. In 2023, the US used over 4,000 TWh of energy, making the annual energy consumption of Chat-GPT a mere 0.005% of the contribution.
While these numbers are tiny on the international scale, this is for a sole company and only accounting for a singular LLM. According to the International Energy Association, the data centers in which the AI’s are trained and operated require roughly 415 TWh to maintain, which is about 1.5% of the global energy consumption in 2024.
Again, this number looks small, but the increased development and increased usage of AI contributes to a projected double in energy consumption within the next five years.
As these LLMs become more and more popular and accessible, the energy they use will only be going up unless immediate action is taken to reduce their staggering cost.
