Energy and AI: How Much Electricity Does ChatGPT Use?

Editor

20th February 2026

Public Sector Leaders

By Koketso Mamabolo

The Energy Demands of AI

“For more than 200 years, economic growth was defined by industry capacity. Now for the first time in history, it is measured in computational power and digital capability,” said Dr Sultan Ahmed Al Jaber, the UAE’s Minister of Industry and Advanced Technology and chairman of state-owned renewable energy firm Masdar, speaking at the Abu Dhabi Sustainability Week earlier this month.

While the majority of businesses are still in the experimentation phase, in a global survey by Mckinsey on AI usage 88% of respondents reported that they use AI for at least one business function, up 10% from last year.

General AI use (20% in 2017) has been rising steadily and the same can be said for generative AI, which was at 37% in 2023 and 79% last year. Unsurprisingly, the bigger the company’s revenue the higher the level of AI scaling.

The technology, media and telecommunications, and healthcare sectors report the highest use of AI agents while software engineering, manufacturing, and IT see the most cost benefits. In terms of revenue, sales, strategy and corporate finance, and product or service development have seen the most benefits.

By 2033, UN Trade and Development predicts that the AI market will reach $4.8-trillion, a staggering increase from the 2023 figure of $189-billion. Research from Microsoft’s AI Economy Institute shows that one in six people now use generative AI tools.

The global north is seeing adoption levels double those of the global south but in the second half of 2025 South Africa had the highest adoption figures on the continent with 21.1%, comparable to developed regions such as the EU and North America. Interestingly, in a survey of 3 000 South Africans, UCT found that 37% of respondents had never heard of AI, while 36% knew very little about it and it is reasonable to expect those numbers to change very soon.

Powering The Future

“Artificial intelligence is rewiring every industry, reshaping every sector and resetting expectations for global growth. And while the world is changing around us, one constant remains and that is energy,” said Dr Al Jaber.

“Every algorithm, every data centre, every breakthrough and advanced technology needs power to drive it. Simply put, there is no artificial intelligence without actual energy.” Last year, oil giants Chevron and Exxon Mobil showed great interest in discussions about using natural gas and carbon capture to provide AI data centres with low-carbon energy.

Jeff Gustavson, president of Chevron New Energies, said: “It fits many of our capabilities – natural gas, construction, operations, and being able to provide customers with a low-carbon pathway on power through CCUS [carbon capture utilisation and storage], geothermal, and maybe other technologies.”

The backbone of the AI world are data centres which are facilities where data is stored and processed. These facilities use about 60% of the electricity needed for AI functions with consumption varying from centre to centre. In 2024 it was estimated that data centres accounted for 1.5% of global electricity consumption, growing 12% per year over the last five years.

According to the International Energy Agency (IEA), “the rise of AI is accelerating the deployment of high performance accelerated servers, leading to greater power density in data centres.” IEA estimates that the figure will double by 2030, with a growth rate four times faster than that of any other sector.

While current data centre electricity consumption in Africa is low (the US and China have the highest South Africa stands out with a consumption rate that is predicted to be 15 times larger than the continental average in the next five years. Research suggests that on the global front renewable energy will meet almost half of the additional energy demand from data centres, followed by natural gas and coal, and the portion of nuclear energy set to increase.

Between 2024 and 2030, IEAs data shows a 22% average annual increase for renewable energy usage in data centres. “This growth is primarily driven by the rising deployment of wind and solar PV in power systems across the globe, with some new capacity financed through PPAs with technology companies.”

The US, the dominant player in the AI space, meets most of its data centre energy demand using fossil fuels, mainly natural gas. Data centres in China,the next big player, are getting 70% of their electricity from coal. By 2035, both countries are expected to have to gradually decrease reliance on fossil fuels and rely more on renewables and nuclear power.

The question remains whether or not the US will remain on this trajectory given the “energy abundance” strategy being touted, which involves doubling down on fossil fuels by leaning into them more to power the AI revolution, especially given the fact that the country is home to 41% of the data centres in the world.

This revolution has been rapid, to say the least, and the industry is hungry for data centres to meet the growing demand for AI services. The amount of money being pumped into development is breathtaking. For example, leading firm Open AI and the US government are planning on spending $500-billion dollars on AI. Chinese firm DeepSeek’s hardware uses less than 10% of the energy OpenAI uses.

In a fascinating investigation on the nexus between AI and energy, MIT’s Technology Review said: “The energy resources required to power this artificial-intelligence revolution are staggering, and the world’s biggest fintech companies have made it a top priority to harness even more of that energy, aiming to reshape our energy grids in the process.”

In total, OpenAI, the US government, Apple and Google plan on spending over $1-trillion on AI infrastructure and development, the bulk of which will go to building data centres.

From the mid-2000s to the late 2010s, the electricity consumption by data centres did not see much of an increase due to improving efficiency. Even with the boom in industries relying on cloud-based services, usage remained relatively stable for over a decade.

Then in 2017 things started changing when technology that seemed a distant reality, and merely the stuff of science fiction, took big leaps. The age of AI had begun. In the matter of six years the need for more data centres to match the developments meant energy consumption doubled.

It is widely believed that our AI footprint is as small as it will ever be. Somewhere along the line we reached the point of no return as AI technology leapt into the future, with ever-expanding use cases and increasing personalisation.

The servers in data centres are where the training data is fed into and computation happens. To train GPT-4, OpenAI spent $100-million and used 50 gigawatt hours of electricity which is enough to power to supply 110 000 average South African households for a month or 9 200 homes for an entire year (equivalent to the power needed for towns the size of Stellenbosch and Makhanda).

While training is cost and energy intensive, it’s the queries and the processing needed at the data centres to answer the queries where a significant amount of consumption is happening. This is called inference and it accounts for 80% of AI’s computing power. While not all data centres are being used for AI (industry figures are hard to come by) there is evidence that the number being built specifically for AI inference is growing.

At the data centres, AI models are loaded onto chips called graphics processing units (GPUs). No matter who manufactures them, which is an incredibly small number of firms, what all chips have in common is that need a lot of energy to run without overheating A typical data centre could have thousands of these chips providing information to other chips called CPUs, and all of this requires cooling fans and large amounts of fresh water to keep operations to prevent overheating.

How much energy? It depends…

The question of how much energy a single query uses is difficult to answer. Firstly, there is very little incentive for AI firms to be transparent about the data. Experts remain in the dark on what’s going on behind the scenes for models like OpenAI’s ChatGPT and Anthropic’s Claude.

Secondly, electricity consumption varies from centre to centre – some have more GPUs than others. Thirdly, the number of parameters (adjustable variables) the AI model has an impact on how much electricity is consumed. It’s not clear how many parameters new AI models have, but they range from as small as 3 billion parameters to as much as 600 billion parameters.

The best available info, via research conducted by MIT Technology Review, is sourced from open source AI models such as Meta’s Llama. The result is that estimates are based on very little information, and while it may be possible to calculate how much energy leading chipmaker Nvidia’s H100 GPU uses, consumption from CPUs and other equipment like cooling fans need to be factored in too.

A paper by Microsoft demonstrated that doubling the amount of energy used by the GPU for large language models (ChatGPT, Claude, Gemini, Meta AI etc.) gives a good estimate of how much energy is required for the entire operation. But this is still just an estimate.

To explain their findings, MIT Technology Review used an example where a person asks an AI model 15 questions, makes 10 attempts to create an image, and 3 attempts to create a five second video. The energy used in the inferencing would be enough to run a microwave for over three and half hours.

Given the fact that last year OpenAI was receiving over two billion queries per day, this an alarming amount of energy. In order to function seamlessly, data centres need power 24/7. Research from the Lawrence Berkeley National Laboratory shows that data centres in the US used 200 terawatt-hours of electricity in 2024, enough to supply Thailand electricity for a whole year.

“By 2028, the researchers estimate, the power going to AI-specific purposes will rise between 165 and 326 terawatt-hours per year. That’s more than all electricity currently used by US data centres for all purposes; it’s enough to power 22% of US households each year,” says the MIT Technology Review.

In terms of emissions, that is equivalent to travelling 482.8 billion kilometres or taking 1 600 round trips from Earth to the Sun. The importance of energy for the AI industry is not lost on the tech giants who have been bringing in talent from the energy sector at a noticeably high rate to ensure that they not only consume electricity but influence generation capacity.

Hiring of energy professionals by tech firms increased 34% year-on-year in 2024 and Amazon is leading the pack with 605 new hires over the last five years. Microsoft has recruited 570 more energy professionals in the same time period. This points to an energy “race” that is being directly driven by the enormous and ever-increasing demand for AI services.

Where to from here?

In a white paper published last year, the World Economic Forum outlined four areas stakeholders need to be focusing on when it comes to the impact of AI on energy consumption:

Deploying AI to optimise consumption and reduce waste
Establishing transparency and efficiency frameworks for electricity usage in the industry
Promoting innovation in data centre infrastructure
Driving collaboration between all stakeholders, from AI developers to electricity providers to governments

According to the Data Center Map, there are 243 data centres in Africa.

The nexus between energy and AI is particularly important in South Africa given our struggles with electricity and the fact that the country is home to 60 data centres, the most on the continent by far.

And there are big plans for the digital infrastructure in Africa, from undersea cables to data centres, and even the continent’s first AI factory. Driven by Zimbabwean telecoms billionaire Strive Maswiya’s Cassava Technologies in partnership with Jensen Huang’s Nvidia (Huang is one of the richest people in the world), the factory will be built in South Africa with 3 000 of Nvidia’s GPUs, before the project expands into other African countries with a further 9 000 GPUs.

AI is expected to bring a spark to the African economy, time will tell if the continent can produce the electricity it needs, and if the world can produce all that energy sustainably.