Third part: fuel and emissions
Now that we have an estimate of the total energy required to perform an AI model to produce text, images and videos, we can determine what it means in terms of emissions that cause climate change.
First, an endeavor data center is not necessarily a bad thing. If all the data centers were connected to solar panels and only worked when the sun shone, the world would speak much less about the EIA energy consumption. This is not the case. Most electrical networks around the world still depend strongly on fossil fuels. Electricity consumption is therefore delivered with an attached climate toll.
“AI data centers need constant power, 24-7, 365 days a year,” explains Rahul Mewawalla, CEO of Mawson Infrastructure Group, which builds and maintains high-energy data centers that support AI.
This means that data centers cannot count on intermittent technologies such as wind and solar energy, and on average, they tend to use dirtier electricity. A pre -impregnated study of the Cha Chan School of Public Health of Harvard revealed that the carbon intensity of the electricity used by the data centers was 48% higher than the American average. Part of the reason is that the data centers are currently grouped in places that have dirty grids on average, such as the heavy charcoal grid in the Atlantic region which includes Virginia, Virginia-Western and Pennsylvania. They also operate constantly, including when cleaner sources may not be available.
Data centers cannot count on intermittent technologies such as wind and solar energy, and on average, they tend to use dirty electricity.
Technological companies like Meta, Amazon and Google responded to this fossil fuel problem by announcing objectives to use more nuclear energy. These three have joined a commitment to triple global nuclear capacity by 2025. But today, nuclear energy represents only 20% of electricity supply in the United States and feeds a fraction of operations of AI data centers – natural gas represents more than half of the electricity produced in VirginiaWho has more data centers than any other American state, for example. In addition, new nuclear operations will take years, perhaps decades, to materialize.
In 2024, fossil fuels, including natural gas and coal, were just under 60% of electricity supply in the United States. Nuclear has represented around 20% and a mixture of renewable energy represented most of the remaining 20%.
The gaps in the power supply, combined with the rush towards the construction of data centers to supply AI, often mean short -view energy plans. In April, the Center X Supercominuting of Elon Musk near Memphis was found, via satellite imaging, to use dozens of methane gas generators who, according to the south of the environment unrealized by energy regulators to complete the energy of the network and violate the clean Air Act.
The key measurement used to quantify the emissions of these data centers is called carbon intensity: how many grams of carbon dioxide emissions are produced for each kilowatt hour of electricity consumed. The nailing of the carbon intensity of a given grid requires understanding the emissions produced by each individual power plant in operation, as well as the amount of energy contributes to the grid at any time. Public services, government agencies and researchers use estimates of average emissions, as well as real -time measures, to follow the pollution of power plants.
This intensity varies considerably from one region to another. The American grid is fragmented and the mixtures of coal, gas, renewable energies or nuclear vary considerably. California’s grid is much cleaner than that of Virginia-Western, for example.
The time of day also counts. For example, the April 2024 data show that the California network can swing less than 70 grams per kilowatt hour in the afternoon, when there is a lot of solar energy available for more than 300 grams per kilowatt-hour in the middle of the night.
This variability means that the same activity can have very different climatic impacts, depending on your location and the time you spend a request. Take this charity marathon runner, for example. The text, image and video responses they asked for up to 2.9 kilowatt hours of electricity. In California, the generation of this amount of electricity would produce an average of around 650 grams of carbon dioxide pollution. But generating this electricity in Western Virginia could inflate the total to more than 1,150 grams.
AI at the corner of the street
What we have seen so far is that the energy necessary to respond to a request can be relatively low, but it can vary a lot, depending on the type of request and the model used. The emissions associated with the amount of electricity given will also depend on the place and the moment when a request is managed. But what is it added to?
Chatgpt is now estimated To be the fifth most visited website in the world, just after Instagram and ahead of X. In December, Openai said that Chatgpt receives 1 billion messages every day, and after the company launched a new image generator in March, it said That people used it to generate 78 million images per day, from studio style portraits Ghibli to photos of themselves as Barbie dolls.
Given the direction that AI is directed – more personalized, capable of reasoning and solving complex problems on our behalf, and wherever we look at – it is likely that our IA footprint today is the smallest it will ever be.
You can make very difficult mathematics to estimate the energy impact. In February, the Epoch AI IA research firm published an estimate of the amount of energy used for a single Chatgpt request – an estimate which, as discussed, made many hypotheses which cannot be verified. However, they calculated about 0.3 watthers, or 1,080 joules, per message. This is between our estimates for the smallest and most important Meta Llama models (and the experts we have consulted say that if anything, the actual number is probably higher, not lower).
A billion of these every day for a year would mean more than 109 gigawatt hours of electricity, enough to supply 10,400 American houses for a year. If we add images and imagine that everyone’s generation requires as much energy as with our high quality images models, this would mean 35 additional gigawattheures, enough to feed 3,300 houses for a year. This is in addition to the energy requirements of other Openai products, such as video generators, and that for all other IA companies and startups.
But here is the problem: these estimates do not capture the close future of how we will use AI. In this future, we will not simply ping AI models with a question or two throughout the day, nor will not make them generate a photo. Instead, the main laboratories run to a world where AI “agents” perform tasks for us without our supervision of their movements. We are going to talk to models in vocal mode, discuss with companions for 2 hours a dayAnd point our phone cameras to our environment in video mode. We are going to give complex tasks to the so-called “reasoning models” that operate logically through the tasks but have been found to demand 43 times more energy for simple problems, or “deep research” models that spend hours creating relationships for us. We will have AI models which are “personalized” by forming our data and our preferences.
This future is on our doors: Openai will offer agents for $ 20,000 per month and will use reasoning capacities in all its models in the future, and Deepseek Catapulted “Chain of Thought” reasoning in the dominant current with a model that often generates nine pages of text for each response. AI models are added to everything, from customer service telephone lines to doctor’s offices, quickly increasing the share of the national energy consumption.
“The few precious numbers that we have can lose a tiny ribbon of light on the place where we are now, but all bets are extinct in the years to come,” explains Luccioni.
Each researcher to whom we have spoken said that we cannot understand the energy requirements of this future by simply extrapolating the energy used in IA requests today. And indeed, the movements by leading the companies of AI to trigger nuclear power plants and to create data centers of an unprecedented scale suggest that their vision of the future would consume much more energy than a large number of these individual requests.
“The few precious numbers that we have can lose a tiny ribbon of light on the place where we are now, but all bets are extinct in the years to come,” explains Luccioni. “The generative tools of AI are done practically in our throat and it becomes more and more difficult to withdraw, or to make enlightened choices with regard to energy and the climate.”
To understand the power of this AI revolution if necessary, and where it will come from, we must read between the lines.