AI's Electricity Bill and the Carbon Footprint of Computing Power
In the early hours of the morning, at the intelligent computing center in Loudoun County, Virginia, the cool white LED lights cast regular light spots between the rows of cabinets, and the dark blue servers emit harsh noises at a constant temperature of 22°C, methodically maintaining the order of the digital world.
This region, known as the world's "data center corridor", gathers the world's most dense cluster of hyperscale data centers, where the intelligent computing centers of tech giants such as Google, Amazon, and Microsoft are all located. These facilities support the high-speed operation of global digital projects.
At 9 a.m. in Beijing, intern Xiao Tao submits a search request for "2025 Global AI Industry Report" on the computer screen. The moment he presses the Enter key, a cross-continental power consumption process quietly starts. Meanwhile, thousands of miles away in London, freelance writer Emma is talking to ChatGPT, asking about "the impact of 19th-century European railway construction on the international landscape". Every "thinking" loading action in the dialog box is converted into heat of server chips in the distant data center.
Behind these two seemingly insignificant clicks lies a shocking set of energy consumption comparisons. When you type a keyword into a search engine, the movement of your fingertips consumes only 0.3 Wh of electricity, which can light up a household LED lamp for two minutes. However, when you turn to ChatGPT to start an in-depth conversation, the energy consumption jumps instantly: a question-and-answer session with 10 rounds of interactions consumes dozens of times more power than a search, and generating an image or video will produce even higher energy consumption.
Different Energy Consumption Curves
Traditional search engines are like efficient librarians. When you give a keyword, it helps you find matching pages in the index library. The amount of calculation involved in this process is relatively limited, because it is mainly doing the work of moving and sorting.
ChatGPT, by contrast, is not looking for ready-made answers, but "creating". It needs to mobilize tens of billions of parameters for probability prediction to simulate human language logic. This shift from search to generation increases the computing density exponentially.
This means that when we get used to asking AI "how to make a beef stew with red wine" instead of searching for the recipe on the web ourselves, we are virtually adding extra heat load to the Earth's atmosphere. Behind this convenience is the transfer of computing power costs across the whole society. We enjoy the intellectual dividend brought by AI, but leave the environmental cost to the future.
What is even more staggering is the "matrix" that supports all this intelligence — it is estimated that training the GPT-4 model (which takes about 90 to 100 days) consumes more than 50 million kWh of electricity. This huge electricity bill is enough to power a fleet of Tesla vehicles to travel thousands of circles around the Earth's equator. Behind every "thinking" loading bar is a computing power storm stacked with heat from silicon chips.
For a long time, the digital economy has been regarded as a "light asset" and "low energy consumption" growth model. Software, algorithms, and platforms all seem to run in a space that is relatively decoupled from the physical world. However, as the scale of AI models expands exponentially, this perception is being rapidly revised.
The training and inference of large models are essentially highly intensive computing processes, and the end of computing is undoubtedly the consumption of electricity. From GPUs, CPUs to dedicated acceleration chips, from server heat dissipation to network equipment operation, almost every link is continuously consuming electric energy. More importantly, AI is not a one-time investment, but a continuously evolving technical system: the larger the model, the more parameters, the wider the application, the higher the demand for computing power, and the deeper the dependence on electricity.
In this sense, computing power is becoming a new type of industrial production capacity, and electricity is the core input factor supporting this capacity. Just as the age of steel cannot do without coal and the age of oil cannot do without crude oil, the AI era also cannot do without a stable, sufficient and predictable power supply.
From the light that "lights up the LED for two minutes" to the scenario where "the fleet travels thousands of circles around the Earth", the energy consumption trajectory of digital technology is not a simple magnitude growth, but an exponential change, and the electric energy it consumes is evolving from a gentle trickle into a greedy behemoth.
Looking back at the early days of the Internet around 2000, the power of a computer was no more than 200 watts, and the server cluster of a medium-sized website consumed less than 100 kWh of electricity per day. Digital energy efficiency was like a tiny capillary, hidden under the huge shell of industrial civilization. At that time, the power consumption of global data centers was very limited, and could not even compare with the monthly energy consumption of the steel industry.
However, with the sweeping wave of AI, this energy efficiency transition has evolved into an energy storm. Reports from the International Energy Agency show that the power consumption of global data centers has soared to 415 TWh in 2024. In this global bill, the United States, China and Europe account for 45%, 25% and 15% respectively. What is even more amazing is that since 2017, the energy consumption of data centers has expanded at an average annual rate of 12%, which is more than 4 times the growth rate of global power consumption.
This expansion presents a frightening "super-linear" characteristic. From GPT-2 to GPT-4, when the scale of model parameters crosses thousands of times, the computing power black hole behind it also expands proportionally: the power consumption required for one GPT-4 training is enough to supply all residents in Beijing for a whole day. This means that every small step of iteration of computing power will widen the energy gap by a large margin.
In 2023, the AI-related energy consumption of cloud giants such as Amazon, Microsoft and Google surged, among which GPU clusters, as the core fuel, consumed most of the electricity. At the same time, the natural growth rate of global power supply is only 2.1%. Even though clean energy is accelerating its development, it still cannot keep up with the bottomless power consumption of AI algorithms.
In fact, when we talk about the cloud and cloud computing, these terms themselves have an obvious misleading tendency. It is not as light and easy as it seems, and what supports cloud computing is the physical entity composed of heavy steel, dense copper wires and brightly lit power plants.
When the evolution speed of digital civilization surpasses the carrying capacity of power infrastructure, power famine is no longer a prophecy, but a survival red line that the intelligent era must face. At this point, we have to re-examine a neglected proposition: electricity is no longer just the blood supporting industrial operation, but the food feeding intelligent civilization.
In the steam engine era, the combustion of coal pushed the gears to rotate, and electricity was only an auxiliary to drive the conveyor belt in the factory. Even if the power was cut off, workers could still rely on manpower to maintain basic production. In the electrical age, electricity entered the home and became the basic guarantee for lighting bulbs and driving refrigerators, but without electricity, humans could still maintain basic life through candles and firewood.
In the intelligent era, electricity becomes the energy source for AI to think, learn and make decisions. Without electricity, ChatGPT cannot generate a single sentence, the lidar of self-driving cars will stop scanning, the traffic lights of smart cities will be paralyzed, and even the AI diagnosis system in hospitals will lose the ability to analyze CT images... The "death" of intelligent systems often starts with the interruption of power. This dependency relationship from "auxiliary" to "essential" has completely reshaped the value orientation of electricity: it is no longer a replaceable energy source, but the premise of intelligent existence.
Why Is AI So Power-Hungry
Looking back at every leap of digital civilization, it is not difficult to find a recurring rule: every improvement in efficiency is accompanied by a synchronous rise in power demand, leaving a clear trajectory in the history of technological evolution.
In 2007, when Steve Jobs released the first generation iPhone at the Moscone Center in San Francisco, most people were immersed in the light feeling of "putting the world in your pocket". However, few people realized that the price of this convenience was about to be passed on to power infrastructure distributed all over the world.
The popularization of smartphones pushed the number of mobile Internet users from less than 1 billion to 5.5 billion in 2024. In order to support this ubiquitous digital ocean, a crazy competition for the construction of global communication base stations has begun. The number of base stations has grown from 500,000 in 2007 to tens of millions in 2024. Especially with the advent of the 5G era, since the coverage radius of high-frequency signals is much smaller than that of 4G, the density of base stations has been forced to increase significantly, and the deployment of 5G base stations alone around the world has exceeded 4 million.
Behind this connectivity revolution is an astonishing 4-fold increase in power consumption over 18 years. In 2024, the total power consumption of global mobile base stations is comparable to the total annual electricity consumption of Egypt. The smooth short videos and millisecond-level social interactions under people's fingertips are actually maintained by tens of millions of never-extinguishing power stations on the surface. However, this is only the beginning of the mobile Internet era. When the AI era arrives, this energy consumption curve changes from a steady rise to an almost vertical sprint.
The Go game at the Four Seasons Hotel Seoul in South Korea in 2016 was not only the moment when AlphaGo defeated Lee Sedol, but also a turning point for the qualitative change of human computing power demand. Before that, Moore's Law was still the golden rule in the technology industry, and after this game, AI's hunger for computing power began to surpass the speed of hardware iteration.
Since 2016, the global computing power demand for AI chips has grown at an astonishing rate of 50% per year. This growth is no longer linear, but exponential. The related power consumption is also surging at a nearly synchronous growth rate of 45%. By 2023, the total power consumption of chips dedicated to AI tasks worldwide has approached 200 billion kWh. This number is beyond imagination, and it even exceeds the total annual electricity consumption of a medium-sized country like Malaysia.
In November 2022, the debut of ChatGPT made generative AI refresh the energy consumption record again. The scale of ChatGPT users exceeded 100 million in just two months. This explosion of user volume brought about the exponential collapse of reasoning cost and the exponential expansion of total energy consumption. At present, the daily power consumption of ChatGPT is equivalent to the daily electricity consumption of about 100,000 ordinary households. When you ask AI to write a poem or debug a piece of code, thousands of top GPUs provided by NVIDIA in the data center thousands of miles away are running at full speed in a state of high heat.
However, the current power supply system has not fully kept up with this qualitative change in demand.
Although major tech giants have promised to achieve carbon neutrality, from the perspective of the current macro energy structure, fossil energy (coal, oil, natural gas) still accounts for about 55% of the global power structure, which means that although we see a clean and efficient AI interface on the screen, the energy background behind it still has a strong carbon footprint.
Data shows that the carbon emissions of global data centers in 2025 are as high as 350 million tons, equivalent to 1/3 of the carbon emissions of the global aviation industry. This puts AI in an awkward paradox — we hope to use intelligent means such as AI to reduce carbon emissions, but AI itself has become one of the important factors exacerbating climate change.
To truly understand why AI consumes so much power, we need to look through its two core links: training and inference. The former is a large-scale parallel computing process that takes months, like making gold in a huge digital blast furnace. Its essence is to exchange computing power for cognition, that is, the process of allowing machines to find rules from data through repeated calculations, just like humans form cognitive abilities through a lot of reading and practice. The latter is real-time response serving hundreds of millions of users, essentially a process of processing newly input information and generating results based on learned parameters and rules, similar to humans using learned knowledge to answer questions. Every click in the reasoning process consumes a tiny amount of power, but it is these little drops that come together to form this unprecedented energy challenge.
In order to ease the expensive power bills and increasing public pressure, tech giants have to build data centers in areas rich in hydropower, wind power and geothermal resources. Microsoft built a data center on the coast of Orkney Islands, Scotland, using the local sea wind with an average annual wind speed of 10 m/s to generate electricity, and using seawater to cool the servers, controlling the PUE (Power Usage Effectiveness) of the data center below the extreme value of 1.1 (PUE is the internationally recognized core indicator to measure the energy efficiency of data centers. The closer PUE is to 1, the higher the energy utilization efficiency, and the global average level is about 1.5). Google built a server cluster in Reykjavik, Iceland, relying on the unique local geothermal resources, which reduces Google's local power cost by 2/3 compared with Silicon Valley, and 100% of the power is clean energy. Amazon is located on the banks of the Columbia River in Washington State, using the hydropower resources of the Columbia River to supply power for the company. At present, the annual average power consumption of data centers in this area accounts for 25% of the total power generation of local hydropower stations.
These layouts chasing energy expose the deep contradiction between intelligent civilization and power supply — when the "food" required by AI is still highly dependent on fossil energy, it means that behind every line of code stands a smoking chimney; when the power supply cannot keep up with the "appetite" of AI, the sustainable development of digital civilization becomes a challenge to be solved urgently.
At present, there is no sign that this steeply rising energy consumption curve will flatten. As larger-scale Artificial General Intelligence (AGI) models enter the training agenda, experts predict that the demand for computing power may multiply several times, thus driving the continuous growth of the power consumption curve.
Invisible Load
What is more challenging is that AI's energy consumption is also spreading to hidden areas, forming a global hidden energy consumption network. In the past, we could clearly perceive the consumption of electricity: the light bulb will turn on when we turn on the light, there will be cold wind when we turn on the air conditioner, and the engine needs to refuel when we drive. But the energy consumption of AI is hidden in the invisible digital interaction, becoming an "invisible load".
When you swipe a 15-second AI-generated short video on your mobile phone, a large amount of energy is consumed by the server to process the video data. When you rely on navigation software to plan the optimal route of 10 kilometers, AI must analyze dozens of dimensions of data such as real-time road conditions, traffic light changes and congestion models in milliseconds. Even a precise advertisement push you receive is the result of high-energy reasoning calculation after the algorithm analyzes your huge browsing traces and consumption habits. These energy consumptions are often ignored by users — few people realize that daily behaviors such as swiping short videos, using navigation, and watching advertisements are all paying for AI's electricity bills.
Behind the hidden computing power is also the consumption of heat. When tens of thousands of chips run at full speed at the same time, the data center is like a huge electric furnace. If the heat is not dissipated in time, the chips will burn out in a few seconds. In order to maintain the constant temperature of 22°C, the cooling system usually consumes about 40% of the total power consumption of the entire data center.
According to the latest calculation by the research team of the University of California, in regions with mild climate, every 20 to 50 conversations of ChatGPT are equivalent to "drinking" a 500ml bottle of pure water. For training large models like GPT-5 or higher-level models, their water consumption is even astronomical. Microsoft's data center in Iowa consumed more than 43,000 tons of water in one month for model training, accounting for nearly 10% of the local total water supply.
This kind of consumption has also triggered unprecedented social conflicts in arid areas. In Chile and Uruguay, Google's data center expansion plan has triggered strong protests from local people, because in years of frequent droughts, the cooling water for AI directly competes with agricultural irrigation and domestic water for residents.
Engineers are trying to break through this physical limit. Some companies sink data centers into the cold seabed of the North Sea, using natural seawater for heat exchange; some companies move them to the Scandinavian Peninsula near the Arctic Circle, using cold wind for natural cooling.
However, these solutions cannot fundamentally solve the problem. As long as the logic of AI is still based on the electronics of silicon-based chips, the generation of heat is an inevitable physical necessity. This heat dissipation anxiety is forcing humans to re-examine the distribution of energy — do we really need to use enough power to support a small town to generate a personalized selfie avatar for a high school student, or to polish a mediocre marketing copy.
The cumulative effect of this AI-driven "hidden consumption" is triggering far-reaching changes in the global energy pattern and bringing potential risks to regional power supply and demand crises. For example, in California, the United States, although there has been no mandatory power rationing caused by AI, the concentrated construction of large-scale data centers has become an unprecedented pressure source for the power grid. California grid operators have issued a warning that the surging peak load will