Artificial intelligence is often discussed as though it exists somewhere beyond the physical world. We talk about models living in the cloud, algorithms thinking, and machines becoming increasingly intelligent, while the infrastructure that makes all of this possible remains largely invisible. Yet AI is not weightless or immaterial. It depends on vast networks of computers, data centres and cooling systems, all of which require one fundamental resource: electricity.
That simple fact deserves more attention. The rapid development of AI is not only a story about software, data and computing power. It is also an energy story. Every time we ask an AI system to generate an image, write a piece of code, analyse a document or produce an answer, computing resources are being used somewhere. At the other end of that process are physical machines consuming electricity, producing heat and drawing on infrastructure that has to be built, maintained and powered.
This creates an interesting tension at the heart of the AI boom. We are building increasingly sophisticated systems that promise to make almost every part of the economy more efficient, while simultaneously creating a new and rapidly growing demand for energy. AI may eventually help us optimise electricity grids, improve battery technology, reduce industrial waste and accelerate scientific research. But the systems that perform those tasks have to be powered in the first place. The question is therefore not simply what AI can do, but what we are prepared to spend in order to make it possible.
For most users, the energy involved in an interaction with AI is effectively invisible. A prompt is typed into a box and an answer appears a few seconds later. There is no obvious physical transaction taking place. Unlike filling a car with petrol or switching on a heating system, there is no immediate reminder that a resource is being consumed. The interface makes computing feel almost free.
That abstraction is one of the defining characteristics of modern technology. We rarely think about the physical infrastructure behind the services we use every day. A photograph stored online does not feel as though it occupies a physical space. A video streamed to a phone does not feel as though it requires a network of servers and data centres. AI takes that abstraction a step further because it gives the impression that intelligence itself can be produced almost on demand.
But there is no cloud in the literal sense. There are buildings filled with machines, connected to electrical grids and supported by complex cooling and power systems. The more AI becomes embedded in everyday life, the more important those physical systems become. Data centres need land, water, connections to the grid and, above all, reliable supplies of electricity. As demand for AI computing increases, those requirements become an increasingly significant part of the technology story.
This raises a question that is easy to overlook amid the excitement surrounding increasingly capable models: how much of what we ask AI to do is actually worth the energy required to do it?
The question is not intended as an argument against AI. There are obvious and potentially enormous benefits to the technology. AI could contribute to medical research, scientific discovery, engineering, education and productivity in ways that are difficult to quantify today. It may allow researchers to process information at a scale that would otherwise be impossible, help businesses identify inefficiencies and give individuals access to tools that were previously available only to specialists.
The problem is that all of these uses exist alongside a vast amount of much less consequential activity. We are also using AI to generate endless images, rewrite sentences that were already perfectly understandable, produce increasingly elaborate pieces of disposable content and automate tasks simply because we can. Individually, these uses may seem insignificant. Collectively, they form part of a much larger demand for computing.
This is where the energy question becomes more interesting than a simple calculation of AI’s environmental impact. It forces us to think about value. We are accustomed to considering the cost of technology in financial terms, but electricity introduces another kind of cost. If computing resources are finite, or if expanding them requires substantial investment in new energy infrastructure, then there is a legitimate question about what those resources should be used for.
We do not normally ask whether sending an email is worth the electricity required to send it. Nor do we ask whether watching another episode of a television programme justifies the energy consumed by the servers, networks and devices involved. Modern life depends on enormous amounts of computing and electricity, and most of us have accepted that consumption as part of the background infrastructure of society.
AI may be different because of the speed at which demand is growing and the scale at which the technology is being deployed. It is not simply another digital service being added to the existing system. It has the potential to become one of the major consumers of computing resources in its own right. If that happens, energy policy and technology policy will become increasingly difficult to separate.
This could change the way we think about the AI industry. For years, the dominant image of a technology company has been one of software engineers, offices and digital products. The physical reality of AI looks rather different. It involves industrial-scale facilities, sophisticated electrical infrastructure and enormous quantities of computing hardware. In that sense, the AI industry may have more in common with traditional heavy infrastructure than the weightless image of Silicon Valley suggests.
The implications extend beyond individual companies. Governments will have to consider where new data centres are built, how they connect to national grids and who ultimately pays for the infrastructure required to support them. Energy companies will have to anticipate new sources of demand. Communities may have to decide whether the economic benefits of hosting large computing facilities justify the additional pressure on local infrastructure and resources.
None of this means that AI should be switched off. In fact, that is probably the least useful way of thinking about the issue. Once a technology becomes sufficiently valuable, society rarely chooses to abandon it simply because it consumes resources. The more realistic challenge is to make the consumption visible and to become more deliberate about what we are using the technology for.
Perhaps the important question is not whether we can unplug AI, but whether we can learn to use it responsibly in an energy-constrained world. That could mean greater transparency from technology companies about the resources their systems consume. It could mean building new computing infrastructure alongside new sources of low-carbon electricity rather than assuming the grid will simply absorb growing demand. It could also mean becoming more conscious as users about whether every possible application of AI is worth pursuing.
There is something revealing about the phrase “artificial intelligence”. It suggests that intelligence is the central resource being created, when in reality AI depends on another, much older resource: energy. The machines may appear increasingly autonomous, but they remain connected to a very physical system of power stations, cables, servers and cooling equipment.
That does not make AI less impressive. If anything, it makes the technology more interesting. It reminds us that even the most apparently intangible developments are ultimately grounded in the physical world.
The AI debate has focused heavily on what machines might become and what they might mean for human beings. Those questions will remain important. But alongside them is a much simpler question, one that may become increasingly difficult to ignore: how much electricity are we willing to dedicate to artificial intelligence, and what do we want in return?
We may not be able to unplug AI. More importantly, we may not want to. But we should at least be conscious that someone, somewhere, has to keep the power running.