There is something slightly odd happening with artificial intelligence. We want more of it everywhere: in search, accounting software, customer service, coding tools and increasingly in the businesses we run. What we seem rather less enthusiastic about is having the infrastructure required to make all of this work built anywhere near us. The word “cloud” was a brilliant bit of marketing. It makes computing sound clean, weightless and somewhere safely out of sight. In reality, the cloud consists of very large buildings full of computers requiring electricity, cooling, cables, transformers, generators and, in some cases, substantial amounts of water. And we are going to need more of them.
The International Energy Agency estimated that data centres accounted for around 1.5% of global electricity consumption in 2024 and expects their electricity demand to more than double by 2030. The precise number may turn out to be wrong. Most forecasts concerning AI probably will be. But the general direction is difficult to argue with. Training increasingly sophisticated models requires enormous computing power. So does running those models for millions of people, integrating AI into businesses and supporting the rest of the digital economy we already take for granted. Which is why simply arguing that we should stop building data centres makes little sense. We cannot demand ever more digital services while objecting to the machinery that provides them. It is rather like buying three electric cars and then campaigning against the local substation.
That does not mean the industry gets a free pass. A large data centre can place considerable pressure on the electricity network, particularly where several developments are concentrated in the same area. New substations, transmission infrastructure and generation capacity may be required. If a hyperscaler creates an exceptional infrastructure requirement, it is perfectly reasonable to ask how much of that cost should be paid by the hyperscaler rather than ending up on everyone else’s electricity bill. Water attracts similar arguments, although the reality is more complicated than some headlines suggest. Data centres use different cooling technologies and local circumstances vary enormously. Water consumption in a water-stressed region is clearly a rather different issue from consumption somewhere with plentiful supply. But where resources are constrained, communities are entitled to ask who gets priority and why.
Then there are the much less glamorous issues of noise, construction traffic, land use and the arrival of enormous industrial buildings on the skyline. These do not receive much attention at global AI conferences. They matter considerably more if the proposed building happens to be at the end of your road. This is where the industry occasionally does itself few favours. Data centres may be essential infrastructure, but describing something as essential does not automatically make every proposed location sensible. Roads are essential too. We do not therefore build motorways wherever somebody finds a convenient field.
There is also a financial question that may eventually prove just as important. Extraordinary sums of money are being committed to AI infrastructure before anybody can know precisely what the long-term economics of AI will look like. The companies making these investments are hardly naïve, and many have both enormous cash flows and very good reasons for investing ahead of demand. But there is still an interesting mismatch between the life of the infrastructure and the technology being installed inside it. A data-centre building may operate for decades. Power infrastructure can last even longer. The processors are rather less considerate. In AI, extremely expensive hardware can become yesterday’s technology remarkably quickly.
That does not mean AI is a bubble. I certainly would not pretend to know. But investors should probably pay at least as much attention to utilisation, replacement cycles and the revenues ultimately produced by AI as they do to announcements of another multi-billion-pound campus. There is another part of the story that receives less attention. While data centres have become bigger, the computers sitting in offices and on desks have also become much more powerful. Nvidia’s DGX Spark, for example, is a compact desktop AI system with 128GB of unified memory. Nvidia says it can run inference on AI models of up to 200 billion parameters and fine-tune models of up to 70 billion. The numbers are Nvidia’s, of course, but the more interesting point is that useful AI computing is starting to move in both directions.
For some work, sending everything to a remote data centre may become unnecessary. A developer testing models, a company analysing confidential documents or a professional firm that simply prefers sensitive information to remain on its premises may increasingly be able to do more locally. There may also be a fairly ordinary economic argument for it. If computing capacity is being used heavily enough, owning some of it can sometimes make more sense than renting it indefinitely. None of this means the data centre is disappearing. Training frontier models still requires extraordinary computing resources. Serving AI applications to millions of users does too. Large organisations need resilience, connectivity and scale that cannot be replicated by putting a clever computer underneath somebody’s desk.
But perhaps we have framed the argument incorrectly. The real choice is not between data centres and communities, or between the cloud and the desktop. It is about becoming rather better at deciding what computing belongs where. The largest workloads will remain in very large data centres. Some processing will move into regional and edge facilities. Other inference, development and privacy-sensitive work may increasingly take place locally. If that happens, the infrastructure supporting AI becomes more distributed, which may ultimately prove healthier both commercially and technically.
The same pragmatism should apply to where we build data centres. Put them where the electricity network can realistically support them. Be transparent about water and power consumption. Make companies creating exceptional infrastructure requirements pay an appropriate share of the cost. Keep improving efficiency. And perhaps speak properly to the people who are going to have several acres of computers built nearby before the planning application arrives on their doorstep. None of this is especially revolutionary. It is roughly how sensible infrastructure planning is supposed to work.
AI is going to require a great deal of computing and that means more data centres. But it does not follow that every proposed development should be built, that every location makes sense, or that every AI workload needs to travel across the country to an enormous server farm. The future is probably not the end of the data centre. It is simply one in which we become a little more intelligent about what we put in them.
