Large language models and the rise of data centres

By Matthew Parish, Associate Editor

Friday 29 March 2026

The artificial intelligence revolution is usually described in ethereal terms. Politicians speak about “innovation”. Investors speak about “scale”. Executives speak about “cloud infrastructure”, as though computation were something floating harmlessly in the upper atmosphere. Yet the reality of large language models is profoundly physical. Every sentence generated by a machine emerges from electricity, copper, silicon, water, steel, concrete and land. Behind the illusion of incorporeal intelligence lies an industrial apparatus of astonishing scale: the modern data centre.

The rise of large language models has therefore inaugurated not merely a software revolution but an infrastructural revolution. Across North America, Europe, the Gulf states and East Asia, colossal warehouse-like facilities are being constructed to house the computational hardware necessary to train and operate increasingly sophisticated artificial intelligence systems. These buildings are becoming amongst the largest consumers of electricity in the world. They are reshaping energy policy, distorting real estate markets, placing pressure upon national electricity grids and accelerating geopolitical competition for energy security. Their emergence deserves far more public scrutiny than it has thus far received.

The central problem is straightforward. Large language models require incomprehensible quantities of computation. Training frontier systems involves processing trillions of parameters across massive clusters of graphical processing units, or GPUs, operating continuously for weeks or months. Once trained, the models must then respond to millions or billions of daily user interactions. Every question asked of an artificial intelligence system requires electricity-intensive inference calculations. The result is an industry whose appetite for power appears almost without limit.

Historically, advances in computing were often associated with efficiency gains. Moore’s Law implied that computers became simultaneously more powerful and more compact. Smartphones replaced entire rooms of analogue equipment. Modern laptops consume vastly less electricity than older mainframes. The artificial intelligence era is reversing that trajectory. Computational demand is now increasing faster than efficiency improvements can compensate. As models become larger, more complex and more commercially integrated, total energy consumption rises dramatically even when individual processors become more efficient.

This phenomenon has alarmed electrical grid planners throughout the developed world. In the United States, regions already struggling with ageing infrastructure now face enormous new industrial electricity demands from artificial intelligence firms. Utilities that anticipated gradual decarbonisation and modest increases in consumption are instead being confronted with requests for gigawatts of additional capacity. A single advanced data centre campus may consume as much electricity as a medium-sized city.

The implications extend far beyond electricity bills. Artificial intelligence data centres are now influencing national energy strategies. Natural gas projects previously considered economically uncertain are being revived because of expected artificial intelligence demand. Nuclear energy is receiving renewed political interest partly because only nuclear reactors can reliably provide the constant baseload electricity these facilities require. Renewable energy alone often cannot satisfy the uninterrupted computational needs of massive model training operations without vast battery storage systems that remain prohibitively expensive.

Consequently the artificial intelligence revolution may paradoxically delay environmental transition efforts. Many governments have publicly committed themselves to ambitious climate objectives whilst simultaneously encouraging the construction of computational infrastructure whose electricity requirements may make those objectives unattainable. There is a growing disjunction between the rhetoric of sustainability and the material realities of artificial intelligence deployment.

Water consumption presents another underappreciated concern. Data centres generate extraordinary quantities of heat. Cooling systems therefore require enormous volumes of water, particularly in hotter climates. Some facilities consume millions of litres daily. Communities already vulnerable to drought increasingly find themselves competing with technology corporations for access to local water supplies. The spectacle of artificial intelligence chatbots consuming fresh water in arid regions whilst governments urge ordinary citizens to conserve household usage illustrates the distorted priorities that can emerge when technological prestige overtakes practical governance.

There is also a geopolitical dimension to this transformation. Artificial intelligence infrastructure depends not merely upon energy but upon highly specialised semiconductor supply chains. Advanced chips are overwhelmingly manufactured in a handful of locations, most notably Taiwan. Hence data centres become part of a broader strategic contest involving rare earth minerals, chip fabrication, submarine cables, electrical grids and energy resources. The artificial intelligence race increasingly resembles an arms race conducted through industrial infrastructure rather than missiles.

This has particular relevance for Europe. The European Union aspires to compete in artificial intelligence whilst simultaneously maintaining stringent environmental standards and reducing dependence upon foreign energy. Yet Europe’s comparatively high electricity prices and fragmented infrastructure place it at a structural disadvantage compared with the United States, China and Gulf states possessing cheaper energy resources. Artificial intelligence may therefore intensify Europe’s long-standing industrial competitiveness crisis.

Ukraine offers a striking contrast. Amidst wartime destruction and electricity shortages caused by Russian attacks upon civilian infrastructure, Ukrainians have become acutely conscious of the fragility and value of electrical power. Rolling blackouts, damaged substations and emergency generators have transformed electricity from an invisible convenience into a strategic necessity. Against that backdrop, the notion that wealthy societies might devote immense electrical resources to generating synthetic essays, advertising copy or conversational novelties can appear morally disorientating.

This does not mean artificial intelligence lacks value. Large language models possess substantial potential for scientific research, medicine, logistics, translation, education and military coordination. Ukraine herself has demonstrated how advanced technology can partially compensate for demographic and material disadvantages in war. Artificial intelligence may ultimately contribute significantly to productivity and human welfare. But utility does not absolve society from evaluating costs honestly.

The deeper concern is that contemporary artificial intelligence development is unfolding without serious democratic discussion regarding limits. Technology corporations routinely frame expansion as inevitable. More data centres must be built because users demand more artificial intelligence services; more electricity must therefore be generated; more land must therefore be acquired; more water must therefore be consumed. The logic becomes self-justifying. Yet societies routinely regulate other industries whose externalities threaten public welfare. Heavy industry, mining and chemical manufacturing all operate within political constraints. Artificial intelligence infrastructure should be no exception.

There is furthermore a psychological dimension to the issue. Modern societies increasingly confuse computational abundance with civilisational progress. The existence of more powerful models is assumed automatically to constitute advancement. But civilisations are not judged merely by their technological sophistication. They are judged by whether technology serves human flourishing in proportionate and sustainable ways. A society that devotes ever larger shares of her energy resources to automated text generation whilst struggling to house citizens affordably or maintain reliable public infrastructure may reasonably be accused of losing strategic perspective.

The architecture of modern data centres itself symbolises this contradiction. These vast anonymous structures often employ relatively few people compared with the enormous resources they consume. Unlike factories of earlier industrial eras, they do not create large working communities. They produce no visible goods. They simply process information ceaselessly, hidden behind security fences and cooling towers, consuming rivers of electricity to sustain the digital appetites of modern civilisation.

One may therefore ask whether the artificial intelligence economy risks reproducing the pathologies of earlier financialised economic models: immense capital concentration, extraordinary energy consumption and growing abstraction from tangible human needs. The danger is not merely environmental. It is civilisational. Societies can become mesmerised by technological systems whose scale gradually escapes meaningful democratic control.

There are historical precedents for such anxiety. Nineteenth-century observers worried about railways transforming landscapes and social rhythms. Twentieth-century critics feared the environmental and psychological consequences of automobile dependency. Those concerns were not entirely misplaced. Modern societies are still grappling with the consequences of infrastructure choices made decades earlier. Artificial intelligence data centres may prove similarly consequential.

The challenge therefore lies not in rejecting artificial intelligence outright but in restoring proportionality and accountability to its development. Governments must ask difficult questions about energy allocation, environmental sustainability and strategic necessity. Citizens must understand that the “cloud” is not immaterial. Every digital convenience possesses a physical footprint somewhere upon the earth.

The rise of large language models is often portrayed as humanity entering a new intellectual age. Perhaps she is. But intelligence without restraint has rarely guaranteed wisdom. If the future of civilisation requires continents increasingly covered with energy-hungry computational fortresses merely to sustain the endless production of synthetic language, then society may eventually discover that she has built something technologically magnificent yet strategically unbalanced.