Artificial intelligence was supposed to be a technology hidden away in the cloud. It was meant to operate in the background: speeding up processes, analysing data, supporting staff, automating decisions and improving productivity. In practice, however, it is becoming increasingly clear that its development depends on very tangible resources: data centres, transmission networks, connection capacity, cooling, capital and energy.
This shifts the conversation about AI far beyond the IT department. The most important question is no longer simply which models a company will choose, what data it will use, and which processes it will automate. An increasingly important question is where this technology will physically operate, how much it will cost to maintain, and whether the energy system will be able to cope with the growing demand.
AI does not merely increase energy consumption. It is beginning to reshape the geography of investment, the logic of infrastructure planning and the way in which nations and companies will think about competitive advantage.
For years, data centres have been the backbone of the digital economy – an important but largely invisible one. The development of generative AI is bringing them to the fore. According to data cited by the International Energy Agency, global energy consumption by data centres stood at around 415 TWh in 2024, representing approximately 1.5 per cent of global electricity consumption. By 2030, this could rise to around 945 TWh.
This figure alone does not mean that data centres will consume the entire energy system. Something else is far more important: their impact will be very significant at a local level. AI is not distributed evenly across the world map. It is concentrated where there is capital, infrastructure, cloud providers, suitable land, networks and access to energy.
That is why the problem is not solely about how much electricity AI will consume on a global scale. The real question is: where will this energy be needed, how quickly, and who will have to adapt the infrastructure accordingly.
Research into the location of AI data centres indicates that the vast majority of projected computing power is set to be concentrated in North America, Western Europe and the Asia-Pacific region. Some locations, such as Virginia, Oregon and Ireland, are already demonstrating just how great the pressure on local electricity grids could be. AI is therefore ceasing to be an abstract technology featured in boardroom presentations. It is becoming a very tangible strain on infrastructure.
From an economic perspective, this is a fundamental shift. Data centres are beginning to compete for energy with industry, cities, electric mobility, district heating and the energy transition. The very same system that is intended to support the decarbonisation of factories, the development of electric transport and the modernisation of buildings must, at the same time, meet the rapidly growing needs of the digital economy.
Where the grid is robust, connections are fast and energy is relatively accessible, new clusters of digital investment may emerge. Where infrastructure is overloaded or being modernised too slowly, the economy may remain primarily a consumer of AI services, rather than a place where its foundations are being developed.
This is reshaping the landscape of competitive advantages. In previous decades, a region’s attractiveness was determined by talent, labour costs, market access, logistics and capital. Now, it will become increasingly important whether a given country or region has the energy, network and capacity to rapidly deploy digital infrastructure.
Energy companies, too, cannot afford the luxury of planning with peace of mind. According to a Capgemini study cited by ITPro, a large proportion of energy suppliers are struggling to forecast demand from AI-powered data centres. This is not a minor technical issue. The quality of these forecasts determines where networks need to be expanded, how much capacity needs to be contracted, and how to avoid a situation where declared demand ties up resources, even though not all projects will ultimately be implemented.
This gives rise to a phenomenon known as ‘phantom load’ — forecast energy demand that may never translate into actual investment. For network operators and regulators, this is an exceptionally difficult challenge. They must prepare the system for growth, but do not always know which projects will actually go ahead, on what scale, or when.
This brings the topic of AI into the realm of economic policy. Governments will have to answer the question of whether data centres are simply energy consumers or strategic infrastructure. If they are strategic, plans must be drawn up for their grids, energy sources, cooling systems, locations and connection rules. If they are treated like any other consumer, another question arises: who will bear the cost of expanding the infrastructure that will mainly be used by the largest tech players?
This dilemma will be particularly significant in Europe. On the one hand, the continent wishes to develop its own capabilities in AI, strengthen its digital sovereignty and attract investment in data infrastructure. On the other hand, it operates in an environment characterised by high energy costs, ambitious climate targets, regulatory tensions and growing public sensitivity regarding the location of large-scale infrastructure investments.
Therefore, the European response cannot be limited to the slogan ‘let’s build more data centres’. Energy efficiency, the use of zero-emission energy, water conservation, equipment reuse and heat recovery will become increasingly important. This is no longer merely an ESG agenda. It is a prerequisite for social, regulatory and economic acceptance of the development of AI infrastructure.
The technology sector is already trying to free itself from grid constraints. There is increasing talk of on-site energy sources, behind-the-meter solutions, energy storage, long-term contracts for renewable energy, gas as a transitional solution, and, in the longer term, small nuclear reactors. Cheaper electricity is not the only thing at stake. What is at stake is predictability, the speed at which new computing capacity can be brought online, and reduced dependence on overloaded grids.
There is also a more optimistic scenario. AI data centres need not be merely a rigid, passive load on the energy system. Some of the computational load can be shifted over time or between locations, responding to energy prices, power availability and grid conditions. In this vision, data centres become more flexible participants in the energy market: they do not merely consume electricity, but are able to adapt their operations to system conditions.
This is still a direction of development, rather than a widespread standard. But it shows that the conflict between AI and the energy sector need not be purely a zero-sum game. Well-designed digital infrastructure can support more informed demand management, better utilisation of renewable energy and new models of cooperation between the technology and energy sectors.
For businesses, the conclusion is simple, though far-reaching. AI should not be viewed solely as a technological project. On a larger scale, it becomes an infrastructure and operational project. The cost of the model, licences or API access is only the visible part of the equation. Added to this are energy, data, integrations, security, availability, redundancy, regulatory compliance and dependence on infrastructure providers.
A CFO should therefore not ask solely how much it costs to implement AI. They should ask what the unit cost of a process, decision, query or automation will be once the system moves from pilot to production. The CIO should not merely ask whether the solution works. They should know whether the organisation has the capacity to maintain it at scale.
The key point, therefore, goes beyond the energy bill alone. The AI economy will not develop solely where the best models and the largest datasets are found. It will develop where it can find energy, networks, cooling, capital, expertise and the consent to build new infrastructure.
Countries and regions with a strong energy base could become the cornerstone of the AI economy. Those that fail to build such a base will be able to use off-the-shelf services, but will find it harder to play a part in laying the foundations of the new digital economy.
So the point is not to reiterate that AI consumes a lot of electricity. That is already well known. What is more important is to understand that AI is becoming a new, highly dynamic player in the race for resources that were already needed by industry, cities and the energy transition.
If this demand is not properly planned, it may drive up costs, delay other investments and exacerbate local tensions surrounding the grid. If it is wisely integrated into energy strategy, it can accelerate infrastructure modernisation, the development of new demand-management models and more informed planning for the digital economy.
The boundaries of the AI economy will therefore be defined not only by algorithms. They will also be defined by megawatts.
