What is driving the surge in AI data centre power demand?

For twenty years, data centres were a quiet success story for electricity planners. Demand grew, but efficiency gains grew alongside it, so the sector’s total consumption stayed remarkably flat even as computing volumes exploded. 

Artificial intelligence has broken that pattern. AI workloads consume far more power per unit of computing, run almost continuously, and are being deployed at a pace that outstrips efficiency improvements. The result is a category of demand growth that European power markets did not plan for and, in several regions, cannot currently accommodate. 

Why data centres are consuming more electricity than ever 

European data centres consumed roughly 96 TWh of electricity in 2024, around 3% of EU demand. Projections indicate this will reach approximately 168 TWh by 2030, an increase of about 75% over six years. 

To put this in context, the comparison is with electric vehicles. Data centre demand growth between 2024 and 2030 is projected at around 72 TWh, exceeding the roughly 67 TWh expected from EV uptake over the same period. Electrification of transport has dominated demand forecasting for a decade; data centres are now growing faster. 

Within a facility, the load divides into computing itself and the cooling required to remove the heat it generates. Efficiency improvements have focused on the cooling side, where the industry has steadily driven down power usage effectiveness. Those gains are real but finite, and they cannot offset an order-of-magnitude increase in the computing load itself. 

The critical point for anyone forecasting demand is that this growth is lumpy rather than smooth. A single hyperscale campus can request several hundred megawatts at one connection point - comparable to a mid-sized city - with a lead time measured in months rather than the years a network needs to prepare for it. 

The capital behind that growth gives some sense of its momentum. European industry estimates put the investment required at roughly €176 billion cumulatively between 2026 and 2031, with around €25 billion a year for large AI-dedicated facilities alone. That is a scale of committed spending which tends to find a way through planning and connection constraints rather than being deterred by them. 

How AI workloads differ from traditional data centre demand 

The difference is not simply that AI uses more power. It uses power with a fundamentally different shape. 

Rack density is the starting point. A conventional server rack draws somewhere between 5 and 10 kW. AI training racks packed with accelerators can draw ten times that or more, which changes the engineering of the entire building. Air cooling stops being viable, pushing operators towards liquid cooling. 

Utilisation is the second key difference, and it matters more commercially. A traditional data centre serves variable user demand, so its load rises and falls throughout the day, and its average draw sits well below its peak. AI training runs continuously at near-full utilisation for weeks, because expensive accelerators sitting idle represent wasted capital. 

That combination is awkward. A flat, inflexible load provides none of the demand-side flexibility that system operators increasingly rely on, and it consumes network capacity around the clock rather than only at peak times. 

Inference workloads serving queries to trained models are somewhat more tractable and can in principle be shifted between regions or times. Whether operators will accept the commercial constraints of doing so is a live question. 

Which regions are feeling the pressure most 

Ireland is the clearest case in Europe. Data centres accounted for around 23% of national electricity consumption in 2025, up from roughly 5% a decade earlier, and consumed almost as much as every home in the country combined. In the Dublin region, the concentration is far higher. 

That trajectory prompted the regulator to impose a de facto moratorium on new Dublin-area connections in 2021. At the end of 2025, it was replaced by a conditional framework: larger new connections must now provide generation located nearby or behind the meter, sized to their connection, be sited in unconstrained parts of the network, and match a substantial majority of their annual consumption with new Irish renewable generation. The moratorium became a set of conditions rather than a prohibition. 

The traditional European cluster - Frankfurt, London, Amsterdam, Paris and Dublin - remains dominant but is losing share, precisely because power and land constraints in those markets are pushing new capacity elsewhere. Projections suggest around half of European capacity will sit outside that cluster by 2035. 

The Nordics have attracted a significant share of that displaced demand, thanks to abundant hydro and wind, cool ambient temperatures that reduce cooling loads, and relatively available grid capacity. Even there, the welcome is becoming conditional, with scrutiny over whether large loads should be allowed to claim access to low-carbon power ahead of other users. 

The Netherlands illustrates how quickly political tolerance can shift. Amsterdam’s early moratorium was followed by sustained public and municipal resistance, and at least one major hyperscale project was abandoned in response to local opposition. Scotland has more recently considered pausing new applications while it assesses the energy implications. 

The pattern across all these markets is that the constraint is rarely purely technical. Network capacity sets the outer limit, but political consent determines where within that limit a project is actually welcome and consent has proved considerably more volatile than engineering. 

What this means for grid capacity and connection queues 

The main challenge is the timing difference: a data centre can be designed, built, and commissioned in about two years, while the necessary transmission reinforcement usually takes nearly a decade. 

Data centres therefore arrive in an already congested queue, competing with generation and storage projects for the same scarce network capacity. In several markets, demand connections are now assessed within the same reform processes that were designed to manage the generation backlog, as covered in What are grid connection queues, and why are they delaying the energy transition? 

Three responses are emerging. The first is conditional connection, in which large loads accept curtailment or flexibility obligations in exchange for earlier connection. The second is bringing your own generation, with operators funding on-site or nearby capacity so the network sees a smaller net load now a formal requirement in Ireland. The third is siting relocation, moving to locations where capacity exists rather than where latency is optimal. 

There is also a growing argument that AI load should be treated as a flexibility resource rather than a passive one. If a fraction of inference capacity could be shifted away from system peaks, a large, flat load would become considerably less problematic. The technical case is stronger than the commercial one at present. 

How power markets are pricing in this new demand 

Forward curves have begun to reflect this. Sustained demand growth in a system with tight capacity margins supports higher prices across the curve, and in markets where growth is concentrated, the effect is visible in longer-dated contracts. 

The peak-to-off-peak spread is where the effect is most interesting. A flat load adds as much consumption at 3am as at 6pm, which lifts off-peak prices disproportionately and compresses the spread. That has direct consequences for assets that earn from volatility rather than price level, as discussed in What is revenue stacking? A beginner’s guide to how battery storage makes money 

Corporate procurement has moved fastest. Hyperscalers have become among the largest buyers of long-term power contracts in Europe, with a strong preference for round-the-clock carbon-free supply over annually matched renewable volumes. That preference has drawn them towards firm low-carbon generation, including substantial commitments to new nuclear capacity. 

Capacity markets are the third channel. When data centre growth tightens adequacy margins, the volume of firm capacity that must be procured rises, and clearing prices follow. Ireland’s situation illustrates this link directly: demand growth concentrated in one region raised adequacy concerns that fed straight into capacity procurement. 

AI has converted data centres from a well behaved, efficiency offset load into the fastest growing source of electricity demand in several European markets, arriving at a speed networks were never designed to match. 

The next developments largely hinge on where the load is situated. Key factors to monitor include whether connection reform can accommodate substantial demand without displacing necessary generation projects, whether siting decisions persist in shifting away from the traditional constrained cluster, and to what extent this demand becomes a lasting structural change on forward curves versus a temporary cyclical build-up. 

The last of these is the genuinely open question. Demand forecasts have been repeatedly revised upward, and forecasting a technology cycle is a different exercise from forecasting electrification. Anyone positioning around this load should hold the projections more loosely than the confidence with which they are usually presented.