Europe can no longer afford to be slow. Too often, Europeans seem resigned to a pace of political decision-making that once symbolised deliberation and prudence but now means stagnation. Slowness is not harmless—it directly undermines European security, prosperity and political stability. The war in Ukraine is a painful reminder: Europe has struggled to ramp up weapons production at the scale or speed needed to support Kyiv or its own defence capacity effectively. Europeans’ inability to act quickly has strategic consequences: slowness erodes not just the continent’s autonomy; it effectively prolongs and aggravates a situation whereby the “kill switches” to Europe’s peace, energy security, economic competitiveness and political stability remain within reach of increasingly hostile powers and private players.
The same inertia is visible in the green transition. Europe once led in environmental innovation, yet it is now losing ground in electric vehicle production and renewable energy. Former European Central Bank president Mario Draghi noted that Europe is in danger of losing up to ten times more renewable energy generation than it loses today because of grid constraint capacity—a symptom of bureaucratic gridlock and insufficient investment coordination. Spain, for instance, has had to cap new energy production because it lacks grid capacity. Meanwhile, high energy costs driven by inefficiencies are deepening voter frustration and fuelling the rise of populist and authoritarian movements that promise action at any cost.
This sluggishness now threatens Europe’s position in the next great arena: artificial intelligence. Competing in AI is not just about algorithms, rare earths and microchips—it requires vast amounts of affordable and reliable electricity. For decades, on both sides of the Atlantic, electricity demand was broadly flat. That era is over. With the rapid deployment of large-scale AI systems, power demand is accelerating sharply.
In the US, electricity demand growth since 2021 has averaged around 1.7% a year, with the overwhelming share of incremental demand coming from the buildout of data centres linked to AI and advanced cloud computing. What may well be moderate projections from the US Department of Energy suggest that data centre power use could rise from around 4% of American electricity demand in 2024 to around 9% by 2030. In some parts of the US, increased demand from data centres used for AI has pushed up consumer electricity prices. In response, US utilities and technology companies are racing to add capacity through a mix of natural gas, renewables and nuclear power.
This is not only an American phenomenon. In its public calls for urgent action on infrastructure, OpenAI has explicitly pointed to the scale at which China is expanding its power system. In 2024 alone, China added 429GW of new generation capacity—more than one third of the entire installed capacity of the US grid—while America added closer to 50GW. Whatever the precise mix of technologies, the geopolitical significance is clear: countries that can build power systems the fastest will shape the AI era. Those that do not will find themselves losing not just economic leverage, but the capacity to sustain their standard of living and the quality of their public services.
Against this backdrop, Europe risks becoming an also-ran. This danger is compounded by the fact that the continent is relatively resource-constrained, with limited domestic sources of oil, gas and coal. At the same time, many member states have made nuclear power politically or administratively difficult to deploy. Meanwhile, as the Draghi report on European competitiveness has underlined, even with strong political support for the European Green Deal, renewable deployment has been slowed by complex planning, permitting and grid-connection procedures. These frictions raise both the cost of capital and the time required to deliver new capacity—often by many years—before even considering the supply chain risks associated with critical raw materials.
This paper argues that, despite the execution failures of recent EU energy policy, Europe has real structural advantages: a deep research base, a large pool of technology companies, an industrial sector bigger than America’s and a single market of more than 450 million relatively wealthy consumers. If Europe uses the AI challenge to force through a programme of “fast energy”—faster permitting, faster grid buildout, faster deployment of clean and firm power—it will strengthen its energy security, lower its long-term energy costs, accelerate decarbonisation and remain competitive in the AI age.
Energy and the AI revolution: The critical factor
The inevitable future
The scale of investment flowing into AI reflects a shared judgment by governments and firms in America and China: AI will be central to future economic growth and geopolitical power. Its effects are already visible in defence—for example, AI-enabled systems deployed in Ukraine; in health, where AI is used to search vast pharmaceutical libraries; and in manufacturing, with highly automated “dark factories” in East Asia. At the physical core of this transformation sit hyperscale data centres. Running them requires enormous and continuous amounts of electricity.
The energy implications of AI-focused data centres differ fundamentally from those of the earlier cloud-computing wave. Between around 2005 and 2017, the number of data centres grew rapidly, yet total electricity use remained relatively flat because efficiency gains offset growth. Cloud services primarily store and move data. By contrast, AI computes intensively—and computation is hungry for energy.
A commonly cited figure is that an AI query requires ten times the electricity of a standard web search. Even that shorthand understates the complexity of the issue. First, energy use varies widely depending on the time of day, the model and the nature of the task—text, images or especially video. Second, unlike mature cloud services, AI systems are still on a steep improvement curve: with more parameters, more training, more inference and therefore structurally rising energy demand.
A simple illustration shows how quickly this scales: consider a short, low-quality, AI-generated video clip—say, a few seconds at very modest resolution. Estimates suggest that even such a basic request can consume several million joules of energy: enough to run a microwave for an hour or power an e-bike for dozens of kilometres. High-definition, photorealistic video requires far more. If such capabilities become cheap and ubiquitous—as consumer technology historically tends to—usage will surge, and so will electricity demand. Crucially, most AI-related energy use does not come from training large language models, but from inference, ie, the day-to-day use of AI by consumers and firms. Various estimates put inference at 80-90% of total AI compute demand. As models improve and are embedded into more products and services, usage will only grow. The direction of travel is towards greater complexity, more personalisation and more continuous interaction—all of which are energy-intensive.
Assessing the precise trajectory is difficult because the industry is opaque. Most leading systems are closed (the data sources are opaque), and detailed energy data are not public. Researchers therefore rely on open models, such as Meta’s Llama, as proxies, which may not be representative. Even so, the trend towards significantly greater energy use remains clear.
In 2024, data centres accounted for roughly 4% of American electricity consumption. Some conservative official estimates suggest this could rise to around 9% by 2030, with half of that power going specifically to AI workloads. Other estimates provide a range, the lower end of which is consonant with reports from official sources. These estimates indicate that electricity used for AI alone could reach the order of 165-325TWh a year by 2028. In contrast, the International Energy Agency estimates that the 2024 consumption of all American data centres was just over 180TWh.
Investment plans are consistent with these numbers. The US government-backed “Stargate” AI infrastructure investment initiative envisages around $500bn invested in a small number of new hyperscale sites. Each site will potentially require several gigawatts of dedicated power—comparable, individually, to the average load of a small European country. Apple has announced plans to spend hundreds of billions of dollars on data and manufacturing facilities over the coming years. Globally, investment banks such as UBS estimate annual data-centre investment will approach half a trillion dollars by the late-2020s.[1]
About the Authors:
Nicu Popescu is co-director of the European Security Programme and distinguished policy fellow at the European Council on Foreign Relations, based in the Paris office.