AI Infrastructure: Why Energy Is Becoming a Bigger Bottleneck Than Chips

For years, the assumption was: whoever owns the best AI chips wins. But the real bottleneck is shifting. Energy is becoming the limiting factor—not only for hyperscalers, but for every company serious about scaling AI. The question is no longer just which model you choose, but where and how you run the compute.

Data Centre Power: The Numbers Behind the Bottleneck

The International Energy Agency (IEA) expects data centre electricity consumption to more than double from 460 terawatt-hours in 2022 to as much as 1,050 TWh by 2026—more than Germany's total annual electricity use. For the US, the IEA projects a 130 percent increase by 2030 (+240 TWh), for China +170 percent (+175 TWh), and for Europe +70 percent (+45 TWh). Energy is becoming the constraint faster than chip production.

China, the US, and Europe: Three Different Speeds

China produced roughly 10,000 terawatt-hours of electricity in 2024—more than twice the US and about four times the EU. US electricity consumption was flat for two decades; AI is now driving growth several times faster. Over the last decade, China built 27 nuclear reactors while the US built only two. Permits for power plants and high-voltage lines in the US often take more than ten years, while China added over 400 gigawatts of new capacity in a single year.

• China: national grid and ultra-high-voltage lines provide stability; data centres are set to grow from 170 TWh (2025) to a planned 800 TWh by 2030—around 6 percent of national consumption.
• US: rapidly expanding data centre boom, but dependence on Chinese transformers and switchgear—already a 15 percent supply bottleneck in 2026.
• Europe: lagging behind in data centre construction; Ireland has more than 20 percent of electricity going to data centres, Dublin has paused permits, and the Netherlands and Denmark restrict growth.

What This Means for Swiss Decision-Makers

Switzerland has a structural advantage with hydropower and nuclear energy—but local data centre capacity is limited. Scaling AI requires a hybrid strategy: Swiss edge infrastructure for latency-sensitive and sensitive workloads, cloud locations in the EU or US for compute-intensive training and bursts. Energy availability becomes a vendor selection criterion. Which provider has long-term access to affordable and sustainable power? How transparent are energy costs? The large infrastructure alliances—such as the 30-billion-dollar fund from Microsoft and BlackRock or reactivated power plants for AI—show how much energy now shapes the agenda. What Google, Alibaba, Microsoft, and BlackRock are really planning makes that shift concrete.

The First Step: Make Energy Visible in AI Planning

Don't start with model selection. Start with a simple inventory: Where do your AI workloads run today? What is the expected power demand as you scale? Which cloud regions are viable? These questions often change the business case more than a new model. The full implementation of an energy-aware AI architecture depends on your specific setup—that is exactly where a structured conversation begins.

For a closer look at how much energy and water AI actually consumes, this analysis of AI's energy hunger and climate impact provides context—including why ChatGPT requires about 2.8 million litres of water per day.