06/01/2026
AI isn’t running on magic. It’s running on megawatts + water.
Here are the numbers defining the Data Center Boom (2015–2025):
Global capacity mix (Q1 2025): Hyperscale 44% | Colocation 22% | On-prem 34%
Hyperscale sites: 1,136 live (end-2024)
Concentration: Top 20 markets = 62% of global hyperscale capacity
Global electricity: 460 TWh (2022) → >1,000 TWh (2026 est.)
U.S. electricity: ~60 TWh (2014–2016) → 176 TWh (2023) → 325–580 TWh (2028 range)
Ireland “grid share” signal: data centers 5% (2015) → 21% (2023) of metered electricity
Water reality (U.S.): 21.2B L (2014) → 66B L (2023) direct water, plus ~800B L (2023) indirect water via electricity
Efficiency isn’t the full story: industry avg PUE 1.65 (2014) → 1.56 (2024) (improved, then plateaued)
AI changes the load shape: inference ~60% (2023) → training ~50–53% (2028) of AI server energy (U.S. estimate)
3 takeaways:
The constraint is shifting from real estate to power availability + grid interconnection.
Water impact isn’t just onsite cooling—the electricity supply chain can dominate.
AI pushes density, accelerating adoption of liquid cooling and smarter siting.
Sources: IEA, LBNL, Uptime Institute, Ireland CSO, Synergy Research Group, CBRE, The Green Grid.
CTA: Comment “DC” if you want the full 7-section infographic version.