Technology · Issue 01
The power behind intelligence
Data centres used 485 TWh in 2025. The IEA’s base case still nearly doubles that by 2030 — and AI sites are growing three times as fast as the rest.
- AI-accelerated load
- Conventional & cooling
- Data-centre electricity, 2025
- 485 TWh
- All data-centre demand, 2025
- +17%
- AI-focused sites, 2025
- +50%
- IEA base case, 2030
- ~950 TWh
A small share of power, a large share of new demand
Data centres are still a few percent of world electricity — 1.5% in 2024, heading toward about 3% in 2030 on the IEA’s base case. That sentence is true and misleading. The IEA’s 2026 update finds consumption rose 17% in 2025, to 485 TWh, and AI-focused sites surged 50%. From 2024 to 2030 the agency still sees a doubling, from 415 TWh to around 945–950 TWh, growing four times faster than electricity demand in every other sector. The world can absorb 3%. Individual grids in northern Virginia, Dublin, Singapore and parts of Texas cannot.
Inference, not training, is the compounding machine
Training a frontier model is a headline: weeks of packed GPU halls. Inference is the business. As models become products — search, agents, coding, video — utilisation stops being a research spike and becomes a 24-hour load. The IEA attributes almost half of the 2024–2030 increase to accelerated servers, growing about 30% a year, versus 9% for conventional servers. Efficiency per query is falling; number of queries is not. That is the same pattern that made global compute explode even as chips became more efficient: Jevons, not conservation.
The bottleneck has moved off the silicon
The 2026 IEA note is unusually blunt: bottlenecks across the energy-equipment and chip chain are cutting the odds of the most aggressive near-term scenarios, even as investment and project pipelines boom. Transformers, turbines, interconnection studies and water rights now gate model deployment as much as HBM supply. The United States and China still account for nearly 80% of the 2030 increase. Europe grows from a smaller base and hits politics faster. This is why hyperscalers are signing nuclear restarts, behind-the-meter gas, and twenty-year offtakes that look more like aluminium smelters than software firms.
Read it as an energy story that happens to be about AI
If you treat AI as a software cycle, you will mis-time the capex. The long view is a new industrial load arriving in a system that spent two decades assuming demand was flat in the OECD. The investment map is therefore power first, chips second, applications third. A model that cannot be plugged in does not ship.
Investing lens
Horizon 5–12 years · Educational, not advice
The scarce goods in the AI buildout are electrons, interconnection, and memory — not another foundation-model wrapper. Position along the physical stack: generation with offtake, grid hardware, liquid cooling, and the memory/logic names actually shipping into accelerated servers.
Where the map points
- Independent power producers and nuclear operators with data-centre offtake
- Electrical equipment: transformers, switchgear, high-voltage cable
- HBM, foundry, and accelerated-server vendors with visible backlog
- Select gas and behind-the-meter generation where interconnection is the binding constraint
What can break it
- Efficiency and utilisation can break linear TWh extrapolations after 2030
- Permitting and community opposition strand announced campuses
- Capex concentration in a handful of hyperscalers makes the cycle binary
CHART does not recommend securities, funds or trades. Figures can be revised by their publishers. Do your own research and consider regulated advice before allocating capital.
Sources
Every headline number traces to a named publisher. Contextual sources inform the essay, not the key stat.
- 01PrimaryInternational Energy Agency2026Key Questions on Energy and AI — Executive summary
Updated path: 485 TWh in 2025, ~950 TWh in 2030. AI-focused sites +50% in 2025; all data centres +17%.
- 02PrimaryInternational Energy Agency2025Energy demand from AI
Base-case data-centre electricity: 415 TWh in 2024, ~945 TWh in 2030.
- 03CorroboratedInternational Energy Agency2025Electricity Mid-Year Update 2025
Global electricity demand +3.3% in 2025 and +3.7% in 2026; consumption over 29,000 TWh in 2026.
- 04ContextualBrookings Institution2026-04Global energy demands within the AI regulatory landscape
Independent reading of IEA, Deloitte and Goldman paths for data-centre power.
Keep reading
The chip supercycle
Semiconductors did $792 billion in 2025. WSTS’s spring 2026 forecast put 2026 at $1.51 trillion — a 90% leap, almost all memory and AI.
$1.51 tn
WSTS spring-2026 forecast for 2026 sales
The century-long reign of coal is ending
In 2025, renewable electricity generated more power than coal for the first time in the modern era. Solar did most of the work.
33.8%
Renewables share of world electricity, 2025
The nuclear return
438 reactors, 397 GWe, a record year of generation — and a political target to triple capacity by 2050. AI offtake is the new customer.
397 GWe
Operable nuclear capacity, Oct 2025
Data-centre water, in a field of almonds
Lawrence Berkeley counted 17.4 billion gallons of on-site water at US data centres in 2023. A 2026 path toward the lab’s 38–73 billion-gallon 2028 range still sits near 28. California almonds, from USDA acres at 3.0 acre-feet, sit near 1,355. US golf applied 531 in 2024. The campus can dry a local aquifer. It does not move a national ledger.
50×
CA almonds vs US data-centre on-site water, 2026 path
The other two hundred and eleven billion
On-site cooling is the political number: 17.4 billion gallons in 2023. The power plants that fed those racks used about 211 billion more. Data-centre water is mostly an electricity mix. Direct is the well. Indirect is the grid.
211
Billion gallons, US data-centre power-plant water, 2023