Powering Intelligence: Energy as the Binding Constraint on AI

Every conversation about AI infrastructure eventually arrives at the same place: a substation. This Deep Dive follows the electrons from generation to the rack, identifies where the constraints are, and maps the suppliers, buyers and policymakers deciding how the AI economy gets powered.

1. The demand picture

Data centres were a modest share of electricity demand for two decades because efficiency gains offset growth. AI changed the slope. Accelerator-dense clusters draw far more per rack, and the campuses being planned are measured in hundreds of megawatts to gigawatts. The International Energy Agency and grid operators in affected regions have documented data centres as a fast-growing component of load, concentrated geographically. Precise forecasts vary widely; the direction does not.

2. The supply response

Three responses are visible. Contracting: technology companies sign long-term agreements for nuclear, gas and renewable output, sometimes underwriting restarts or new builds. Co-location: campuses are built next to generation or with on-site plants to avoid the interconnection queue. Geography: capital flows to regions with surplus clean firm power, from the Nordics to parts of the American interior and the Gulf.

3. Who pays

Grid upgrades to serve large loads are expensive, and the question of who bears the cost is now political. Utilities argue that new load spreads fixed costs and lowers rates for everyone. Consumer advocates worry that speculative projects leave ratepayers holding the bill. Regulators are writing large-load tariffs that require developers to commit capital up front. The outcome will shape where AI is built as much as any technology decision.

4. The technology

On the supply side: gas turbines for speed, nuclear restarts and small modular reactor proposals for firm clean power, renewables paired with storage for cost. On the demand side: liquid cooling to reduce overhead, higher-voltage distribution inside the campus, and software that schedules flexible workloads around power availability and price. Grid-enhancing technologies and long-duration storage sit in between.

5. Market map

The value chain runs from generation developers and utilities through transmission and equipment makers to the data centre developers, hyperscalers and the accelerator and cooling suppliers inside the building. Equipment makers with long lead-time products, such as transformers and switchgear, hold unusual pricing power in this cycle.

6. Risks

• Stranded capacity if AI demand disappoints or shifts geographically.
• Emissions growth if gas fills the gap that clean firm power cannot.
• Political backlash if residential rates rise.
• Water and local environmental constraints on siting.
• Supply chain concentration in critical equipment.

7. Opportunities

• Clean firm power developers and nuclear supply chains.
• Grid equipment manufacturers and grid-enhancing technologies.
• Storage and flexibility providers.
• Regions marketing surplus power to AI developers.
• Software for power-aware compute scheduling.

8. What happens next

Expect large-load tariffs to spread, behind-the-meter projects to multiply, and the first public disputes over who pays for grid expansion. Flexible compute will move from concept to contracts. And the split between training sites, which follow cheap power, and inference sites, which follow users, will become visible on the map.