Power Is the New Bottleneck. AI’s 945 TWh Problem Reprices Energy, Compute and Capital (2026–2030)

The base case, from the IEA. Global data centres consumed about 415 TWh in 2024, 1.5% of world electricity, roughly Germany plus Austria. By 2030, that doubles to ~945 TWh, just under 3% of global consumption, growing ~15% per year, four times faster than all other sectors combined.

Inside that number, the split is the signal. Total data-centre demand grew 17% in 2025. AI-specific demand grew ~50% in the same year, and the IEA projects it to triple to ~465 TWh by 2030. AI servers are expanding at ~30% annually within the data-centre footprint, and Gartner expects AI-optimized servers to rise from 21% of data-centre power use in 2025 to 44% by 2030.

1. The money is already moving

Hyperscaler capex exceeded $400 billion in 2025 and the IEA projects roughly +75% in 2026, a scale now exceeding global oil and gas investment. Amazon alone has been reported at $200bn for 2026. This is no longer a software investment cycle; it is an energy and physical-infrastructure cycle.

2. Where the constraint binds

Morgan Stanley warns of a potential 20% US electricity shortfall for data centres by 2028, a 44 GW deficit. The United States hosts ~45% of global data-centre consumption (183 TWh in 2024); DOE/LBNL projections put US demand at 325–580 TWh by 2028, a range wide enough to be a planning crisis in itself.

The binding constraint is not capital, it is time. A data centre is operational in two to three years; grid interconnection, transmission and generation take far longer. Grid interconnection queues are the leading indicator to watch: the gap between announced and operational capacity is determined by permitting, not by money. State-level moratoria (e.g., Maine) are the first political response.

3. The second-order trades

Generation mix. Renewables are projected to meet roughly half of new data-centre demand growth to 2035 via corporate procurement, but the immediate gap is filled by natural gas turbines and nuclear restarts, with SMRs as the 2030s option. Long-duration PPAs become strategic assets, not procurement line items.

Cooling. GPUs above 700W and racks at 50–140 kW have ended the air-cooling era. Liquid cooling (direct-to-chip and immersion) grows from $870m (2024) to ~$10.7bn by 2030, a ~52% CAGR, covering a projected 76% of servers by 2026. Microfluidic cooling, etching coolant channels into silicon, removes heat 3x more effectively than cold plates in lab tests.

Geography. Power availability, not fibre, now selects sites: the US and China drive ~80% of growth to 2030 (US +240 TWh, China +175 TWh), while Europe grows a modest +45 TWh, a competitive and regulatory divergence with direct market consequences.

4. What this means for markets

The repricing runs in both directions. On the demand side: utilities with interconnection capacity, gas turbine manufacturers, nuclear restart stories, cooling pure-plays, and power-adjacent real estate reprice on scarcity. On the risk side: AI-capex equities carry an unpriced physical assumption — that the power arrives on schedule. A 44 GW deficit says it may not. The scenarios investors should be running treat energy delivery, not model capability, as the binding variable on AI earnings through 2030.

Energy is AI’s physical constraint. Model capability, compute architecture and software efficiency all operate inside an envelope set by how much power can be delivered, cooled and sustained. Capital alone does not compress a grid queue.

STEELLDY’s cross-asset stress framework tracks the energy-compute-capital nexus, from power-market tightness to liquidity stress. Methodology: steelldy-indices.com.