AI data centres: Powering through the constraints

The US power system may support substantial AI data-centre growth through 2030, but project delivery will be constrained more by transmission, equipment, labour and political factors than by generation capacity alone.

• Columbia Threadneedle’s bottom-up model uses announced utility and independent-producer projects through 2030, plus improvements in transmission, power-flow management, storage optimisation and grid utilisation.
• Natural-gas generation and grid-optimisation measures are expected to provide roughly 70% of incremental US power supply through 2030, while nuclear power is not expected to have a material impact before then.
• Assuming 65% of incremental capacity is available for data centres, the grid could theoretically support about 31 GW of additional data-centre capacity in 2026 and around 169 GW cumulatively by 2030.
• The most severe bottlenecks are community and political opposition, EPC and construction labour, and power transformers; gas turbines, grid interconnection and 800V DC architecture are significant constraints, while cooling and electrical equipment constraints are still emerging.
• Behind-the-meter solutions such as gas turbines, reciprocating engines, fuel cells and batteries can bridge delayed grid connections; fuel cells can be deployed in under a year but are more expensive and mainly serve requirements below 100 megawatts.

Takeaway: Investors should look beyond electricity generation to regulated utilities, electrical-equipment and cooling suppliers, construction companies and hybrid-power providers, while recognising that the ultimate scale and timing of the AI build-out remain uncertain.