The Data Center Boom and the Race for Power

The AI data-center boom is creating a severe power bottleneck, making behind-the-meter generation, power equipment and gas infrastructure potential beneficiaries.

• Data-center construction is constrained by equipment lead times and grid access; nearly 2,600 GW of generation and storage capacity was stuck in the interconnection queue at the end of 2023, with typical waits exceeding four years.
• Morgan Stanley estimates that planned AI semiconductor sales require about 97 GW of power in 2026-2028, while only 40 GW is already being built or can draw on the grid, leaving a 57 GW shortfall.
• NVIDIA's Vera Rubin architecture delivers up to 10 times more inference throughput per watt than Blackwell, but lower computing costs are encouraging additional deployment; Morgan Stanley expects Feynman-era power density to raise U.S. data-center demand by another 40% in 2029.
• Political resistance is increasing: in August 2026, 54% of Republicans and 68% of Democrats said they would be uncomfortable with a new data center nearby, while the share viewing data centers as bad for household energy costs rose to 50% from 38% in January.
• Behind-the-meter solutions include fuel cells, reciprocating engines, gas turbines and dedicated gas supply; SemiAnalysis estimates that more than half of new U.S. data centers could use them from 2028 onward, often under contracts lasting at least 15 years.
• The power race is also geopolitical: BloombergNEF's base case has China adding almost six times as much power-generation capacity as the United States over the next five years, despite U.S. companies leading in AI-model development.

Takeaway: Investors should look beyond AI-chip makers toward the companies supplying onsite generation, power equipment and natural-gas infrastructure, although political opposition, grid constraints and the disputed effect on electricity bills remain important risks.