The consensus AI trade stops at silicon. The harder, more durable bottleneck sits one layer beneath the chips: electricity, specifically firm power delivered fast. A modern AI datacenter campus can demand as much continuous power as a small city, and it needs that power available 24/7, near the load, and energized on a two-to-four-year timeline. That is a physical problem, not a software one, and it is where the real supply constraint now lives.
Why gas turbines, and not the grid or clean power? Interconnection queues in the US and Europe stretch for years; utilities cannot add high-voltage capacity on a hyperscaler's schedule. Renewables are queue-bound and intermittent, batteries shift energy but do not generate it, and new nuclear, including SMRs, is a decade away at scale. Heavy-duty gas turbines are the only technology that can drop gigawatts of dispatchable, always-on power next to a datacenter within the window operators actually have. Hyperscalers and their utilities are responding exactly this way, ordering behind-the-meter and utility-scale gas capacity.