In September 2024, Microsoft signed a deal to restart Three Mile Island, a nuclear plant shuttered since 2019. The press called it an energy story. It was an infrastructure story: Microsoft needed power for data centers and could not get it from the grid fast enough.
That gap between when AI infrastructure needs power and when the grid can supply it is the constraint shaping everything right now. Not chips. Not money. The wire from the substation to the building.
The numbers make the picture concrete. ERCOT, the Texas grid, has 233-plus gigawatts of large-load requests queued for interconnection. Projects entering PJM, the largest U.S. grid operator, wait more than seven years on average before getting power. Globally, the IEA counts 2,500-plus gigawatts of generation stalled in interconnection backlogs. Total U.S. generation capacity is roughly 1,200 gigawatts. The queue is twice the grid.
What the panel sees: Three of four analytical frameworks predict that power and grid interconnection becomes the single binding constraint on AI deployment by 2027-28 (35% probability). Another 25% says memory friction persists alongside the grid problem, a two-front bind. Together, those paths carry a 60% probability that power is part of the constraint picture regardless of how chips resolve. The remaining 40% splits between an efficiency-driven demand slowdown (20%) and an architecture or geopolitical shock that reshuffles the stack (20%).
| Scenario | Probability |
|---|---|
| Grid Claims the Cascade: power/interconnection dominant by 2028 | 35% |
| Multi-Front Deadlock: HBM friction + grid bind simultaneously | 25% |
| Efficiency Dampens the Cascade: AI demand decelerates before 2028 | 20% |
| Contingency Reshuffles the Stack: architecture or geopolitical pivot | 20% |
No calls have resolved. This is the panel's opening position.
How the constraint got here
Figure 1: The constraint migration, 2020-2026.
The constraint on AI infrastructure has been migrating up the stack since 2020. Foundry capacity first, then packaging and memory, then grid and power. Each time one layer cleared, it accelerated arrivals at the next. The cascade has not resolved. It has compounded.
The dissent worth taking seriously
Wardley Mapping (Wardley) argues that when a mature utility service is constrained, markets find workarounds. Hyperscalers have already started: Microsoft restarted a nuclear plant, Meta is building a 6.6-gigawatt nuclear portfolio, and brownfield industrial sites with existing power are being bought up. Wardley's prediction is that these mechanisms will partially circumvent the grid constraint before it reaches maximum severity, making the clean single-bottleneck outcome less decisive than the other three frameworks suggest.
What to watch
Two early checks matter. By October 2026, HBM lead times should show signs of normalizing (target: below 26 weeks by Q1 2027), and CoWoS should be tracking toward a supply gap below 5% (target: Q2 2027). If both clear on schedule, the freed capacity flows straight into the grid queue and the 35% scenario gains weight. If HBM stays stuck, the 25% two-front scenario becomes more plausible.
The grid check comes later: FERC/ISO median approval times falling below three years would be the first sign of real institutional relief, with a first read in January 2027.
The queue is twice the grid. That sentence is the forecast.