The Interconnection Queue Isn't Saving the Grid - It's Stranding Hyperscaler Capex

The headline says slow data center hookups offer summer relief for US power grids. It should say: slow data center hookups are about to turn hundreds of billions in AI infrastructure capex into stranded assets. Both are technically true. One matters for your portfolio.

The "relief" is an execution failure disguised as a grid win

Here's what actually happened. NERC's January 2026 Long-Term Reliability Assessment revised peak demand forecasts upward by 24% over the next decade. Summer peak demand is projected to climb from 127 gigawatts in 2026 to 143.7 gigawatts by 2035. NERC flagged MISO, PJM, and ERCOT as the highest-risk regions - the same regions where hyperscalers have been burning cash on data center land and shell buildings.

Meanwhile, the average time from interconnection application to commercial operation has ballooned from under two years in 2008 to over eight years today. AI data center projects that finally reached service in 2025 took more than seven years on average, according to PJM grid data. If a data center relies solely on the existing grid for power, energization can take up to ten years in some regions.

That is not a grid victory. That is a buildout failure.

The grid isn't being saved because data centers are getting more efficient or because demand is lower than expected. The grid is being saved because the hyperscalers simply can't plug in. The interconnection queue is acting as an accidental circuit breaker - and it is about to blow through the AI capex narrative.

The numbers that invert the "relief" story

Nearly 2,300 gigawatts of projects are sitting in interconnection queues nationwide. Of the roughly 12 gigawatts of new US data center capacity planned for 2026, approximately 7 gigawatts - close to half - are expected to be delayed or canceled. That Bloomberg estimate from April is not a rounding error. It means the compute schedules every hyperscaler CFO has presented to their board are built on delivery timelines that are physically impossible.

Microsoft canceled data center leases with at least two private operators in the US, totaling hundreds of megawatts, and blamed power delays. Whether you read that as honest engineering reality or strategic cover for a scaling problem, the result is the same: capex that was already committed is being unwound because electrons can't arrive on schedule.

This is the unit-economic inversion the "grid relief" headline skips. The interconnection delay doesn't reduce the grid's future problem. It defers it. Every megawatt sitting in queue for eight years is a megawatt that will eventually arrive on a grid that hasn't had eight years to build transmission capacity. The bottleneck compresses, not dissolves.

What the market is still getting wrong

Three things the consensus narrative has backwards:

The queue is a temporary problem. It is structural. The grid interconnection process wasn't designed for a demand surge of this magnitude - summer peak demand growth of 224 gigawatts over the decade, per NERC, represents a 69% increase over the 2024 forecast. That is not a backlog that clears with a few new substations. It is a system that would need to be rebuilt faster than political, environmental, and permitting realities allow.

Hyperscalers will build their own generation. They are trying - coal plants, nuclear SMRs, natural gas peakers - but those projects face their own interconnection and permitting timelines. The energy infrastructure buildout has the same delay mechanics as the load it's supposed to serve. You cannot shortcut a ten-year queue by building the generator yourself if the transmission line to it is also in the queue.

This only matters for data center operators. It matters for every chip company whose revenue depends on these racks being powered and running. Nvidia's roadmap assumes thousands of Rubin and Vera racks deployed on schedule. AMD's MI450 thesis assumes customers can absorb the hardware into live data halls. If 7 gigawatts of 2026 capacity gets pushed back or canceled, those revenue schedules need to be rewritten. Not the chips. The power.

The investor-grade implication

The cross-currents are clear. On the grid side, NERC is explicit that mutual aid between regions may not be sufficient during simultaneous emergencies - meaning PJM can't necessarily borrow from MISO when both are stressed. On the hyperscaler side, the capex commitments are larger than the delivery pipeline can support. On the semiconductor side, the chip supply story has overtaken the chip demand story as the constraint.

The interconnection queue is the single largest unpriced risk in the AI infrastructure trade. Every stock that depends on data center deployment timelines - NVDA, AMD, VRT, EQIX, DLR, the whole supply chain - is trading as if the electrons will appear when the racks do. They won't.

That doesn't mean the AI infrastructure buildout is dead. It means the schedule is the product, and the schedule is broken. The companies that can navigate the queue - through direct utility partnerships, on-site generation with fast-track permits, or geographic arbitrage into less-constrained regions - will capture disproportionate market share. The ones that assumed power was a solved problem will write down commitments they can't fulfill.

The summer grid relief headline is real. It's also a trap. The relief is temporary, the deferral is certain, and when that queue eventually works through - or collapses into cancellations - the market will reprice every assumption about how fast AI compute can scale. The question is no longer whether data centers are the future of electricity demand. It's whether the future can wait eight years.