Seventy percent of the grid is near end of life. AI arrived at the worst possible moment.

Reporting on grid readiness puts roughly 70% of infrastructure approaching the end of its life cycle, at the same moment data centre demand is projected to rise from about 180 TWh toward 400-600 TWh by the end of the decade.

Either of those alone is a manageable engineering programme. Together they compound, and the compounding is what matters for anyone planning capacity.

Replacement and expansion compete

Grid operators have finite crews, finite equipment lead times and finite outage windows. Every hour spent replacing aging transmission is an hour not spent connecting new load, and the two draw on the same constrained pool of transformers, switchgear and skilled labour.

This is why interconnection queues lengthen even where there is political will and capital. The constraint is not primarily money or permission. It is physical throughput in an industry that was sized for steady replacement, not simultaneous replacement and expansion.

The queue is not a policy failure. It is a supply chain running at capacity on two jobs at once.

What it means for reliability, not just availability

The availability conversation dominates, but aging infrastructure carrying record concentrated load is also a reliability question. Systems designed for diffuse, predictable demand behave differently under large step changes, and the margin for absorbing a fault narrows.

For organisations running consequential workloads, that argues for treating power as a dependency with a realistic failure model rather than an assumed constant. What is the actual reliability of the supply at your site? What is the failover, how long does it hold, and has it been exercised under load rather than tested on a schedule?

The planning horizon problem

The mismatch that catches people is temporal. AI capacity decisions run on eighteen-month cycles; grid infrastructure runs on five-to-fifteen-year ones. A deployment plan that assumes power arrives when needed is implicitly assuming the slower system will accommodate the faster one.

It will not, and the organisations that adjust first will be the ones that put interconnection status into the same review as model selection — which sounds absurd until the first programme slips a year waiting for a substation.

Written by the Devence Lab research team.