All stories
The Power Crunch: When AI's Bottleneck Became the Electrical Grid (2025)  -  Moments | AI History by CobuildX
CobuildX
2025Moments

The Power Crunch: When AI's Bottleneck Became the Electrical Grid

By 2025, the limiting factor on how much AI the world could build was no longer chips or data. It was how fast you could get a gigawatt of electricity to a data centre.

AI data centersStargate projectpower gridAI infrastructurenuclear power AIcompute bottleneckhyperscalers

Through 2024 and 2025, OpenAI, Microsoft, Google, Amazon, and Meta collectively committed hundreds of billions of dollars to new data centre capacity. OpenAI's Stargate project, backed by Oracle and SoftBank, alone announced plans for up to $500 billion in US AI infrastructure investment over several years, while utilities in Virginia, Texas, and Georgia began reporting that data centre power demand requests exceeded the capacity their grids could actually deliver.

A single frontier training cluster can draw hundreds of megawatts, comparable to a mid-sized city, and training runs need that power delivered reliably and continuously for months at a time. Unlike a factory, a GPU cluster cannot easily shift its load to off-peak hours, so data centres started competing directly with residential and industrial users for scarce transmission capacity, not just for new generation.

Hyperscalers responded by signing direct deals with power generators to bypass the grid bottleneck entirely, including deals with nuclear operators: Microsoft agreed to fund the restart of a reactor at the former Three Mile Island site, and several companies invested directly in small modular reactor startups to build power dedicated to a single data centre campus. Chip supply, the binding constraint of 2023, increasingly took a back seat to permitting and grid interconnection timelines, which run on a scale of years, not quarters.

Many had assumed the ceiling on AI's growth would be set by chip availability, the subject of this archive's CUDA-GPU story, or by training costs falling further, as told in the DeepSeek R1 story elsewhere in this archive. Instead, physical infrastructure, transformers, transmission lines, cooling water, and interconnection queues became the binding constraint, exactly the kind of problem no amount of algorithmic efficiency can route around.

The industry's growth curve is now explicitly tied to energy infrastructure buildout timelines measured in years, reintroducing a slow, capital-intensive, physical constraint into an industry that had spent the better part of a decade bottlenecked mostly on software and silicon.

What it actually meant

AI's growth hit a constraint that software cannot optimise away. Algorithmic efficiency, better chips, and cheaper training all help, but electrical grid capacity and interconnection timelines are governed by physical construction schedules, not Moore's Law or scaling laws.

What enterprises get wrong today

Enterprises planning AI-dependent products assume compute availability and API pricing will keep improving on the curve of the past few years, and build roadmaps that assume unconstrained capacity. Energy and compute scarcity can reintroduce latency, cost spikes, and rationing a purely software-side plan never accounts for.

Related capabilities

AI History

Ask about this story

Explore this history

By 2025, the thing actually limiting how much AI infrastructure companies could build wasn't chips, it was electricity. Want to know how bad the power crunch got, or what hyperscalers are doing about it?