AI and the Grid: The Defining Interplay of the Decade
Shubham Yadav, Strategy Lead, Esyasoft
Where the world's most disruptive technology meets its most constrained piece of infrastructure
Part 1/7 — The Grid is the Bottleneck
For two decades, the energy transition was a story about generation. The world has answered emphatically and the solar costs are down more than 80% in a decade[1]. Renewables now make up most of the new capacity added each year. Battery storage is scaling beyond every forecast. And yet the transition is slowing on the measure that increasingly matters - getting clean electrons to where they are needed, in time. The constraint has shifted. The world is building clean energy faster than it can move, manage, or connect it to demand.
**The grid has become a bottleneck.**And it is the kind of bottleneck that does not yield to faster construction or more capital alone.
Two forces, one strained system
Onto a system designed for one-way flows from a small number of large generators, two forces are now landing simultaneously - each faster than the grid can adapt.
The first is the renewables backlog. In the United States alone, more than 2000+ GW of clean generation and storage (roughly twice the size of the entire existing US power system) sits in interconnection queues[2]. Average wait times have climbed from under two years in 2008 to nearly five years today; in California, it is the highest[2]. When PJM, the grid operator covering 13 US states, opened its first reformed queue cycle in April 2025, it received 800+ applications totaling 220 GW in a single intake window[3]. Construction-ready capital sits in suspension, waiting not for money or technology but for the engineering capacity to integrate it.
The second is the AI demand shock. A single AI training campus now consumes one to two gigawatts, which is the peak demand of a small country. In Northern Virginia, peak load could nearly double over the next 15 years, driven almost entirely by data centres[4]. Ireland's data centres already consume 20+% of national electricity, prompting EirGrid to pause new Dublin connections[5]. AI demand is unlike past load growth in three ways: it is geographically concentrated, it arrives in eighteen months rather than years, and it is willing to pay extraordinary premiums for speed. The clearest signal of how acute this has become is that nuclear power plants once on the path to retirement are now being brought back specifically to serve AI. Microsoft has contracted to restart Three Mile Island[6]. Amazon has acquired a data centre campus adjacent to the Susquehanna nuclear plant for direct nuclear power[7]. In the Netherlands, TenneT has begun selling time-dependent connection contracts to large loads willing to accept curtailment.
Why does the bottleneck not yield easily
The bottleneck is structural. A new high-voltage transmission line takes seven to ten years in the US, ten to fifteen in Europe and most of it is permitting, not engineering[8]. Transformers have multi-year manufacturing backlogs. The specialised workforce is in short supply globally. More capital does not shrink any of these timelines.
This reshapes the strategic landscape. For hyperscalers, grid access has become the most important variable in AI infrastructure planning - more than chips, more than talent. For industrial decarbonisers, the grid determines whether electrification plans are executable at all. For utilities, the operating model itself is being rewritten under pressure. For governments, electricity capacity is becoming inseparable from AI competitiveness and national security.
How the bottleneck reaches you
The AI-driven demand shock propagates through every category of electricity user, in different forms.
For large industrial customers, the constraint is becoming existential. Sectors building decarbonisation roadmaps around electrification (steel, cement, ammonia, chemicals, etc.), are discovering that the clean power they planned around may not be deliverable on their investment horizons, in part because AI compute is competing for the same scarce grid capacity. New industrial sites are being routed away from preferred locations, while data centre developers willing to pay premiums move to the front of the queue[9]. The result is a quiet reshaping of industrial geography around where grid capacity remains available.
For commercial and mid-sized business customers, the impact shows up as rising wholesale prices in regions absorbing AI demand, longer connection lead times, and increasingly complex tariffs as utilities introduce demand charges and flexibility requirements. What was once a back-office utility bill is becoming a strategic variable in site selection and operations.
For retail and residential consumers, the bottleneck arrives more slowly but more visibly. Electricity prices in affected regions have begun to rise, and several US states have seen utility filings explicitly citing data centre load growth as a driver of rate cases. The political question of who pays for the grid investments AI demand requires - hyperscalers, industrial users, or households - is becoming very contested.
For hyperscalers themselves, being a grid customer is no longer simply about buying power. It now involves long-dated agreements with new projects, co-investment in transmission, and negotiations with regulators over connection rights, curtailment, and tariff design. The largest AI buyers have moved upstream into the energy system in ways the frameworks built around traditional load were never designed to handle.
Across all these segments, the pattern is the same. The grid bottleneck, intensified by AI, is no longer happening upstream of the customer. It is actively shaping which industries can grow, where they can locate, and at what cost. For a serious operator in any sector, understanding the grid-AI interplay is becoming as important as understanding their market.
The bottleneck is the prerequisite
What makes the moment unusual is that the same technology now intensifying the bottleneck, i.e. AI, is also emerging as the most powerful tool we have ever had to relieve it. None of this removes the need to build new wires. But it changes what existing infrastructure can do, how quickly new infrastructure can connect, and how flexibly the system can respond.
In a world where physical grid expansion takes a decade and demand arrives in months. The question is no longer only how fast we can build. It is how intelligent we can make what we already have. And the first place that intelligence is quietly beginning to show up is at the very gate of the grid - at the door through which everything new must pass.
References:
[1] International Renewable Energy Agency (IRENA), Renewable Power Generation Costs in 2023 (September 2024) on long-term solar PV cost declines. https://www.irena.org/Publications/2024/Sep/Renewable-Power-Generation-Costs-in-2023 [2] Lawrence Berkeley National Laboratory, Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2024 (April 2025) on US interconnection queue capacity and average wait times. https://emp.lbl.gov/queues [3] PJM Interconnection, 2025 Cycle 1 Application Window Update on the first reformed queue cycle volume. https://insidelines.pjm.com/over-800-new-generation-projects-seek-to-connect-under-pjms-reformed-process/ [4] Dominion Energy, 2024 Integrated Resource Plan on projected Northern Virginia peak load growth [5] Central Statistics Office Ireland, Data Centres Metered Electricity Consumption 2023 on Irish data centre electricity consumption share. https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2023/ [6] Constellation Energy, Crane Clean Energy Center Announcement (September 20, 2024) on the Three Mile Island restart agreement with Microsoft. https://www.constellationenergy.com/newsroom/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html [7] S&P Global Market Intelligence, Talen Energy sells Pa. datacenter campus to Amazon Web Services for $650M (March 2024) on the Susquehanna-adjacent data centre campus acquisition. https://www.spglobal.com/market-intelligence/en/news-insights/articles/2024/3/talen-energy-sells-pa-datacenter-campus-to-amazon-web-services-for-650m-80711401 [8] International Energy Agency, Average Lead Times to Build New Electricity Grid Assets in Europe and the United States, 2010–2021 on transmission infrastructure build times. https://www.iea.org/data-and-statistics/charts/average-lead-times-to-build-new-electricity-grid-assets-in-europe-and-the-united-states-2010-2021 [9] Forbes, As AI Booms, Data Centers May Create Electricity Scarcity Among Users (December 15, 2025) on competition for grid capacity between data centres and other industrial loads. https://www.forbes.com/sites/kensilverstein/2025/12/15/as-ai-booms-data-centers-may-create-electricity-scarcity-among-users/
