The energy demands of artificial intelligence are not a future challenge. They are an immediate structural disruption that no infrastructure developer, grid operator, or policymaker can afford to treat as a distant problem. The question for organisations building or advising on AI infrastructure today is not whether a power crisis is coming — it is how to plan, site, and finance compute facilities in an environment where grid capacity has become the single most consequential constraint on AI deployment.
This piece examines the scale of the demand shock, the structural reasons why conventional grid development cannot keep pace, the geographic concentration of risk, and the practical responses available to infrastructure developers and their advisers.
The Scale of the Demand Shock
The numbers are now well-documented, but they bear repeating in aggregate. A single hyperscale AI training campus — the kind being planned in Northern Virginia, Texas, and across the Gulf Cooperation Council — may draw 500 megawatts to more than a gigawatt of power continuously. To contextualise that: 500MW is roughly equivalent to the constant residential electricity demand of a city of 400,000 people.
Individual rack densities have also increased by an order of magnitude in three years. A conventional enterprise data centre rack draws 5–10 kilowatts. A modern GPU cluster configured for frontier AI training draws 80–200kW per rack, with leading hyperscalers planning for configurations beyond 300kW. The thermal engineering and power distribution implications of this shift are not incremental — they require fundamentally different facility designs.
What distinguishes AI infrastructure demand from previous waves of data centre construction is its speed and concentration. Cloud computing grew quickly, but it was distributed across hundreds of facilities and geographies. The current AI build-out is concentrated in a small number of very large campuses, all competing simultaneously for grid interconnection in markets where that interconnection queue stretches years into the future.
Why Grid Development Cannot Keep Pace
The fundamental problem is structural: transmission infrastructure is built on decade-long planning horizons, while AI infrastructure demand has materialised within a three-year window. No grid was designed with the expectation of gigawatt-scale single-site loads appearing with 18 months of notice.
The interconnection queue in the United States — the formal process by which a new large electricity customer secures a connection to the grid — has grown from approximately 1,000GW of pending requests in 2021 to over 2,500GW by late 2024. The median study and approval process now exceeds five years for large transmission-level interconnections. This is not a regulatory bottleneck that can be legislated away quickly; it reflects genuine physical constraints on transformer manufacturing capacity, skilled labour availability, and planning consent processes.
In Europe, the situation varies materially by market. The Netherlands has imposed a de facto moratorium on new large-scale data centre connections in certain regions of the national grid. Ireland — home to approximately 25% of European data centre capacity — has faced successive warnings from its national grid operator about the viability of continued growth at current rates. The Nordic markets, which offer genuine renewable power and cooler climates, have become acutely contested.
"The grid cannot be the afterthought in AI infrastructure planning. It must be the first question — site selection without a credible power pathway is not site selection at all."
Geographic Concentration and Its Consequences
Three US markets have absorbed a disproportionate share of AI infrastructure investment: Northern Virginia (Loudoun County and the surrounding area), the Dallas-Fort Worth metroplex, and the Phoenix metropolitan area. Each faces distinct but related challenges.
Northern Virginia — which hosts the largest concentration of data centre capacity of any single market on Earth — has effectively exhausted available grid capacity in its core zones. Dominion Energy, the primary utility, has published interconnection timelines of five to seven years for new large-load connections. Developers are now expanding into adjacent counties in Virginia and West Virginia, accepting longer fibre paths and less mature ecosystems in exchange for the possibility of near-term power.
In Texas, the ERCOT grid's independence from the wider US interconnected grid is both an advantage and a constraint. It allows faster interconnection in some circumstances, but it also means the grid cannot import power during periods of scarcity — a risk that became viscerally apparent during Winter Storm Uri in 2021 and which continues to be priced into power purchase agreement structures.
In Europe, the pressure is re-routing investment to markets that were previously considered secondary: Spain, Poland, the UK's regions outside London, and the Gulf Cooperation Council — particularly the UAE and the Kingdom of Saudi Arabia, where sovereign AI programmes are providing both demand anchor and capital underwriting for large-scale infrastructure build.
The Practical Responses: Behind-the-Meter, PPAs, and SMRs
Faced with grid constraints that cannot be resolved within AI infrastructure development timescales, developers and their advisers have pursued three principal strategies, each with distinct risk profiles.
Behind-the-Meter Generation
The fastest route to power certainty is to generate it on-site and avoid grid interconnection entirely — or to use grid connection as a partial and supplementary source. Combined heat and power (CHP) gas turbines, in particular, can be procured and commissioned in 18–24 months, compared to five-plus years for grid interconnection. The tradeoff is carbon intensity: natural gas generation does not meet the net-zero commitments of the large technology companies that are the primary occupants of AI campuses, creating a structural tension that has not yet been fully resolved.
Long-Duration Power Purchase Agreements
Twenty-year power purchase agreements with dedicated renewable generation — primarily utility-scale solar and wind — provide long-term cost certainty and allow hyperscalers to credibly claim renewable matching. The constraint is that PPAs for the volumes required by large AI campuses (200MW+) are not readily available in all markets and require substantial counterparty creditworthiness. They also do not solve the physical challenge of 24/7 power reliability, since solar and wind generation is intermittent.
Small Modular Reactors
The most discussed — and most distant — solution is co-location with small modular reactor (SMR) nuclear generation. SMRs offer 24/7 carbon-free power at high capacity factors, directly addressing the limitations of renewable intermittency. Microsoft, Google, and Amazon have all announced SMR procurement programmes. The challenge is that no SMR has yet achieved commercial operation in the Western world, and the most optimistic first-of-kind deployment timelines place commercial power delivery no earlier than the early 2030s. SMRs are a credible long-term solution; they are not a solution to the current power constraint.
What This Means for Infrastructure Advisers
The implications for anyone advising on AI infrastructure — whether as a developer, investor, lender, or government programme sponsor — are clear. Power availability is not one factor among many in site selection and project evaluation. It is the threshold criterion. A site without a credible, costed, and timetabled route to reliable power at the required scale is not a viable AI infrastructure site, regardless of its other characteristics.
This reorientation requires a different kind of analysis at the front end of any project. Grid capacity studies, utility relationship mapping, interconnection queue position analysis, and backup generation feasibility all need to precede, or at minimum accompany, the traditional questions of land cost, planning consent, and fibre connectivity.
It also requires honest assessment of timelines. A project that assumes grid power in 24 months in a constrained market is not a conservative assumption — it is an error of judgment that will cascade through every subsequent milestone in the project programme.
AI Advisory Services LLC provides independent advisory on AI infrastructure strategy, power and energy planning, and site selection across the USA, Europe, and the Middle East. If you are facing a significant infrastructure decision, we welcome a confidential conversation.