Datacenter Power Constraints: The New Cloud Capacity Bottleneck

For most of the cloud era, the constraint on datacenter growth was capital and chips. An operator that could finance the building and procure the servers could bring capacity online roughly as fast as it could pour concrete. That assumption no longer holds. Across the major North American and European markets, the binding constraint on new cloud capacity has shifted decisively to a single resource: electrical power.

The shift is not subtle. In the largest datacenter markets — Northern Virginia, Dublin, Frankfurt, Santa Clara — utilities and grid operators are now the gatekeepers of cloud expansion. An operator with land, capital, and a signed server order can still wait years for a grid connection capable of feeding the facility. The question for infrastructure planners in 2025 is no longer “can we build it” but “can we power it, and when.”

How Power Became the Bottleneck

Two trends converged. First, the per-rack power density of modern infrastructure has climbed sharply. A conventional enterprise rack drew 5 to 10 kilowatts. A rack densely packed with current-generation accelerators for AI training and inference can draw far more — densities that were exotic a few years ago are becoming routine in purpose-built AI halls. Higher density per rack means higher power draw per square meter of datacenter floor, which translates directly into larger demands on the local grid.

Second, the absolute scale of new datacenter projects has grown. Where a large facility once meant tens of megawatts, hyperscale campuses are now planned in the hundreds of megawatts, and multi-site buildouts are discussed in gigawatt terms. A single gigawatt is roughly the output of a large nuclear reactor — the scale at which datacenter demand stops being a rounding error in a regional grid’s planning and starts driving it.

The result is the interconnection queue: the backlog of generation and large-load projects waiting for utilities and regional grid operators to study, approve, and physically connect them. In several markets these queues now stretch for years. Some utilities in the most constrained regions have paused or restricted new large-load connections entirely while they assess transmission capacity.

The Interconnection Queue Problem

Connecting a large load to the grid is not a matter of running a cable. The grid operator must study whether existing transmission can carry the additional load without compromising reliability, whether substations need upgrades, and how the new demand interacts with every other project in the queue. These studies take time, and the queue is processed in a way that means a single large, complex project can hold up others behind it.

For cloud operators, this has several consequences:

Site selection is now power-led. The availability of grid capacity — or the credible ability to procure it — drives where datacenters get built, often overriding traditional factors like proximity to network peering or cheap land.
Timelines are dictated externally. An operator can no longer fully control its own delivery schedule. The grid connection date is set by the utility, and it frequently lands well after the building and equipment are ready.
Existing markets are saturating. The traditional clusters built up over two decades of network and ecosystem advantages, but several have reached the limits of their local transmission. New capacity is being pushed toward markets with spare generation and transmission headroom, reshaping the geographic map of cloud infrastructure.

On-Site and Dedicated Generation

Faced with grid constraints, operators have begun pursuing power directly rather than waiting for the grid to catch up. The most visible move has been toward long-term power purchase agreements and dedicated generation, including nuclear.

In late 2024, the restart of a previously shuttered nuclear unit at Three Mile Island was announced specifically to supply a hyperscaler’s datacenter demand under a long-term agreement — a striking signal of how far operators will go to secure firm, carbon-free power. Other operators have signed agreements for output from existing nuclear plants, and several have publicly committed to advanced nuclear and small modular reactor projects as a future supply source, though those remain years from delivery.

The appeal of nuclear for this use case is specific: datacenters want power that is both firm (available around the clock, unlike intermittent solar and wind without storage) and low-carbon (to meet the sustainability commitments most operators have made). Few sources satisfy both at the scale required. This is why operators are willing to enter agreements of a kind that would have been unthinkable a few years ago.

On-site generation — gas turbines, fuel cells, and large-scale battery storage — is also being deployed, sometimes to bridge the gap until grid capacity arrives, sometimes as a permanent hedge against grid unreliability. The trade-off is cost and, for fossil generation, carbon: an operator that meets its load with on-site gas is undermining the emissions goals that drove it toward nuclear in the first place.

The Stargate Signal

The scale of forward demand was underscored in January 2025 with the announcement of a large multi-party initiative to build AI datacenter infrastructure in the United States, with stated ambitions measured in tens of gigawatts of capacity and hundreds of billions of dollars of investment over several years. Whatever the eventual delivered reality of such announcements, the framing is itself instructive: the headline number was power and capital, not chips. The industry has internalized that the chips can be procured if the power and the buildings exist to house them.

This reframing matters for how infrastructure teams should think about the next several years. The growth of AI workloads is, at the margin, a question about the electrical grid and the pace at which generation and transmission can be added — a question that sits largely outside the cloud industry’s direct control and operates on the multi-year timelines of energy infrastructure, not the multi-month timelines of server procurement.

Implications for Infrastructure Planners

For organizations that consume cloud capacity rather than build it, the power constraint has indirect but real effects:

Regional availability will tighten. Capacity in the most constrained regions may become scarcer and more expensive, and new instance types — particularly accelerator-heavy ones — may launch first in markets with power headroom rather than in the traditional hubs.
Sustainability reporting gains substance. As operators sign nuclear and renewable agreements to feed their growth, the carbon characteristics of cloud regions will diverge more sharply. Organizations with emissions-reporting obligations will find the choice of region increasingly material.
Efficiency returns to the agenda. When power is the scarce input, the efficiency of a workload — how much useful work it extracts per watt — becomes an economic and capacity question, not only an environmental one. This favors architectural attention to utilization, scheduling, and right-sizing that the abundant-capacity era let slide.

This connects to a longer theme in distributed-systems research: the management of constrained, shared resources across a federation of sites. The scheduling and placement problems that grid computing tackled — matching demand to available capacity under hard physical limits — are reappearing at the level of the electrical grid. The constraint has simply moved from CPU cycles to kilowatt-hours.

The Outlook

The power bottleneck is not a transient supply-chain hiccup that will resolve in a quarter or two. Building generation and transmission is slow, capital-intensive, and subject to permitting and regulatory processes measured in years. The demand from AI infrastructure, meanwhile, is growing on a much faster curve.

The likely outcome is a sustained period in which power — its availability, its cost, and its carbon intensity — is the primary variable shaping where and how fast cloud capacity grows. Operators that secured firm power early will have a durable advantage. For everyone planning infrastructure around cloud capacity, the practical lesson is to treat regional power availability as a first-class planning input rather than an assumption.