A new industry forecast warns that most of the electricity being requested to power America’s artificial intelligence data centers will never actually be delivered, because a large share of the demand is inflated or speculative. The projection estimates that more than two-thirds of the power sought will fail to materialize, undercutting some of the boldest claims about the scale of the AI buildout. The finding has real consequences for utilities, grid planners, and ordinary ratepayers.
The forecast behind the two-thirds figure
The projection comes from energy research firm Wood Mackenzie, which examined the enormous volume of power that data-center developers have requested from utilities. According to the report’s findings, grid operators and utilities are likely to commit to only about 28 percent of the roughly 1,066 gigawatts requested for data-center projects. That leaves the great majority of the demand unlikely to result in an actual, energized facility.
A gigawatt is an immense amount of power, enough to supply hundreds of thousands of homes, so a request totaling more than a thousand gigawatts represents an extraordinary claim on the nation’s electricity supply. The forecast suggests the true figure that will get built is a fraction of that headline number.
The problem of phantom projects
Much of the gap stems from what the industry calls phantom applications. Some developers pitch the same project to multiple utilities at once, planning to move forward only with whichever application lands the best deal and the fastest approval. As an account of the projection explained, that practice inflates the apparent demand because a single planned data center can show up as several separate power requests across different providers.
The result is a distorted picture in which the sum of all requests vastly overstates how many facilities will genuinely be built. Duplicate and long-shot proposals pile onto legitimate ones, making it difficult for planners to distinguish real demand from speculative interest.
Why inflated demand hits electricity bills
The overstatement is not merely an accounting quirk; it complicates planning and budgeting for utilities and grid operators. When providers prepare to serve demand that never appears, the capital costs of upgrading transmission lines, substations, and generation are typically recovered through rate increases. That means the expense of gearing up for phantom projects can translate into higher utility bills for households and businesses that had nothing to do with the AI industry.
The dynamic creates a genuine risk: build too little and real data centers cannot get connected, but build for demand that never arrives and the public foots the bill for stranded infrastructure. Threading that needle requires accurate forecasts, which is precisely what phantom applications make harder to produce.
Why developers pitch the same project everywhere
The practice of submitting a single project to multiple utilities is a rational response to a slow, uncertain approval process. Connecting a large data center to the grid requires securing power commitments, permits, and infrastructure upgrades, any of which can stall. By approaching several providers at once, a developer improves the odds of finding the fastest, cheapest path to power and can then abandon the other applications. From the developer’s vantage point it is prudent hedging, but from a utility’s it looks like real demand that may never arrive.
The trouble is that utilities and grid operators cannot easily tell which requests are serious and which are placeholders. Every application must be treated as potentially real for planning purposes, because failing to prepare for genuine demand risks leaving legitimate projects unable to connect. That asymmetry, developers holding the knowledge of which projects are firm while planners must account for all of them, is what allows a modest number of actual data centers to generate a towering, inflated total of requested power.
The grid strain behind the numbers
Even setting aside phantom applications, the electricity the AI boom genuinely requires is straining the power system. Data-center demand has been pushing forecasts of national electricity use to record highs, forcing utilities to weigh expensive new generation and transmission. The core difficulty is distinguishing the portion of that demand that will truly materialize from the speculative surplus, because building for the wrong figure carries costs in either direction.
If planners underbuild, real projects stall and economic activity is lost; if they overbuild for demand that evaporates, ratepayers are left covering the cost of idle infrastructure through higher bills. The forecast’s warning that most requested power will never be delivered is therefore not a dismissal of the AI buildout but a caution about how to size the response to it. Getting that calibration right has become one of the central puzzles facing utilities, regulators, and the technology companies driving the demand.
How much of the pipeline may actually get built
The Wood Mackenzie warning aligns with other analyses questioning the pace of the buildout. According to research on the data-center pipeline, a substantial portion of projects slated for 2026 may be canceled or delayed rather than coming online as scheduled, with a meaningful share unlikely to be finished before year’s end. Constraints on power, equipment, and construction are converging to slow projects that once looked imminent.
Taken together, the findings paint a more measured picture than the sweeping announcements that have accompanied the AI boom. Demand for computing power is real and growing, but the raw tally of power requests appears to dramatically overstate how much generation and grid capacity will ultimately be needed. For utilities and the customers who fund them, distinguishing the genuine buildout from the speculative one has become a central challenge of the AI era.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
