Utilities and grid operators across the United States are sitting on requests for 1,066 gigawatts of new electricity capacity to feed the AI data center boom — but according to new projections from energy research firm Wood Mackenzie reported by Bloomberg on August 12, only about 28% of that requested power, roughly 298 gigawatts, is likely to actually get built and connected to the grid. The remaining 768 gigawatts, Wood Mackenzie says, is effectively “phantom load”: speculative, duplicative, or non-binding requests that have no realistic path to construction.
How phantom demand piles up
The inflated numbers stem largely from how data center developers approach the interconnection process. Rather than committing to a single site and utility, developers routinely submit power requests to multiple utilities simultaneously for the same project, hedging their bets on which location will secure land, water, chips, and permits fastest. Each of those parallel submissions shows up separately in a utility’s interconnection queue, meaning the same gigawatt of planned capacity can be counted two, three, or more times across different grid operators — inflating the headline demand figures utilities use for long-term planning without representing genuinely new, additive electricity need.
Why this matters beyond spreadsheets
Utilities use these interconnection queues to justify major capital decisions: building new transmission lines, upgrading substations, and in some cases delaying the retirement of fossil-fuel plants to preserve capacity margin. If a large share of the demand behind those decisions is phantom, ratepayers risk paying for grid infrastructure sized for a data center boom that only partially materializes. Wood Mackenzie’s analysis lands alongside a broader wave of reporting this year on rising electricity bills in regions with heavy data center buildout, where state regulators and consumer advocates have begun questioning how much of the new infrastructure cost being passed to households is actually justified by real, committed data center load.
The background: how AI supercharged power demand forecasts
Electricity demand from U.S. data centers was widely forecast in 2025 to nearly triple by 2035, driven by the compute needs of large language model training and inference. That forecast triggered a rush among hyperscalers — including projects like OpenAI’s newly announced Project Camellia data center campus in Effingham County, Georgia, and AWS’s expanded $48 billion investment plan in India that includes $21 billion for cloud and AI infrastructure through 2030 — to lock down power commitments years ahead of actual construction. The scale of those ambitions is part of what has driven interconnection queues to record highs, even as the industry acknowledges that chip supply, water availability, and permitting delays mean many announced projects will be delayed, downsized, or canceled outright.
Industry response: risk management, not deception
Data center developers and their trade representatives argue that filing parallel, non-binding requests across multiple utilities is standard, rational risk management in a capital-intensive, multi-year siting process — not an attempt to mislead grid planners. They note that utilities themselves apply “at-risk” discount factors to raw interconnection queue numbers precisely because they understand a large share of requests will not convert to built capacity, and that treating every megawatt in a queue as guaranteed demand was never the industry’s expectation.
The skeptical view from grid planners
Critics, including some state utility regulators, argue that even accounting for standard attrition, the scale of overstatement Wood Mackenzie describes — roughly 72% of requested capacity failing to materialize — is large enough to distort multi-billion-dollar infrastructure decisions and delay the retirement of aging power plants that utilities keep online “just in case” the phantom demand turns real. Some have called for stricter deposit and penalty requirements on interconnection applications to force developers to signal genuine commitment rather than optionality.
What to watch next
Several states are now moving to reform interconnection rules specifically to filter out speculative data center requests, including proposals for higher upfront financial commitments and shorter windows before an unbuilt project loses its place in the queue. As AI companies continue announcing new campuses through the back half of 2026, the gap between announced gigawatts and built gigawatts will be the clearest signal of whether the AI infrastructure boom is proceeding as advertised — or running well ahead of the physical grid capacity that will ultimately have to support it.
