The tensions caused by artificial intelligence data centers do not stem solely from the actual electricity consumed. Projects that are still uncertain may reserve capacity from multiple grids, lengthen connection waitlists and distort planning. The debate therefore centers just as much on the maturity of the projects and the sharing of costs that on growth, which is very real, the energy demand of digital technology.
What Just Happened
On October 13, 2026, The Wall Street Journal described the growing difficulty U.S. electric utilities face in distinguishing between data center projects that are ready for construction and those that are still in the exploratory phase [1]. In this context, the term “ghost data center” refers to a project announced as part of a connection queue, but for which the financing, land, equipment, timeline, or even the final location have not yet been sufficiently established.
This classification does not necessarily mean that a project is fraudulent. A single developer may explore multiple regions to compare timelines, electricity prices, or regulatory constraints. The same potential load is then counted multiple times across separate portfolios. Conversely, an overloaded grid can slow down credible projects and lead grid operators to consider upgrades that may not all ultimately be necessary.
A connection request measures an intention and a maximum need considered. It does not constitute neither observed consumption nor a guarantee of construction, nor a certain forecast of future charge.
What’s Really New
Speculative projects in the power grid are nothing new. What has changed with AI is their scale, geographic concentration, and speed. According to the WSJ, a typical request made to certain suppliers has increased from about 50 MW to 1,000 MW or more [1]. A single project can therefore have the same impact as a large industrial facility, even though its capacity is sometimes drawn upon gradually over several years.
This dynamic is complicated by a timing mismatch. A data center can be built in a few years, whereas a high-voltage transmission line, a power substation, or new generation capacity often require longer approval processes. To secure an option, developers are advised to file their applications early. To protect consumers, however, utilities need evidence of project maturity before investing. The novelty lies less in the queue itself than in the gap between the scale of demand and the number of projects that have already demonstrated their ability to succeed.
How Does a Connection Queue Work?
When a project requires high power capacity, the manager analyzes its location, consumption profile, start date, and the work required. The manager may need to upgrade a substation, a transmission line, or flexibility resources. The project moves forward through a series of technical studies, financial guarantees, permits, and contracts. Therefore, being in the queue does not mean that capacity is immediately available.
It is also important to distinguish between demand queues—which involve both consumers and data centers—and generation queues, which include power plants and renewable energy facilities seeking to feed electricity into the grid. The two interact, but their rules and metrics are not interchangeable.
Why This Matters for the AI Economy
A bloated pipeline has effects even before a substation is installed. It influences load forecasts, substation siting, utility investments, land-use decisions, and sometimes rates. If infrastructure is built for a project that is subsequently abandoned, part of the cost may be passed on to other customers. Conversely, if operators underestimate demand, actual projects may be delayed and the reliability of the power supply compromised.
The environmental challenge, therefore, goes beyond the number of kilowatt-hours consumed by the models. An overly high forecast can lead to the development of oversized infrastructure, the use of materials, and the reservation of land. An overly low forecast can encourage temporary solutions with higher carbon emissions, such as on-site power generators. The quality of connection data thus becomes a component of AI’s material footprint.
This is also an industrial issue. Access to reliable power determines where computing capacity is located, and thus a region’s attractiveness for cloud computing and AI. Screening applications should not mean blocking investment, but rather prioritizing projects based on criteria that are transparent, comparable, and compatible with local capacity.
What the Data Actually Shows
These figures should not be added together, as they describe different areas, dates, and scopes. Nevertheless, they point to two common observations. The growth in electricity demand from data centers is measurable. At the same time, the gross volume of applications does not constitute a reliable projection without information on duplicates, studies conducted, guarantees submitted, and the ramp-up schedule.
Liability Issues
Emerging responses rely on deposits, longer-term contracts, maturity milestones, and documentary reviews. Exelon has reduced its high-probability portfolio after requesting financial commitments [1]. In Texas, ERCOT has implemented a batch processing system to verify the eligibility of large loads and better incorporate sufficiently advanced projects into its studies [5]. In the United Kingdom, Ofgem proposed in 2026 to better screen viable projects, notably through milestones and commitments designed to limit speculative applications [6].
However, these mechanisms must remain proportionate. A guarantee that is too onerous may favor the largest groups at the expense of smaller players, even credible ones. Nor does a purely financial criterion provide insight into social utility, carbon footprint, water use, or the local impact on prices. Governance should therefore combine technical maturity, financial responsibility, environmental transparency, and the protection of other consumers.
- Report separately the requested, planned, contracted, energized, and actually drawn power;
- Report requests submitted for the same project in multiple jurisdictions, without disclosing trade secrets;
- Require the applicant to bear a reasonable share of the risk associated with reinforcement and removal;
- Plan for a phased ramp-up, with a review of unused capacity;
- Assess local impacts on rates, water, emissions, employment, and grid resilience.
What to Watch for Now
The first useful metric will be the conversion rate of the queues: how many megawatts requested actually reach the planning, contracting, energization, and consumption stages? It will also be necessary to track the gap between reserved capacity and observed peak demand, cancellations after a security deposit has been posted, the duration of connections, and the portion of costs ultimately borne by data centers.
International comparisons will be instructive. Texas is experimenting with a more formalized selection process for major projects, while the United Kingdom is reforming its application process. Their effectiveness will be judged by their ability to expedite sound projects without unduly shifting risks to households, businesses already connected to the grid, or local communities.
Ghost data centers therefore do not prove that AI’s electricity demand is fictitious. Rather, they reveal a flaw in measurement: an intention to connect is often treated as future consumption. To plan for the growth of AI rigorously, we will need to replace the spectacular figure from the backlog with a set of verifiable indicators, ranging from the exploratory proposal to the megawatts actually used.
To learn more, visit the aivancity blog
Find four additional analyses on infrastructure, energy and impacts environmental aspects of artificial intelligence .
Sources
[1] Hiller, Jennifer. (October 13, 2026). “Phantom” Data Centers Are Demanding More From the Power Grid Than They Need. The Wall Street Journal. Read the article
[2] International Energy Agency. Energy and AI, Energy demand from AI. Data on global data center energy consumption. View the report
[3] International Energy Agency. Energy and AI Observatory. Issues Related to Site Selection and Capacity Planning. Visit the observatory at
[4] ERCOT. (April 2026). Interconnection and Grid Analysis Update. Monitoring of major loads and the share of data centers. View the document
[5] ERCOT. (2026). Large Load Integration. Eligibility Documents and Verification Process for Large Loads. View the documentation
[6] Ofgem. (July 29, 2026). Ofgem acts to free up grid capacity by tackling speculative data center projects. View the press release
[7] Ofgem. (2026). Proposed data center connection reforms. Consultation on data center connection reforms. View the consultation