The claim that “half of AI data centers aren’t getting built” is rhetorically potent, and directionally plausible. It also runs into a basic problem: there is no single national dataset tracking “AI data centers proposed vs. built,” and there is no authoritative ledger of which projects are real, which are duplicative, and which are simply hedges.

What does exist—and what increasingly drives headlines and planning fights—are utility and grid-operator queues. Two queues matter most: large-load interconnection/service requests (demand-side) and generator interconnection requests (supply-side). Both can become inflated. Both can be misread.

A helpful baseline comes from the generator side, where the data is stronger and the pattern is unmistakable. Berkeley Lab’s Queued Up: 2025 Edition examined projects that submitted interconnection requests from 2000–2019. By the end of 2024, only 13% had reached commercial operation, 77% had been withdrawn, and 10% remained active. Those are not marginal miss rates; they are a structural feature of how queues work. (Berkeley Lab, Queued Up: 2025 Edition.)

The point isn’t that data centers will follow the same percentages. Data centers aren’t generators, and loads can move faster than power plants. The point is more basic: a queue is a filing system, not a construction schedule. If you treat it like a build forecast, you will overbuild plans—and possibly costs.

Public debate often collapses two distinct processes into one ominous phrase—“the interconnection queue.” That shortcut obscures what’s happening and who bears the risk.

Generator interconnection queues exist to study how a new power plant—or storage project—will affect the grid, and what upgrades are required. The bottleneck is notorious. Projects enter early to hold a place, learn costs, and preserve options. Many withdraw when upgrades are expensive or financing fails.

The statistics from Berkeley Lab put a spine under what grid engineers already know. When 77% of queued capacity ends up withdrawn, the queue becomes less a pipeline and more an arena. Every withdrawal can force restudies and delays for projects behind it, cascading into years of friction.

Regulators have tried to bring order. FERC Order No. 2023, issued July 28, 2023, mandated major reforms: cluster studies, a shift toward “first-ready, first-served,” and tighter readiness and financial commitments. Order 2023-A, issued March 21, 2024, adjusted details while keeping the core framework. (Federal Energy Regulatory Commission explainers.)

These reforms reflect a clear diagnosis: a system that treats speculative entries as harmless paperwork eventually becomes an expensive traffic jam.

Large-load interconnection/service requests are the demand-side cousin. They’re used to connect new factories, refineries, and now data centers—especially the biggest ones—at either the transmission or distribution level.

The problem is familiar, but the evidence is more dispersed: speculative and duplicative load requests can flood a utility’s planning process. Some developers “shop” requests across multiple utilities, regions, or states, seeking the best mix of price, speed, and incentives. Bloomberg, syndicated through Yahoo Finance, reported that utilities and analysts have struggled to size future capacity because a single data center project may pitch power requests to multiple utilities, distorting projections.

When that happens, utilities face a planning dilemma. Underbuild and risk reliability problems if a wave of projects turns real at once. Overbuild and risk stranded upgrades—costs that someone will try to recover.

If you want a real-world illustration of how queue numbers can overwhelm intuition, look at North Texas.

Oncor, the major transmission and distribution utility in the Dallas–Fort Worth region, has disclosed queue figures that are difficult to reconcile with any near-term build reality. According to The Dallas Morning News, as of June 30, 2025, Oncor’s large commercial and industrial interconnection queue had 552 requests, including roughly 186,000 MW (186 GW) from data centers. Later, in a company release, Oncor reported that at the end of 2025 its active LC&I interconnection queue included 650 requests, including ~255 GW from data centers.

These numbers are staggering on their face, but their real significance is procedural. A queue that large is not merely a signal of demand; it becomes an input to:

- where transmission upgrades are proposed,
- how distribution substations are reinforced,
- how utilities argue for capital plans,
- how state regulators evaluate “prudency” in grid spending.

The queue can influence the direction of billions in investment even if a large share of requests never materializes as actual load.

Multiple perspectives matter here. From a utility’s standpoint, the queue is a risk register. If even a fraction turns into real projects, delays in infrastructure can create reliability crises and public backlash. From consumer advocates’ standpoint, the queue is a bargaining chip—one that can be used to justify “just-in-case” spending that shows up in rates while the promised jobs and tax base remain hypothetical.

Critics sometimes describe an “interconnection-queue trick,” implying a single maneuver. The reality is less cinematic and more consequential: queue volumes can become persuasive evidence in planning and ratemaking even when the underlying requests are option-like, duplicative, or premature.

The mechanism is straightforward. A developer files one or more large-load requests, sometimes in multiple territories, to preserve flexibility. Utilities, in turn, are required to plan for reliability and serve load under regulatory obligations. In a world of long lead times for transformers, substations, and transmission lines, planners are paid to worry.

Those worries move quickly into formal documents—integrated resource plans, transmission plans, and rate cases at state public utility commissions. Even a cautious utility has incentives to say: We are seeing unprecedented demand; we need to invest now to keep the lights on.

An April 6, 2026 report from the Information Technology and Innovation Foundation (ITIF) made the concern explicit. ITIF argued that interconnection-queue data can overstate real needs because speculative and duplicative filings inflate projections. In some regions, ITIF warned, projected AI load alone can trigger near-term cost increases via forecast- or reservation-driven mechanisms—before a single facility is built.

That critique doesn’t require conspiracy. It requires only a system where early signals are treated as near-certainties, and where the cost of being wrong is socialized.

Queue inflation persists because it has rational roots:

- Developers want options; grid access is a scarce asset.
- Utilities must plan for reliability; underbuilding is punished.
- Regulators demand evidence; queues are among the few visible data sources.
- Lead times are long; by the time certainty arrives, it may be too late.

The question, then, is not whether queues will be imperfect. It’s whether the rules can distinguish between seriousness and speculation without freezing legitimate growth.

The most contentious part of the AI-data-center power story is not the technology. It’s cost allocation.

Grid upgrades are not free, and they’re rarely small. When planners see a wall of load requests, they begin to design reinforcements. If the load arrives, those upgrades may look prescient. If the load vanishes—or arrives later, smaller, or elsewhere—utilities will still seek to recover investments.

That is where rate cases become the battleground. Utilities make the argument from obligation and reliability: they cannot gamble with the grid. Consumer advocates and industrial customers often counter with a fairness claim: existing customers should not bankroll speculative projects.

The dynamics also vary by region and regulatory structure. In some contexts, large customers can be required to fund interconnection-related upgrades; in others, costs can be spread broadly, especially when upgrades are deemed system-wide benefits. The dispute is less about a single rule than about interpretation—whether a project is “incremental” or “public.”

Practical implications for readers are immediate:

- Residential customers may see higher delivery charges if grid investments accelerate.
- Small businesses can face the same squeeze, often without the ability to negotiate tariffs.
- Local communities may be promised jobs and tax revenue, but those gains depend on projects that—by the nature of queues—may not be committed.

The uncomfortable truth is that even a rational attempt to prepare for growth can become a mechanism for shifting risk from private developers to public ratepayers.

The generator interconnection world offers a partial roadmap. The system became clogged; regulators responded.

FERC Order No. 2023 and Order 2023-A are designed to reduce speculative entries and speed serious projects. The reforms include cluster studies and stronger readiness requirements—tools meant to discourage “free options” in the queue.

Berkeley Lab’s findings—13% built, 77% withdrawn for older cohorts—help explain why regulators acted. A queue dominated by projects that never materialize is not merely inefficient; it’s destabilizing.

Large-load queues are now producing the most politically explosive numbers, yet they often lack the standardized, public reporting that exists for generators. When the public hears “255 GW of data centers,” the natural assumption is that 255 GW of construction is imminent.

A more responsible framework would push toward:

- clearer disclosure of duplicative requests (without exposing competitive secrets),
- readiness milestones for large-load customers analogous to generator reforms,
- cost responsibility rules that align incentives (serious projects proceed; speculative ones pay for optionality),
- planning that distinguishes between “expression of interest” and “service-ready commitment.”

None of these changes require hostility toward data centers. They require precision about what a queue entry means—and what it doesn’t.

Readers don’t need a power engineering degree to approach AI-load claims with skepticism and clarity. A few simple questions can separate signal from noise.

When you see a number tied to data centers and the grid, ask:

Even if a large share of requests never become real data centers, queues can still reshape reality:

- Utilities may accelerate transmission and distribution upgrades “just in case.”
- Regulators may approve spending based on reliability arguments.
- Markets may respond to the expectation of higher load through capacity and congestion dynamics.

The paradox returns: speculative paperwork can produce real-world costs.

The healthiest public posture is neither panic nor dismissal. It’s insistence on better definitions: a queue entry is a request, not a guarantee; a forecast should be audited against attrition; costs should follow commitments.

Electricity planning has always involved uncertainty. What’s new is the scale and speed of load requests arriving under the AI banner—and the temptation to treat them as destiny.

The better analogy is financial, not technological. Many queue requests function like options: relatively low-cost filings that preserve the right—but not the obligation—to build later. Options are not inherently bad. They become dangerous when the rest of the system mistakes them for firm contracts.

If policymakers want to avoid either overbuilding or underbuilding, they’ll need to do the unglamorous work: standardize reporting, tighten readiness requirements, and ensure that those who reserve scarce grid capacity bear a fair share of the risk.

Readers should watch for a simple tell. When a debate leans heavily on queue totals without discussing attrition, duplicative filings, or readiness, it is not a debate about power. It is a debate about persuasion.

No single national dataset tracks proposed-versus-built AI data centers, so that exact claim is hard to verify. The more defensible point is that requests and early-stage proposals can greatly exceed what gets built, especially when developers file duplicative requests across multiple utilities. Queue volumes should be treated as interest, not as construction commitments.

An interconnection queue is a formal process for studying and approving connections to the grid. There are generator queues (for new power plants and storage) and large-load queues (for big new customers like data centers). Both can be crowded with speculative entries. A queue position signals a request to be studied—not a guaranteed project.

Because queue entries can function like low-cost options. Developers may file early to learn upgrade costs, preserve flexibility, and compete for scarce grid capacity. On the generator side, Berkeley Lab found that for projects entering queues from 2000–2019, 77% were withdrawn by end-2024 and only 13% reached commercial operation—showing how often early filings don’t translate into completed projects.

Media reporting has described “phantom” load requests where a single project may pitch power demand to multiple utilities in multiple states, complicating planning. A concrete illustration comes from Oncor in North Texas: disclosures cited 186 GW of data-center requests in its queue as of June 30, 2025, and about 255 GW by end of 2025—figures that underscore how queue totals can dwarf near-term build realities.

Utilities plan and invest based partly on expected future load. If queue totals are treated as firm demand, utilities may propose “just in case” upgrades and seek cost recovery in rate cases. Even if many projects never materialize, the planning response can still produce near-term spending pressure. That’s why cost allocation—who pays for upgrades tied to speculative demand—has become central.

On the generator side, FERC Order No. 2023 (July 28, 2023) and Order 2023-A (March 21, 2024) required major reforms, including cluster studies and stronger readiness and financial requirements. Those reforms aim to reduce speculative queue entries and speed serious projects. Comparable, standardized reforms for large-load queues are less developed and remain a policy gap.

No. The mistake is treating every request as inevitable, not taking AI load seriously. Large-load growth can be real and fast, and underbuilding can threaten reliability. The sensible approach is better transparency and readiness standards—so planners can distinguish between exploratory filings and committed projects, and so costs track actual commitments rather than optimistic paperwork.