A power request can look remarkably solid on paper while remaining little more than an option held against future infrastructure. That problem became impossible to ignore as large-load requests in ERCOT moved from roughly 48 GW in 2023 to more than 474 GW by mid-2026, a surge driven heavily by proposed data center demand rather than equivalent growth in energized consumption. The number does not represent 474 GW operating on the system, nor does it mean that every requested megawatt will require transmission and generation on the same schedule. Instead, it exposes a planning problem created when the volume of requests grows faster than the system can establish which projects have credible schedules, financing, sites, equipment, and customers. ERCOT’s own reporting separates projects that have submitted no studies from projects under review, approved to energize, and already operating, showing why a queue total cannot function as a direct forecast.

The planning challenge becomes sharper when a single queue combines projects at radically different stages of development and treats their requested capacity as part of the same visible demand universe. A proposed AI site with an identified location and construction schedule carries different planning value from a concept that has not completed studies or established a commercial path, yet both can contribute to the headline queue. ERCOT’s March 2026 operating overview showed approximately 429.9 GW of projected large-load growth through 2033, while more than 310 GW remained in the category for projects with no studies submitted, illustrating how much of the apparent pipeline still required validation. By July, reported requests had risen beyond 474 GW, reinforcing how quickly the queue itself could change before physical demand caught up.

When 48 Became 474 In 24 Months

The jump from about 48 GW to more than 474 GW is striking because the queue expanded far faster than physical electricity consumption could reasonably follow. Large-load requests grew as developers pursued sites capable of supporting increasingly dense computing deployments, while the interconnection process became the place where planners recorded future power requirements before those requirements became construction commitments. That behavior changes the statistical character of the queue because a request represents an intention to seek service, not proof that a facility will reach energization. AI infrastructure has introduced a rapidly expanding pipeline of large-load requests, with ERCOT reporting that the volume of proposed large loads had grown beyond the capacity of its earlier review process. Consequently, the headline number became more useful as a measure of market pressure than as a direct estimate of future electricity consumption.

The speed also exposed a timing mismatch between private project development and public infrastructure decisions. A large computing site may change its capacity target as hardware availability, customer commitments, financing conditions, and deployment economics change, while transmission upgrades require engineering, permitting, procurement, and construction over much longer periods. ERCOT acknowledged that historical processes no longer matched the unprecedented level of large-load interconnection activity and subsequently moved toward batch evaluation for qualifying projects. The approved process groups eligible large-load requests so planners can assess their combined reliability effects instead of repeatedly studying projects in isolation. That change matters because planners cannot evaluate a 1 GW request only as an isolated customer when dozens of similarly sized projects seek capacity in the same electrical area. Planning therefore has to examine geographic concentration, commissioning timing, operating behavior, and network constraints together rather than treating every request as an independent future load.

The Queue Learned To Hoard

The rapid increase in large-load requests made the qualification of proposed capacity increasingly important because the interconnection process had to distinguish projects at different stages of development. A project that enters the interconnection process can remain at an early stage of development while additional technical and commercial information is gathered before it advances toward energization. The result can create multiple competing claims against limited grid capacity without requiring every claimant to reach the same level of financial or construction readiness. This does not mean every request is speculative, because serious projects can legitimately enter an interconnection process before every commercial detail becomes final. It does mean that the planning system must identify which project attributes have actually advanced since the original request entered the queue. Site control, completed studies, financing evidence, equipment procurement, customer commitments, construction milestones, and credible energization schedules can provide stronger signals than requested megawatts alone.

Site availability creates another layer of uncertainty because a proposed location can support a request long before the physical and commercial conditions required for construction become firm. A proposed site can enter the large-load process before all of the technical, financial, and construction-related requirements needed for energization have been completed. The same uncertainty makes project qualification important because proposed large loads can remain at different stages while the grid evaluates their technical requirements and readiness for interconnection. That behavior makes a simple queue count particularly weak as a forecasting instrument because the count records claims while the system ultimately needs to serve physical facilities. ERCOT’s large-load process already requires detailed project information and study inputs, while the 2026 verification effort placed greater attention on project readiness and the credibility of proposed demand.

Forecasting Models Had No Filter For Hype

Conventional load forecasting works best when historical relationships provide a reasonable guide to future consumption, but AI infrastructure weakens those relationships because project announcements can precede actual demand by several years. Residential growth usually develops through measurable population and housing trends, while established industrial demand often follows observable production capacity, employment, and investment patterns. Large computing sites can behave differently because a single project may request hundreds of megawatts before its ultimate workload, customer mix, and hardware deployment schedule become fixed. The forecasting challenge is therefore not simply choosing a higher growth rate; it is determining how much confidence each demand signal deserves at each development stage. A model that gives equal weight to an early request and an approved project effectively converts uncertainty into apparent certainty.

Useful signals can come from the sequence of actions that precedes energization rather than from the requested load itself. A stronger model can assign greater confidence when a project has advanced studies, secured site control, demonstrated financing, ordered long-lead electrical equipment, established construction milestones, and provided a credible commissioning schedule. Geographic concentration should also influence the forecast because multiple large-load projects in the same area can create different transmission and reliability requirements from an equivalent amount of capacity distributed across separate locations. Operating characteristics matter as well, since an AI computing load with a predictable commissioning curve creates different requirements from a project that could rapidly expand after initial energization. The model should also preserve uncertainty instead of forcing every request into a binary real-or-fake classification, because credible projects can still experience delays, resizing, or phased construction.

ERCOT’s 474GW Lesson In One Line

The most important number is not 474 GW; it is the gap between what a queue records and what a grid must eventually deliver. A queue can reveal where developers are seeking access to power, but it cannot independently establish how much of that requested capacity will ultimately become energized demand, when that demand will arrive, or how quickly individual projects will advance through the interconnection process. That difference became visible when ERCOT’s reported queue reached a level several times larger than the system’s historical peak demand, prompting a review of projects before they could advance through the new process. The response did not reject the underlying AI demand story, and it did not assume that every request represented an imminent load. Instead, it moved the process toward broader assessment, project verification, and stronger information requirements before scarce grid capacity gets committed.

The durable lesson is that future demand planning must become evidence-driven without becoming blind to early signals. AI can create legitimate electricity demand at a pace that historical forecasting models have not experienced, yet legitimate demand still requires a physical site, electrical connection, equipment, capital, construction activity, and an operating plan before it becomes a grid obligation. A robust planning process should therefore maintain separate views for requested capacity, technically viable capacity, commercially credible capacity, construction-ready capacity, and energized demand. The 474 GW episode ultimately demonstrates that the challenge is not forecasting growth too aggressively or too conservatively, but knowing which pieces of the forecast have earned enough evidence to influence irreversible infrastructure decisions. In one line, queues measure interest, while infrastructure must be built around intent that has survived technical, commercial, and physical scrutiny.