AI Data Center Power Demand Shrinks 17 GW in Texas as Develo

Texas's largest power queue has lost more than half its volume in a single step. American Electric Power watched interconnection requests in its Texas service area fall from roughly 30 gigawatts to about 13 GW once the utility required developers to cover 85% of the capacity they were asking for. The 17 GW that disappeared over recent months is the sharpest measurement yet of how much of the AI data center power demand utilities were planning around was never committed capacity.

The reduction sits inside a wider reckoning. ERCOT, the grid operator covering most of Texas, is carrying 474 GW of interconnection requests, more than five times the state's record peak demand. Governor Greg Abbott directed the Public Utility Commission of Texas and ERCOT in August to freeze new data center connections and audit every request already in the queue. The first batch of projects under review covers about 200 GW by itself.

Why AI Data Center Power Demand Shrank by 17 GW

Interconnection requests were cheap before the guardrails arrived, so reserving capacity worked as a low-cost option on future growth. A developer could stake a claim on grid capacity for a campus it had not financed, had no signed tenant for, or might never build. When operators attach money, engineering studies, and evidence of a real customer to a request, much of the queue stops making sense to hold.

AEP's take-or-pay condition applies that pressure directly. Developers must pay for 85% of the capacity they reserve whether or not the load ever shows up, which turns a free option into a commitment sized in the tens of millions of dollars for a mid-sized campus. What survived the change, 13.02 GW, is the portion developers were willing to put money behind.

The scale of the withdrawal reads more clearly in proportions than in raw numbers. More than half of the capacity AEP had been planning against was withdrawn once payment entered the picture, which means the utility's growth projections for its Texas territory rested on requests that developers declined to stand behind at cost.

AEP's requirement governs its own Texas service area rather than the whole ERCOT footprint, so the 30-to-13 GW figure measures one utility's queue rather than the state's. ERCOT's 474 GW total still includes requests filed with other utilities and cooperatives, each carrying its own threshold for what a developer must fund.

Other states are running the same experiment. Pennsylvania Governor Josh Shapiro signed a comparable order in August. Ohio lifted its fee for grid interconnection studies to $100,000, and reported data center demand in the state fell after the change.

Jurisdiction	Guardrail applied	Reported effect
AEP Texas	Developers pay for 85% of reserved capacity	Queue fell from about 30 GW to 13.02 GW
ERCOT statewide	Freeze on new connections plus audit of pending requests	474 GW under review; first batch about 200 GW
Ohio	Interconnection study fees raised to $100,000	Reported data center demand declined

Take-or-Pay Becomes the Binding Constraint

An 85% floor changes who carries the risk of a forecast being wrong. The utility gains a revenue guarantee for most of the capacity it reserves, which supports the transmission and generation it builds around that reservation. The remaining exposure lands elsewhere: on the 15% developers are not obliged to pay for, and on any infrastructure built for a project that never energizes.

Reservation charges and minimum bills are ordinary tools in power procurement, but they have rarely been applied at this scale to load with no customer behind it. Utilities historically managed speculative requests through queue position, deposits, and letters of intent, arrangements that cost a developer little to abandon. An 85% obligation is a different instrument: it prices the option rather than merely ranking it.

That residual is where ratepayers sit. Utilities lock in capacity prices against their forecasts of AI data center power demand, and those prices flow into the bills of every customer on the system. When forecasts built on speculative requests run ahead of the load that materializes, the cost of the unused headroom does not disappear; it is spread across the customer base.

Take-or-pay narrows the gap between what is planned and what is paid for, but it does not close it. A developer holding an 85% obligation on a project that stalls still leaves the utility with capacity, land, and interconnection work that someone must absorb.

What the National Numbers Say

The stakes scale well beyond Texas. Data center electricity use in the United States rose about 80% between 2020 and 2025, and current demand sits near 176 terawatt-hours a year. Major technology firms put more than $200 billion into AI infrastructure during 2024.

The planning problem is the range. Projections for 2030 run from roughly 200 TWh to more than 1,000 TWh, a five-fold spread. A utility that builds for the high end and gets the low end strands capital; one that builds for the low end and gets the high end fails to serve load. Take-or-pay contracts narrow that spread by discarding the requests nobody will fund.

The spending that is real is also concentrated. The $200 billion invested in 2024 came from a small set of balance sheets, while the queue mixed those commitments with requests from developers carrying far less capital and far less certainty about tenants. Generation and transmission projects take years to permit and build, so a utility choosing a capacity figure today is committing to it long before the tenants behind the requests are verified.

What the Audits Will Reveal

Texas should get a clearer picture of the real pipeline once the review of the first batch of projects finishes. With roughly 200 GW in that group alone, even a modest reduction shifts current estimates of AI data center power demand in the state. The audit standard matters as much as the audit itself: requests that cannot show site control, financing, and a named customer are the ones most likely to fall away.

The review is expected to move through the queue in batches rather than all at once, which leaves utilities planning against an incomplete number in the interim. That gap matters for procurement, because turbine orders, transformer lead times, and transmission siting decisions are locked in months or years before audit results arrive.

Utilities face a two-sided cost. Building around every request in the queue risks billions of dollars in spending on facilities that never appear, while building around too few risks failing to serve load from projects that do arrive.

Local opposition is adding a second filter. Projects are being delayed or blocked over environmental concerns and the cost impact on nearby customers, which affects how much of the audited queue can actually be built even where the demand is genuine.

For developers, the practical change is timing and cost. Capacity that once could be reserved on paper now requires payment, engineering studies, and proof of a customer before it can be counted, which raises the price of holding land and power options while a site is still being selected. Financing, land control, and tenant agreements have to be in place earlier, because a queue position no longer substitutes for them.

Why This Matters

The 30 GW to 13 GW collapse shows that a large share of the AI power crunch was optionality rather than committed load, and that utilities now have a tool for separating the two. The consequence for data center economics is that take-or-pay terms, not land or fiber or chip supply, increasingly decide which projects move forward. For ratepayers, the exposure that remains is the capacity utilities build against forecasts that the payment structure only partly filters out.

Photo by BENOIT LAMARCHE on Unsplash

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.