Texas Freezes Data Center Power Over 'Ghost Demand'

Texas, the state that has courted more AI data center investment than almost anywhere in the country, has slammed on the brakes. After connection requests to its grid ballooned to an estimated 474 gigawatts — more than five times the grid’s record peak demand — the state paused new large-load interconnections and ordered a project-by-project audit to figure out how much of that staggering number is real. The episode has become a national test case for a problem regulators are calling “ghost demand.”

The freeze crystallized a reckoning that has been building all year: across the middle of the United States, data centers have requested roughly as much electricity as it takes to power every home in the country. Much of that, officials increasingly suspect, is an illusion — the same project counted many times over, or proposals from developers with neither the funding nor the customers to build.

What Texas actually did

On August 3, 2026, Governor Greg Abbott directed the Public Utility Commission of Texas (PUCT) and the grid operator, the Electric Reliability Council of Texas (ERCOT), to audit every data center sitting in the interconnection queue. Any project that fails to comply with the audit’s disclosure requirements is to be denied connection to the Texas grid.

The scale is the story. ERCOT is weighing roughly 474 GW of load interconnection requests, with data centers accounting for about 90% of that queue. For context, that figure dwarfs ERCOT’s all-time peak demand — the grid has never come close to needing that much power at once, and no credible forecast says it will this decade. The audit examines each project’s power demand and generation strategy, water use, public incentives, community impact, and ownership, so regulators can separate committed builds from placeholders.

Analysts at BloombergNEF estimated the review could delay roughly 49.8 GW of genuine data center load and put up to $15 billion of projects at risk of slipping — a reminder that a cleanup aimed at fake demand still catches real projects in its net.

The legal machinery: SB 6

The freeze did not come from nowhere. It rests on Senate Bill 6 (SB 6), a law Abbott signed in June 2025 that overhauled how large-load customers connect to the ERCOT grid. SB 6 set a 75-megawatt threshold for large-load interconnection standards and — critically for the ghost-demand problem — required applicants to disclose substantially similar interconnection requests they have filed elsewhere in Texas, along with any on-site backup generation.

That disclosure requirement is the key. A major driver of phantom queue volume is developers filing the same project at multiple sites to preserve optionality, then building at most one. By forcing applicants to reveal duplicate filings, SB 6 gives regulators a tool to deflate the queue to something closer to reality. The August audit is SB 6’s enforcement moment: the point where disclosure obligations get checked against actual plans.

Independent evidence suggests how much air is in the number. Of the more than 100 data centers proposed in the state, only about 20 have applied for the permits needed to move forward, and most proposals have not secured a power source or an anchor customer — the two things that separate a financeable project from a speculative land grab.

Why “ghost demand” matters

The term captures a specific failure mode of the AI infrastructure boom. When every hyperscaler, neocloud, and speculative developer files enormous interconnection requests, grid operators and utilities are forced to plan — and spend — against numbers that may never materialize. That means building transmission, procuring generation, and in some cases justifying higher rates on consumers, all to serve load that could evaporate.

The gap between requested and real capacity distorts the entire planning stack. It inflates demand forecasts that utilities use to justify new gas plants and transmission lines. It lets speculative developers tie up grid capacity that credible builders need. And it shifts risk onto ratepayers, who can end up paying for infrastructure sized to a queue that was always partly fictional. This is the same tension we traced in the economics of AI data centers: the physical grid moves far more slowly than the capital chasing AI compute, and the mismatch has to resolve somewhere.

Texas is not acting alone. Other states are moving to fix the same problem. Pennsylvania Governor Josh Shapiro signed a data center executive order on August 18, 2026, imposing stricter permitting requirements on projects of 25 MW or more and demanding greater disclosure of project plans and end users. The common thread is a shift from courting data centers at any cost to demanding that developers prove demand is real before the grid commits to serving it.

The collision with the AI buildout

The freeze sits directly athwart the largest infrastructure buildout in a generation. Texas has been a magnet for exactly the kind of multibillion-dollar campuses driving the hyperscaler capex boom — including the $35 billion Anthropic–Lambda campus we covered in that Texas data center deal. Those are precisely the projects — well-funded, with named anchor tenants — that should sail through an audit designed to weed out speculation.

The risk is timing. Even legitimate projects now face a review that adds months of uncertainty, and the BNEF estimate of ~50 GW of potentially delayed load shows the collateral cost of a blunt instrument. Power, not chips, has quietly become the binding constraint on AI expansion — which is why the industry is pouring money into everything from on-site generation to more efficient data center power delivery. A connection queue that can’t be trusted makes that constraint worse, because no one can plan against a number that is five times too large.

What it means

Texas has done something the rest of the country will be watching closely: it has said out loud that the demand figures underpinning the AI infrastructure story are not fully trustworthy, and it has built the legal machinery to test them.

The queue is not the demand. The single most important takeaway is that 474 GW is a request number, not a forecast. Anyone — investors, utilities, policymakers — who has been extrapolating grid stress or generation needs from headline interconnection totals has been working from an inflated base. The audit’s real product is a defensible estimate of how much of the AI power demand is genuinely financed and committed.

Winners and losers. Well-capitalized hyperscalers with signed tenants and secured power win: the audit clears the speculative clutter blocking their path and validates their projects. Ratepayers win if the review stops utilities from over-building against phantom load. The losers are speculative developers who filed duplicate or unfunded requests to hold grid position — and, in the near term, legitimate projects caught in the delay.

What to watch next. First, the audit’s findings — the gap between the 474 GW requested and the capacity that survives scrutiny will be the most honest number yet on real AI power demand, and it will reset forecasts nationally. Second, contagion: expect more states to copy the SB 6 disclosure model, turning “prove your demand” into the default posture. Third, whether the delay meaningfully slows the AI buildout or merely reprices it, as developers race to lock in power and permits ahead of tighter rules. Texas spent years saying yes to data centers. The more consequential shift is that it — and a growing list of states — now wants to see the receipts first.