Texas has 474 gigawatts of data-center demand and a grid that peaks at 91. It just stopped believing the number.

The most important number in the American power grid this month is one that cannot be true. Texas, the fastest-growing market in the world for the giant warehouses of computers that run artificial intelligence, is sitting on roughly 474 gigawatts of requests to connect new large loads to its grid, about nine-tenths of it data centers. The grid those requests want to join has never, in its history, drawn more than about 91 gigawatts at once — a record set on a hot afternoon this July. The queue is more than five times the size of the grid. So the state did the only sensible thing you can do with a number that size. It stopped believing it.

In early August the governor directed the state's utility regulator and its grid operator, ERCOT, to audit every data-center project moving through the interconnection process before any new one could advance. In plain terms, Texas froze the queue to find out how much of it is real. The grid operator now aims to finish the count by December. This is not a story about a power shortage. It is a story about a measurement problem, and the measurement problem is expensive.

What an interconnection request actually costs

To see why the number got so large, price the thing being counted. An interconnection request is not a purchase. It is closer to a reservation — a developer telling the grid, 'I may build a large facility here, please begin studying what it would take to power it.' For years, in most of the United States, filing one cost very little: some paperwork, a modest fee, no binding commitment to ever pour concrete. When a reservation is nearly free and the thing being reserved — guaranteed grid capacity in the middle of an AI build-out — is scarce and valuable, you do not get a demand forecast. You get a land rush.

So developers did what the incentives told them to. Since 2023, requests from data centers and other large users to connect to the Texas grid climbed from about 48 gigawatts to more than 474. A single project shopped to three different sites shows up three times. A speculative site a developer is not sure it will ever build shows up at full size, because there is no cost to overstating it and a real cost — losing your place in line — to being modest. More than 1,800 projects now sit in the queue, and by the state's own account many of them lack the basic elements a real project has: financing, a signed tenant, land control, a plausible path to actually being built.

When a reservation is nearly free and the thing reserved is scarce and valuable, you do not get a demand forecast. You get a land rush.

This is what the word 'ghost demand' is pointing at. Not fraud, exactly — just the predictable output of a system that let people reserve enormous quantities of a valuable thing for almost nothing. The 474-gigawatt figure is not a lie. It is an honest count of requests, most of which were never priced to mean anything.

Who pays for the demand that never arrives

Here is why the fog matters, and it comes down to who holds the bill. A grid operator does not just watch the queue; it plans against it. Transmission lines, substations, new generation — the poles and wires and power plants that take years to build — get sized to the demand the operator expects. If the operator believes even a fraction of 474 gigawatts, it builds toward a grid several times larger than the one Texas has. That construction is not free, and the people who pay for it are not the developers who filed the requests. They are ratepayers, the households and businesses whose monthly bills fund the shared infrastructure.

The cost that matters here is the one that gets socialized. Build transmission for a data center that arrives, and the load helps pay for the line. Build it for a data center that was always a placeholder, and the line still gets built, still gets paid for, and the payer is everyone else — spread across the bills of people who never filed a request and will never draw a watt of that capacity. Overbuild against phantom demand and you raise the cost per kilowatt-hour for the whole system to serve load that does not exist. That is the real exposure Texas just paused to measure. Not 'will the lights stay on,' but 'are we about to make everyone pay to build for ghosts.'

The governor's order reaches for exactly these levers. It asks the auditors to document not just whether a project is real but whether it leans on taxpayer-funded incentives, how much water it would draw, and whether it plans to bring its own on-site generation rather than lean on the shared grid. Those are the questions of someone trying to separate load that pays its own way from load that quietly bills the public. The audit is, at bottom, a cost-allocation exercise dressed as a reliability review.

The number that decides it

Strip away the drama and one number decides whether Texas has a crisis or a clerical problem: the conversion rate. Of 474 gigawatts requested, how many will actually energize and draw power? Nobody credibly knows yet — that is the entire point of the audit — but the gap between the queue and the grid tells you the answer is a small fraction. If even a fifth of it were real, Texas would still be facing the largest load growth in its history. If it is a tenth, the story is manageable. The queue does not tell you which. It only tells you the number is too big to plan against honestly, which is why planning against it honestly required stopping to look.

A request that costs nothing to file tells you nothing about demand. It only tells you the option was free.

This is the unglamorous lesson under the headline, and grids that have always been tight learned it long ago. In markets where power is genuinely scarce — much of the world, in fact — you cannot afford to build ahead of demand you have not verified, because the capital is too dear and the ratepayers too stretched to absorb a mistake. A connection request in those systems tends to come with money attached: deposits, milestone payments, penalties for holding a place you do not use. Make the reservation cost something, and the queue shrinks to roughly the size of the real demand, because the speculative requests stop being free. Texas is arriving, late and at enormous scale, at a discipline that constrained grids never had the luxury to skip.

Not just Texas

The reason this is worth your attention beyond one state is that the incentive that produced the ghost queue is everywhere. Pennsylvania's governor moved in the same direction weeks earlier, signing an order that tightens permitting for data centers above 25 megawatts and demands more disclosure of who the end users really are. Interconnection queues across the country are swollen with the same mixture of real projects and free options, because the rules that let you reserve grid capacity for almost nothing are broadly similar. Texas is simply the first big hub to freeze the line and count, because Texas is where the mismatch grew most absurd — five times peak demand is hard to look away from.

What comes out of the December audit will matter less as a Texas headline than as a template. If the fix is to make interconnection requests carry a real financial commitment — a deposit that hurts to forfeit, milestones you have to hit to keep your place — then the giant queue numbers that have been quoted all year, in Texas and everywhere else, will quietly deflate to something closer to the truth. That would be good news badly disguised as a slowdown. The AI build-out is real, and it needs an enormous amount of power. But you cannot build a grid against a wish list, and the first honest thing you can do with a wish list is stop pretending it is a forecast.

• The mismatch: ~474 GW of large-load requests, ~90% data centers, against a grid that peaked at ~91 GW in July 2026 — more than 5x.
• The growth: requests rose from ~48 GW in 2023 to >474 GW, across more than 1,800 projects, many lacking financing, tenants or land.
• The mechanism: an interconnection request has been nearly free to file, so a valuable, scarce reservation drew speculative and duplicated claims — 'ghost demand'.
• The exposure: grids get built against the queue; overbuild for load that never arrives and ratepayers fund the poles and wires, raising cost per kilowatt-hour for everyone.
• The fix, likely: make reservations cost something — deposits, milestones, penalties — so a request means demand, not a free option. ERCOT aims to finish its audit by December.

The freeze will be reported as Texas slamming the brakes on data centers, and in the near term it does slow new connections. But read it as an accounting decision and it is more interesting than that. Texas did not discover that it lacks power. It discovered that it could not tell its real demand from a set of free reservations, and that the cost of guessing wrong lands on the wrong people. The 474-gigawatt number was never a measure of how much electricity the state needs. It was a measure of how little it used to cost to ask.