Phantom Power: What 700 Gigawatts of Data Center Requests Reveal About AI

On September 1, 2026, Reuters published a number that sounds like a typo: data centers and other large consumers have requested grid connections exceeding 700 gigawatts in parts of the United States. According to Reuters, that is more than ten times what the industry actually consumes in data center power today. And the requests from the middle band of the US roughly equal the total power demand of all US households combined. I read the analysis twice because I did not believe it at first. Then I checked the numbers against official sources and found confirmation: the US energy agency expects a record consumption for 2027, Texas has frozen new grid connections for data centers, and one utility after another slashed its demand forecast the moment it started requiring deposits. Exactly this mechanism is the story, and I gave it a name: phantom power.

My thesis is uncomfortable: the AI industry has handed the power grids a bill it will not pay itself. Not out of malice, but because the market currently confuses requests with ambition. When a request is free and a deposit only becomes due once you are serious, you file as many requests as possible. Grid operators then plan lines and power plants, and in the end electricity customers pay for capacity that no server ever touched. Reuters calls it an illusion, Texas treats it as a screening process, I call it phantom power. And I counter it with the only operating budget I know completely and can measure myself: my own.

The analysis is the core of the report, so I will summarize it here. Reuters evaluated grid data and statements from utilities. The requests from data centers and other very large power consumers in the middle band of the US, meaning parts of the Midwest, the Mid-Atlantic region and the South, add up to more than 700 gigawatts. For scale: total current US data center capacity is estimated by the industry at roughly one tenth of that sum. For Texas alone, grid operator ERCOT saw requests grow from about 48 gigawatts in 2023 to more than 474 gigawatts. Outside Texas, ten of the largest utilities in these regions, among them AEP Ohio, Southern Company and PPL, reported a combined roughly 270 gigawatts of requests, according to Reuters from their recent quarterly investor calls. The industry's core problem is stated in the report in black and white: there is no uniform standard for how utilities count these requests. Some count only projects with a signed contract, others count every request without any commitment. Whoever compares such counting methods is not comparing numbers but moods.

The most interesting finding is what happens when grid operators demand seriousness. Texas has become the first major data center location to freeze new grid connections for data centers and started an audit of which projects are real at all. Governor Greg Abbott pulled the emergency brake in early August 2026, as reported by Tagesschau. The effect was immediately measurable: the US energy agency's forecast for Texas power demand in 2027 fell from plus 14 percent to plus six percent. In Ohio, AEP Ohio, after new state rules introduced study fees for grid studies of up to 100,000 US dollars, saw its data center power request pipeline shrink by more than half. And Exelon cut its classification of probable data center demand by around 40 percent to 11 gigawatts in an investor presentation dated July 30, 2026, after the company demanded stricter security deposits. The pattern is so simple it is almost insulting: demand an upfront payment, and the fantasy requests disappear. I covered the Texas building freeze and the grid connection backlog in a separate post, Texas pauses data center permits, and this analysis adds the new numbers to it.

The sequence can be pinned to dates. In 2023, the request backlog at Texas grid operator ERCOT stood at roughly 48 gigawatts. Then came the boom: according to Reuters, planned spending on AI data centers rises to over 700 billion US dollars this year. Companies and landowners with access to power or grid connections began translating that wave into requests, sometimes without money, without a project and without expertise, as the report puts it. In March 2026, Thomas Gleeson, head of the Texas regulator Public Utility Commission, told an industry conference in substance: if you do not know what is real, you do not know what to build the infrastructure for. In July 2026, New York became the first state to temporarily pause construction of further hyperscale data centers. On July 30, 2026, Exelon cut its probable demand by 40 percent. In early August 2026 came the Texas building freeze, and on September 1, 2026, Reuters published the full analysis. Six weeks, four interventions, one pattern.

Why would anyone file requests for 700 gigawatts without owning a single acre? Because the application is free and the queue rewards whoever files first. Reuters describes companies and landowners speculating on the boom, projects without financing, and requests parked at several utilities at once because the applicant provisionally secures everywhere. The incentive is asymmetric: a rejected request costs nothing, but whoever arrives late loses the connection. Regulators therefore warn that unrealistic forecasts raise electricity prices and complicate grid planning, because power plants and lines are built based on these numbers. For me this is the same mechanism I keep observing in this industry: numbers get reported because they are useful, not because they are verified. Benchmarks, user counts, now grid connections. The only difference is that in the end the electricity customer pays, through grid fees, for capacity that no server ever touched.

Germany cannot watch this topic comfortably from the sidelines. The Federal Network Agency now expects 78 to 116 terawatt hours of power consumption for data centers in 2037 in its scenario framework, up to ten percent of German electricity consumption, and has revised the estimate sharply upward recently. The AI gigafactory the German government wants to attract is not even included in those numbers. Germany has around 2,000 data centers, more than a third of them in the Frankfurt region, and there major new connections will not be available until the mid-2030s according to the grid operator; I wrote about this in Frankfurt is full: why data centers wait. The climate balance comes from a study by Allianz Trade: data centers caused 286 million tons of CO2 worldwide in 2025, 57 percent more than earlier estimates by the International Energy Agency, more than 70 percent of it from power consumption alone. Location matters: German data centers emit 329 grams of CO2 per kilowatt hour according to the study, while Norway and Sweden stay below 30 grams. That more than half of respondents in a survey worry about the water consumption of data centers no longer surprises me given these numbers.

Now for the counter-evidence I can contribute, because it stands on my desk. I run my agent workload on a used Google Pixel 6a that cost 45 euros, without root, and I have been measuring the operation for weeks instead of claiming it. The long-term test at a stretch: 88 hours of runtime from August 28, 2026, 7:40 PM, to September 1, 2026, 12:18 PM, with zero OOM kills since the hardening on August 28. Over 7 days the liveness rate was 99 percent, 170 of 171 hourly checks. In the snapshot of September 7, 2026, 7:00 AM, we stand at 100 percent liveness, 306 of 306 checks, zero kills and 917 MiB of available RAM with 11 running jobs.

The price of this operation is small when set against the debate. Processing a complete 54-minute video cost 0.5 percent of my 5-hour window on September 4, 2026, according to a dashboard comparison. My container environment needs around 380 MB of RAM as a snapshot, measured without a defined worker load, that is the honest limit of the claim. On this path, two projects have been verified with a combined 36 of 36 passed tests. I am of course not claiming a phone replaces a data center: training happens in data centers, and for that the world continues to need the large facilities. But agent orchestration, verification and everyday work run on this deck in megabytes and percent windows, not in gigawatts. The entire professional agent stack therefore already runs on a 45 euro phone, and Android killing processes when things get tight taught me more about this than any forecast. Every gigawatt that evaporates as phantom makes the math for central mega-facilities shakier and the math for decentralized edges more sensible.

So this does not remain a question of belief, here is my test plan. First, the quarterly investor calls of the large utilities, starting with Exelon and AEP Ohio, and the question whether the pipelines keep shrinking after deposits. Second, the queue of Texas grid operator ERCOT and the outcome of the audit of which projects survive screening. Third, the Federal Network Agency's scenario framework and its next revision. Fourth, the short-term forecast of the US energy agency, which still expects 4,391 terawatt hours for 2027 and thus holds the record regardless of the phantom dispute. I commit myself: I expect the next deposits to bring the next cut. If requests remain stable despite deposit requirements, there was more genuine demand than I assume, and then I will write a correction here.

I am Marcel, a graphic designer, and I run the HUNTER cyberdeck on a used Google Pixel 6a in daily operation. This post draws on the Reuters analysis of September 1, 2026, on the energy forecasts of the US energy agency and the Federal Network Agency, and on the stability snapshots and measurement log of this deck, as of September 7, 2026. Last technically reviewed on September 7, 2026.