Two-Thirds of AI Data Center Power Demand Is Phantom and Won't Materialize

The number that has dominated the AI infrastructure conversation for the past two years — the terrifying, grid-breaking total of gigawatts requested for new data centers — turns out to be mostly fiction. Not because AI’s power appetite is fake, but because the way electricity gets requested and the way it actually gets built are two very different processes, and the gap between them is enormous. Bloomberg put a number on that gap, and it changes the story.

How much of the requested power is actually going to be built?

About 28%. Bloomberg reported on August 12, 2026, that US grid operators and utilities are likely to actually commit to only about 28% of the 1,066 gigawatts requested for data center projects. More than two-thirds of the electricity sought for the AI data center boom is not likely to materialize.

That 1,066-gigawatt figure is the number that has been driving headlines, policy debates, and genuine public anxiety about AI’s impact on the power grid. It is a real number — those requests were actually filed. But Bloomberg’s reporting reveals that the relationship between “requested” and “built” is not the straightforward pipeline most coverage has treated it as. The majority of those requests are what the reporting describes as “phantom” projects and long-shot pitches — filings that were never going to become operational data centers.

Does this mean AI’s power demand is not a real problem?

No. This is a precision story, not a dismissal story. The nuance matters: AI’s power demand is real, significant, and growing — but the specific number most commonly cited to quantify it is a poor proxy for what will actually get built and stress the grid.

The 28% that grid operators expect to commit to is still a very large amount of electricity. Twenty-eight percent of 1,066 gigawatts is roughly 300 gigawatts of likely committed capacity — which would be a massive expansion of the US data center footprint by any historical standard. The communities, utilities, and grid operators dealing with actual construction in their jurisdictions are facing real challenges around water use, electricity costs, and infrastructure strain.

What the Bloomberg reporting corrects is the framing, not the underlying concern. When a headline says “the AI industry is requesting a thousand gigawatts,” the reader reasonably concludes that a thousand gigawatts of data centers are being built. They are not. A fraction of that — still enormous, still consequential — is what the grid will actually have to absorb.

Metric	Figure
Total gigawatts requested for US data center projects	1,066 GW
Percentage likely to be committed by grid operators	~28%
Percentage unlikely to materialize (“phantom” demand)	~72%
Implied actual committed capacity	~300 GW

Why do developers file so many phantom requests?

Because filing is cheap, building is expensive, and optionality is valuable. The dynamics are not unique to data centers — they resemble how land developers file permits for projects they may never build, or how airlines request landing slots they may not use.

The result is a requests pipeline that functions more like a funnel than a conveyor belt. Everything enters at the top — every serious project, every hedge, every long shot — and the number at the top tells you very little about what comes out the bottom. Grid operators and utilities know this, which is why their expected commitment rate is 28%, not 100%. The public conversation has been reading the top-of-funnel number as if it were the bottom.

What are the major players actually doing on the ground?

While the aggregate demand picture is inflated, individual commitments continue. OpenAI announced Project Camellia, a long-term data center project in Effingham County, Georgia. For a company that has historically relied on Microsoft’s data center capacity, a direct infrastructure commitment signals that major AI labs are beginning to build their own physical compute footprint rather than renting it.

Meanwhile, in Q2 2026 earnings calls, Microsoft, Alphabet, and Meta reportedly shifted their language in a telling direction. Rather than leading with aggregate capital-expenditure totals — the “we’re spending $X billion on AI infrastructure” headlines that dominated 2024 and 2025 — they reportedly moved toward discussing time-to-energy (how quickly they can bring power online for new capacity), large-scale networking, power procurement, and how fast they can convert infrastructure into revenue-generating compute.

What does this mean for the “AI is eating the grid” narrative?

It means the narrative needs a precision upgrade, not a retraction.

The concern about AI’s impact on the power grid is legitimate. Data centers consume real electricity, real water, and real land. Communities are right to scrutinize proposed projects and demand honest accounting of the costs. The 28% commitment rate does not make those concerns disappear — 300 gigawatts of new data center capacity is still a generational expansion of the country’s power infrastructure.

But the 1,066-gigawatt headline number has been doing work it was never suited for. It has been cited in policy debates, used to justify emergency grid planning, and invoked to argue that AI’s power appetite is an existential threat to the electrical grid. Bloomberg’s reporting shows that the number includes so much phantom demand that it is closer to a wish list than a forecast. Policy made on the basis of a wish list — whether to approve or oppose data center construction — is policy built on sand.

1. The headline number is real but misleading

   1,066 gigawatts were genuinely requested for US data center projects. That number is not fabricated. But “requested” and “will be built” are not synonyms, and treating them as such has distorted the public conversation.

2. The actual grid stress is significant but smaller

   About 28% of requests — roughly 300 gigawatts — is what grid operators expect to actually commit to. That is still an enormous number that will strain grids, raise electricity costs in some regions, and require real infrastructure investment. It is just not a thousand gigawatts.

3. The demand is real; the panic-number is not

   AI data centers are being built, will consume significant power, and will affect communities. The correction is to the scale of the projections, not to the existence of the demand. Better data leads to better policy — both for communities evaluating proposals and for developers planning infrastructure.

Why this matters for anyone following AI infrastructure

Because the most commonly cited number in the AI power debate turned out to be roughly three and a half times larger than reality. That matters for grid planning, for community opposition movements, for utility rate-setting, and for anyone trying to forecast how fast and how expensive AI compute becomes over the next decade.

The bottom line

The AI data center power story needed a reality check, and Bloomberg provided one. More than two-thirds of the electricity requested for the US AI data center boom is phantom demand that will never materialize. The remaining 28% — about 300 gigawatts — is still a massive expansion that will reshape grids, raise legitimate community concerns, and require billions in infrastructure investment. The story is not that AI’s power demand is fake. The story is that the number everyone has been using to describe it is roughly three and a half times too large, and better numbers lead to better decisions — for communities, utilities, policymakers, and the AI industry itself.