Who really needs all these data centers?

The efficiency decade

The 2020 study in Science that recalibrated the numbers found three things happening at once from 2010 to 2018: processors got far more efficient and wasted less power idling; work consolidated from small, badly cooled company server rooms into hyperscale buildings where the overhead for cooling and power delivery fell to a few percent; and storage moved to drives that use a fraction of the energy per terabyte. Compute rose 550%. Electricity rose 6%. Earlier forecasts, including some that predicted data centers would consume a fifth of world electricity by the 2020s, had simply extrapolated the growth of computing without the efficiency.

What changed in 2019

Training and running large AI models uses graphics processors that draw 700 to 1,200 watts each, packed into racks that draw 100 kilowatts or more where a conventional rack drew 5 to 10. LBNL's 2024 report attributes most of the growth from 2017 to 2023 to this class of server. Its 2025 update puts the 2030 reference case at 649 TWh, 11.8% of U.S. electricity, with a range of 9.5% to 15.3% depending on how many chips ship and how hard they run.

Why the forecasts disagree so much

Two inputs dominate: how many AI chips ship, and how much of the time they run. A shorter assumed chip lifetime cuts LBNL's 2030 estimate to 590 TWh; higher idle power raises it to 782. The IEA, modeling the same country, lands at 426 TWh for 2030. The honest statement is that nobody knows within a factor of two, which matters when utilities are using the high end to justify thirty-year investments.

Global context

Worldwide, data centers used about 415 TWh in 2024, 1.5% of electricity, and the IEA expects about 945 TWh by 2030. The U.S. is 45% of the world total, more than China and Europe combined, which is why the fight over who pays is mostly an American fight.

How a request becomes a forecast

A developer with a site option and no tenant files a request for 500 megawatts with the local utility. It files the same request in two other states to see who offers the cheapest power and the biggest tax break. Each utility adds 500 MW to its forecast. The forecast goes into the regional grid operator's plan, which goes into the capacity auction (see §3), which sets the price every household pays. Three "phantom" data centers have now raised your bill before a shovel has moved. Grid Strategies found that utility forecasts collectively overstate data-center-driven growth by about 40% against a proprietary database of real projects.

The incentive

A regulated utility earns a guaranteed return, usually 9 to 11%, on every dollar it invests in plants and wires. More forecast means more investment means more profit. The industry's trade group expects its members to spend $1.1 trillion from 2025 to 2029, nearly matching the previous ten years combined, and is open that data centers are the reason. A forecast that turns out to be too high is not the utility's loss; the plant is built and ratepayers pay for it either way. A forecast that is too low gets a utility blamed for blackouts. The asymmetry points one direction.

What a real forecast looks like

Ohio made AEP require new data centers to pay for 85% of the power they asked for whether they used it or not. The request list fell from 30 GW to 13 GW almost immediately. Texas ordered an audit of every project in its 474 GW queue in August 2026 and then halted new environmental permits for data centers in September until it finishes. Federal regulators ordered every regional grid operator in June 2026 to justify how it handles large-load requests. The pattern in every case: ask developers to put money behind the number, and the number shrinks.

Who the load factor helps

Utilities assume data centers run flat out, 24 hours a day, near 100% of their rated power. Real facilities, even AI ones, average well under that. Grid Strategies flagged that some forecasts use "unrealistically high load factors," which inflates energy forecasts even more than peak forecasts. Every percentage point of assumed utilization is another plant someone gets to build.

How a capacity market moves your bill

Your utility buys enough "capacity" to cover its share of forecast peak demand plus a reserve. When the forecast jumps because of data center requests, the utility must buy more capacity at auction, supply is fixed in the short run, and the price clears at the cap. The charge is passed through to every customer in proportion to use. A household did nothing different and pays more. The market monitor estimated data centers caused 63% of the 2025/26 price increase ($9.3 billion) and 40% of the 2027/28 cost ($6.5 billion).

The rules that decide who pays

The counter-evidence, honestly

Rutgers' policy lab reviewed residential bills across 24 states in 2026 and found no large, detectable increase attributable to data centers yet; its title was "Mostly Not. Yet." An EPRI study argued that from 2015 to 2024, data centers slightly lowered average rates by spreading fixed costs over more sales. Both are consistent with the PJM story: the costs are front-loaded into capacity and transmission charges that reach bills on a lag, and they are concentrated in the regions building the most. The question is not whether you are paying today but whether the rules will stop you from paying tomorrow.

What else shows up

Gas plants. Entergy will build ten gas-fired units, more than 7 gigawatts, to power Meta's Louisiana campus. Meta pays for the first 15 years; the plants are financed over 30. Consumer advocates asked who pays the second half. In Memphis, xAI ran dozens of gas turbines beside its facility before permits for most of them existed. Water: U.S. data centers consumed about 17 billion gallons directly in 2023 and 211 billion indirectly through the power plants that feed them. Google's Oregon facilities alone used 550 million gallons in 2025, about 40% of The Dalles' total.

The stadium parallel

Economists have studied publicly funded sports stadiums for forty years and reached one of the strongest consensuses in the field: they do not pay for themselves. In a 2005 survey of American Economic Association members, 85% said state and local governments should stop subsidizing franchises; in a 2017 University of Chicago panel, 80% agreed subsidies cost taxpayers more than they return. Cities build them anyway, because the ribbon-cutting is visible and the cost is spread thin across decades. Data centers are the stadium with the ribbon-cutting removed: a billion-dollar building, a few hundred jobs, a thirty-year tax abatement, and no one in the seats.

Why data centers are a worse deal than stadiums

• The subsidy scales with the hardware, not the jobs. A sales tax exemption on chips grows every time chips are replaced, which in AI facilities is every three to five years. Texas's exemption grew from a $29 million projection to a $1.3 billion cost because of chip purchases nobody modeled.
• Half of the programs don't require a single job. Sixteen of 36 state data center subsidies had no job-creation requirement at all. Of Texas's 138 exempt facilities, 20 have been audited and six of those failed the job rules.
• Local governments lose money they never voted on. A state exemption usually wipes out the local share of sales tax too. Georgia's counties and cities are projected to lose $1.1 billion in 2026 to a break the state legislature granted.
• The subsidy isn't what decides the location. Industry surveys find 3% of operators call incentives the biggest siting factor; power and land are. North Carolina's governor put it plainly in 2026: "The market is already delivering the incentives."

The Foxconn precedent

In 2017 Wisconsin promised about $4 billion in incentives for a Foxconn factory that was to employ 13,000 people in Mount Pleasant. The village and county borrowed hundreds of millions to prepare the site. The factory was never built as promised and employment peaked at about a tenth of the pledge. Microsoft now owns part of the site and is building a data center on it. In September 2026, under its "community-first" pledge, Microsoft waived the $5 million annual incentive the village had offered and redirected it to local tax relief, the first time a Wisconsin economist could recall a company returning one. It is also a measure of how far the ground has shifted: the largest tenant in the industry now advertises that it will pay full property tax and its own grid costs, because communities stopped saying yes.

What the chips are really for

A hyperscale AI building is 55 to 60% chips by cost. Those chips are bought from one company at a 71 to 75% gross margin, depreciate in a few years, and are tax-exempt in 37 states on every replacement cycle. The building that holds them is tax-abated for decades. The electricity that runs them is, in Texas and several other states, also exempt from sales tax. Strip the exemptions out and a 1-gigawatt campus is a $40 to $50 billion purchase of taxable equipment; with them, it is a $40 to $50 billion purchase on which the public collects almost nothing and provides the roads, the water and, through the capacity market, part of the power plant.

The utility business model, in one paragraph

An investor-owned utility does not make money selling electricity; it makes money owning things. Regulators let it earn an allowed return, typically 9.5 to 10.5%, on the capital it has invested in plants and wires, recovered through rates. Growth in that "rate base" is the whole earnings story. NiSource targets 9 to 11% rate base growth a year through 2033 and credits data centers. Sempra plans a record $65 billion through 2030, mostly in Texas, and expects its rate base to nearly double. The trade association expects $240 billion of utility capital spending in 2026 alone. Every one of those dollars earns a return from ratepayers whether the data center that justified it ever draws power.

Who owns the buildings

Most hyperscale capacity is leased, not owned, by the tech companies. The landlords are data center REITs (Equinix, Digital Realty), private equity-backed developers (QTS under Blackstone, Vantage, Crusoe, Aligned) and, increasingly, former bitcoin miners who already had power contracts. The leases run 15 to 20 years with investment-grade tenants, which makes the buildings easy to finance with debt. If the tenant's AI business does not materialize, the lease is still owed; if the tenant walks, the building is a very expensive shed in a county that gave up its property tax to get it.

The cancellations

Microsoft was reported to have walked away from leases totaling a couple of gigawatts in early 2025, which analysts read as oversupply against its own demand forecast. Oracle and OpenAI dropped the Abilene expansion in March 2026. Industry trackers put the share of U.S. 2026 capacity that has been delayed or canceled somewhere between a third and half, though the figures mix opposition, permitting, financing and power delays. None of this has slowed guidance: every hyperscaler raised its 2026 spending plan at least once during the year, and the stated reason is often not demand but the rising price of memory chips.

The jobs number, honestly

Construction employment is real and large: 7,500 at the peak of Meta's Louisiana build, thousands at Abilene, enough to cause a housing crisis in a city of 130,000. It is also temporary and largely imported. The permanent payroll of a hyperscale campus is a few hundred technicians and security staff. A typical measure of what the public gets per dollar, permanent jobs per billion invested, runs from 4 to 50 for data centers against roughly 1,000 for an auto plant.

The asset was never the bitcoin

A bitcoin mine is a power contract with computers attached. Miners went where power was cheapest and least regulated, mostly Texas, signed long fixed-price contracts with generators, and enrolled in the grid's demand-response programs, which pay large customers to drop load when the grid is stressed. In a heat wave, a miner earns more by selling its contracted power back to the grid than by mining. Senators Warren and Markey called the payments subsidies; the industry calls its load a "demand-side battery." Both are describing the same transfer: from everyone's bill to a company that agreed to turn off something it only turned on because the power was cheap.

The tax break came free

Most state data center sales tax exemptions were written before anyone imagined a building full of ASICs, and most do not exclude crypto mining. A miner in a qualifying state buys its machines and often its electricity tax-free, like any other "data center." Good Jobs First's first recommendation for any state that keeps its exemption is to disqualify crypto facilities from it.

The pivot

When the AI boom arrived, miners discovered their power contracts and interconnections were the scarcest asset in the industry. Core Scientific, TeraWulf, Cipher, IREN and Riot converted mining halls into GPU hosting and signed leases with AI companies that dwarf their old revenue. Riot's first tenant was AMD at 50 MW; its second was a 20-year, 191 MW lease with Anthropic worth about $9.1 billion through 2048. The county that approved a bitcoin mine with a tax abatement now hosts an AI campus with the same abatement, and the municipality that paid to upgrade a substation for a business that could switch off at any time now serves one that cannot.

The noise

Mining and AI facilities both run tens of thousands of fans. Residents of Granbury, Texas recorded about 85 decibels, the level of a running blender, from Marathon's 300 MW mine beside a gas plant, and reported migraines, vertigo and hearing damage. A court allowed their suit to proceed in 2025. Texas does not allow counties or its environmental agency to regulate noise, which is one reason the mine is in Texas.

Two different loads

Inference, answering your prompts, is small per task and large in aggregate only at planetary scale. Even a heavy user's day is under a tenth of a kilowatt-hour. If every American ran the calculator's "typical" day, the country's AI use would be roughly 7% of what data centers already consume. Training is the opposite: a handful of runs a year, each drawing 100 megawatts or more continuously for months, with the next generation planned at a gigawatt or more. Training is what the gigawatt campuses in Texas, Louisiana, Georgia and Indiana are for, and it is a bet on models that do not exist yet.

Built ahead of demand

The chips are bought before the customers are. Microsoft said in late 2025 that it had chips in inventory it could not plug in because buildings and power were not ready; Oracle built the Abilene campus for one customer whose revenue cannot yet cover the contract; Meta's Louisiana campus is being sized to 5 gigawatts for products nobody has used. None of that is a shortage in the sense a household experiences one. It is capacity purchased on the expectation that demand will arrive, financed partly by the public, with the downside held by ratepayers and counties if it does not.

What the per-task numbers hide

Published per-prompt figures are mostly for text and for a company's own median; a long reasoning answer can use ten to fifty times the median. Video generation is the category to watch: measurements span a 3,000-to-1 range depending on the model and settings, and the expensive end is the one being marketed. Image and video are also where the "generated imagery" demand is growing fastest. The fair summary: text is nearly free, images are cheap, video is not, and the aggregate depends on how much of it companies can convince people to make.

What would "need" look like?

A build that followed demand would show chips running near capacity before the next campus is announced, forecasts that match shipments, and tenants paying for their own power. What the record shows instead is utilization nobody publishes, forecasts 40% above shipments, canceled expansions, and the public asked to underwrite the gap. Whether the bet pays off is a question for investors. Who holds the downside is a question for voters.

Not every "restriction" restricts. Several of the 2026 state laws are ratepayer-protection measures that make the buildout easier to approve by moving its cost off households, which is the industry's preferred outcome. The projects that actually stopped were stopped by zoning boards, county commissions and city councils, usually over water, noise, farmland or an abatement, and usually after residents found out what the NDA was hiding.