Five companies—Microsoft, Alphabet, Amazon, Meta and Oracle—will spend between $660 billion and $690 billion in capital expenditure this year, almost all of it on data centers built to train and serve artificial intelligence. Amazon’s share is roughly $200 billion. Alphabet’s is $175 billion to $185 billion. Meta’s is $115 billion to $135 billion. The numbers keep climbing, and the markets financing them have barely started pricing the risk and costs that comes with them.
Stargate, the OpenAI-led joint venture announced in January 2025, said last September it had reached nearly seven gigawatts of planned capacity across Texas, New Mexico and Ohio. It has since blown past nine gigawatts and picked up Wisconsin and Michigan along the way. Set that against any private infrastructure program in American history—the transcontinental railroads, the fiber buildout of the late 1990s—and this one is bigger, faster and far more concentrated.
“Compared to more traditional load growth,” Federal Energy Regulatory Commission member David Rosner told the commission at its June open meeting, “the large loads seeking to connect to the grid today are larger, sometimes by orders of magnitude, and more concentrated.” Federal regulators do not reach for “orders of magnitude” casually.
The pro formas circulating through investment committees right now look like beautiful plans. Land, shells, chips, cooling systems, power purchase agreements, depreciation schedules, offtake risk; all modeled, all sourced, all footnoted. What they rarely model is the part of the bill that never touches the developer’s balance sheet. These include transmission upgrades spread across people who have never bought a GPU, water pulled from basins that were already having a rough decade, thirty years of municipal tax revenue waved cheerfully goodbye, and a political backlash that has spent 2026 rewriting the rules on projects that are already pouring concrete.
Those costs are real, and 2026 is the year they started coming home for big tech. Large-load tariffs, repealed tax abatements, groundwater permits and federal interconnection reform are all doing the same thing: moving the bill off the public ledger and onto the projects themselves.
The timing is unkind. This repricing arrives as hyperscaler debt has tripled as a share of capital spending and the industry’s own accountants are arguing about how fast a GPU dies. The rest of this story is that bill, item by item.
Power Imbalances
Start with the capacity price, because it is the one number nobody can spin. PJM Interconnection, the grid operator for more than 67 million people across 13 states and the District of Columbia, pays generators to promise they’ll show up when demand peaks. In the 2024/25 delivery year that price was $28.92 per megawatt-day. For 2026/27 it is $329.17—an increase of more than 1,000%, essentially all of it in a single auction.
Monitoring Analytics, PJM’s independent market monitor, attributed 63% of the jump in the 2025/26 auction to data centers—roughly $9.3 billion recovered from customers in higher rates. A broader estimate, using a different measure over a longer period, puts data center demand behind some $23 billion in customer price increases and expects the pressure to persist until at least the end of 2028.
The utilities aren’t pretending otherwise. “Residential supply costs in the Mid-Atlantic have increased by up to 80% or more over the past five years,” Exelon chief executive Calvin Butler told analysts in May. “Without addressing supply constraints, affordability challenges will persist.” Translated: your bill went up, it will keep going up, and the utilities would like the record to show it wasn’t their idea.
The politics arrived on schedule. Speaking at PJM’s own annual meeting that same month, Maryland Governor Wes Moore told the room that “the single largest driver of capacity price increases is strain on the system from large loads and data centers,” and that “data centers must pay their own way.”
This can be harder that it looks. Developers shop the same project to several utilities at once, and grid operators can’t easily tell a financed campus from an optimistic spreadsheet. Brian Fitzsimons, whose firm GridUnity sells software into this market, says one of the largest U.S. utilities watched nearly 30% of its 2024 applications evaporate. Utilities plan, procure and build against those forecasts anyway. If the load never turns up, the transmission is still there, still expensive, and somebody still amortizes it.
Historically that somebody has been the ratepayer. “Ohio households should not be asked to subsidize speculative infrastructure investments for extraordinary private load growth,” Maureen Willis, director of the Office of the Ohio Consumers’ Counsel, told a state legislative committee in May.
The rules are being rewritten to stop precisely that. In June the Federal Energy Regulatory Commission sent “show cause” orders to the six regional grid operators, opening what Rosner called “a dialogue.” The intent, in his words, is that “if new infrastructure is built to accommodate a data center, and that data center doesn’t show up, residential customers are not left on the hook to pay the costs.” Nothing is in force yet, and the industry will fight it, because of what the fix actually does: it moves the stranded-asset risk onto the data center developer, and onto whoever financed the developer.
Water Hoarding
Water is where both sides of this argument are least informed. The current crop of U.S. data centers directly consumed about 17.4 billion gallons in 2023. Lawrence Berkeley National Laboratory puts the water consumed upstream—generating the electricity they drew—at another 211 billion gallons, twelve times as much. That figure includes evaporation off hydropower reservoirs as well as thermal plant cooling, and the Information Technology and Innovation Foundation, the industry-friendly think tank that most recently compiled it, believes the number is probably too high. The direction survives the quibbling: most of a data center’s water footprint sits in the power plants behind it rather than in its cooling towers.
But the smaller number is the one that lands in somebody’s actual aquifer. Roughly two-thirds of data centers fully built or under development since 2022 sit in water-stressed areas. That is the entire problem in one sentence: a facility drawing a million gallons a day is a rounding error nationally and a crisis locally. Easier siting options exist, but developers keep targeting places where water is already at a premium—Virginia, Ohio, and further inland, Kansas and Iowa.
Virginia did the arithmetic. A state environmental study released in July concluded that “under current conditions, it appears unlikely that a data center with evaporative cooling technology (or any comparable water user) would find a reliable, sufficient groundwater supply anywhere” in the eastern coastal plain. Virginia currently hosts 371 operating data centers, with another 438 planned, including one just announced in Fairfax. You can read those two sentences in either order. Neither improves the other.
The deeper problem is that nobody knows what’s actually being used, in part because so few will say. Of the 341 facilities the Texas Water Development Board surveyed for its 2025 water use report, 17% replied. The board’s water supply planning director, Temple McKinnon, described reconstructing the rest as “forensic accounting.” Asked at a June hearing what governs a data center pumping groundwater in a county with no conservation district, she didn’t hedge: “There would be no entity in place to regulate that use.”
Representative Brad Buckley summed it up: “We don’t have the fundamental data we need to make decisions moving forward, nor does the Texas Water Development Board, nor does ERCOT, nor does PUC.” Texas is planning its water future on an a industry where 83% of the participants operate in the dark.
If you’re holding paper on any of this, that’s the finding that matters. A campus underwritten on evaporative cooling, in a basin that can’t support it, doesn’t keel over. It gets redesigned at cost, or moved at greater cost. Air-cooled alternatives use no water but burn more power running fans, and lose efficiency at precisely the moment it gets hottest outside.
Tax Break and Build
When it comes to tax subsidies, Texas keeps better receipts. And they are shocking,
When the legislature created its data center sales tax exemption in 2013, the state figured it would cost about $14.6 million across the 2014–15 budget cycle. The comptroller’s 2025 forecast put it at $3.2 billion over the following two years—a number the comptroller’s office conceded was almost certainly an underestimate, given how many new facilities were lining up. The Senate Finance Committee now projects $3.3 billion for 2028–29.
The statute was written for server farms. AI campuses are orders of magnitude larger, measured against thresholds that never moved. A facility above 100,000 square feet must create 20 permanent jobs paying 120% of the area’s median salary and invest $200 million over five years. Twenty jobs.
Of 138 certified recipients, 20 have been audited; six were found in breach, at least four of them for failing the jobs requirement. Asked at a Senate Finance Committee hearing in July how many jobs a data center actually creates, the comptroller’s chief revenue estimator, Brad Reynolds, replied: “That, I don’t know. I just know that clearly some of them have difficulty achieving the 20.”
Some of them have difficulty achieving the 20. “No tax exemption should operate on autopilot,” committee chair Senator Joan Huffman told the room. Autopilot would be an improvement. Autopilot implies somebody once set a course.
The national pattern is consistent. By Good Jobs First’s count, 16 of 36 state data center subsidy programs require no job creation whatsoever, and the three states that measured their returns are losing between 52 and 91 cents on the dollar.
The industry’s figures tell a different story. A PwC study commissioned by the Data Center Coalition credits data centers with $65.8 billion in Texas GDP in 2024 and more than 103,000 direct jobs—428,000 once indirect and induced employment is counted. Both things can be true at once: the economic activity is real, the permanent on-site employment is thin, and the reporting was structured to reflect neither.
What changed in 2026 is that legislatures started reading their own statutes. Massachusetts halted its incentive in June, Arizona enacted a three-year pause, Illinois and Ohio suspended theirs, and Governor Greg Abbott—who spent a decade recruiting these projects—has said that if reelected he’ll work to repeal the Texas exemption in 2027. Capital committed against a twenty-year abatement is being repriced in year four.
The Land, the Noise and the NDAs
The local objections stopped being local some time ago.
Data Center Watch—a research project of 10a Labs, a consultancy that works for AI companies, which is worth knowing before you weigh its numbers—counted 75 projects blocked or delayed in the first quarter of 2026, disrupting some $130 billion in project value. That is roughly five-sixths of the $152 billion it logged across all of 2025, in a single quarter. It puts active opposition groups in 49 states, up from 42 at the end of last year. More than 300 data center bills were filed in state legislatures in the first six weeks of the year, and of 63 local moratorium actions introduced, considered or adopted, 54 have passed.
Some of the grievances are sitting. Virginia’s legislative auditors found that a third of the state’s data centers sit near residential areas; in Fairfax County, 55% are within 200 feet of somebody’s house. Prince William County residents have complained of noise routinely topping 60 decibels. When the county tightened its industrial limits in 2,025 it landed on 73 decibels by day—above the 67 its own consultants recommended.
More of it is a process. Roughly 80% of Virginia localities hosting data centers have signed non-disclosure agreements with developers, and one agreement in Bessemer, Alabama required city officials to destroy records.
Communities are routinely learning how large a project is only after the land is optioned and the incentives are signed. That sequencing is converting a zoning question into a political one. “This is Appalachia. We have a very long memory here,” Rachel Wilson told a town hall in Boyd County, Kentucky, in June. “Most of that memory is boom and then bust, so you want us to get excited about a boom. We’re all concerned about how much that boom is going to cost us, and when it is going to bust, and who’s going to pay for that bust.”
One popular objection, though, doesn’t hold up. There is no good statistical evidence that proximity to a data center depresses home values—and neither of the two existing studies settles much. A George Mason University analysis of Northern Virginia sales found no effect, though its author notes an average result is compatible with individual homes getting clobbered; an Indiana study showing near-facility homes appreciating was commissioned by a developer with a live application pending, and three of its four counties underperformed. The costs here—noise, sightlines, traffic, and above all trust—are real. On the evidence, they aren’t equity. And the campuses now proposed are far larger than anything either study measured.
Datacenter Defense
Of course, the datacenter industry has its own data to cite.. Start with rates, where causation is most acrimoniously contested. Analysts at E3 examined PJM’s 2025/26 capacity auctio. They attributed roughly half the price increase to load growth of every kind, the remainder to changes in market parameters, supply constraints and the assumed cost of new generation. That is a materially smaller share than the market monitor’s 63%, and it is the more carefully constructed piece of work. It is also, as the report itself discloses, a whitepaper funded by the Data Center Coalition and reviewed by the coalition before publication.
Virginia’s legislative auditors likewise found no evidence that data centers had historically shifted costs onto residential customers—while projecting that a typical Dominion residential customer could pay $14 to $37 more per month in real terms by 2040, and warning of cost shifts “that may be inequitable” without rate reform. E3 was the contractor on that analysis too. The two studies that agree with each other are from one firm, cited twice.
The structural response has moved faster than the argument about it. At least 38 large-load tariffs, which oblige data centers to cover the cost of serving them, were established between 2018 and 2026, thirty of them in the last two years. Two Southern Company utilities have agreed to rate freezes the utility industry credits partly to data center growth.
“In Southern Company territory, this growth is helping us freeze rates in many jurisdictions for the next few years,” said Chris Womack, Southern’s chairman and chief executive, at an Edison Electric Institute event in April. The freezes cover base rates only; fuel and storm recovery come through the door as usual.
On water, the argument is that everybody has the scale wrong. Microsoft has designed a data center that eliminates evaporative cooling and says the design will avoid more than 125 million liters per facility each year. The first two such sites, in Phoenix and Mount Pleasant, Wisconsin, don’t come online until late 2027; until then, every existing Microsoft data center keeps evaporating as usual.
Satya Nadella told attendees at Microsoft’s Build conference in June that in those closed-loop designs a facility’s “daily water usage over the course of an entire year is roughly equivalent to what a single restaurant would use.” That is true of on-site cooling—and it quietly excludes the water embedded in the electricity those campuses will draw.
The efficiency argument is the weakest of the lot. Between 2010 and 2018, data center electricity consumption rose 6% while computing output rose 550%, a statistic the Data Center Coalition still keeps in circulation. The trend has since broken: Berkeley Lab finds consumption more than doubled between 2017 and 2023 on the back of AI servers, reaching 4.4% of U.S. electricity and headed for 6.7% to 12% by 2028. Citing 2018 efficiency figures in defense of a 2026 AI buildout is an exercise in nostalgia.
What survives all of it is the concession, and it is a real one. “Especially when tech companies are so profitable,” Microsoft vice chair and president Brad Smith wrote in January, “we believe that it’s both unfair and politically unrealistic for our industry to ask the public to shoulder added electricity costs for AI.”
Microsoft’s commitment, in his words: “We’ll pay our way to ensure our datacenters don’t increase your electricity prices.” That is the most creditable thing anyone in this industry has said. It also concedes, in the politest available language, that until recently they weren’t paying.
The Reckoning
None of this will stop the buildout, and nobody building it thinks otherwise. “There’s just nowhere near enough compute for all the demand,” Mark Zuckerberg told analysts in July. Alphabet’s Sundar Pichai, a week earlier: “We continue to be supply constrained.” Whatever else is wrong here, nobody is short of customers. They are short on showing revenue.
What’s contested is how much of the announced pipeline is real. Sightline Climate, tracking 190 gigawatts across 777 projects, found that of the 16 gigawatts slated for delivery in 202,6 only about five were physically under construction, and judged 30% to 50% of the pipeline unlikely to come online this year. SemiAnalysis called Sightline’s denominator “hugely flawed, off by multiples,” noting that its own North American hyperscaler forecast had shifted by roughly 1% over six months. Both firms sell forecasts for a living. The disagreement is itself the finding: outside the developers’ own offices, nobody reliably knows which announcements are projects and which are press releases with a site plan attached.
So what’s changing is who absorbs the costs. Large-load tariffs, repealed abatements, groundwater permitting, local moratoriums, FERC’s proposed interconnection reform. Every mechanism now in motion is doing the same thing, moving costs off the public ledger and onto the project’s. In the main, that is the right outcome. It is also a repricing, and it arrives just as the projects are becoming harder to finance. And it won’t stop at the developer’s books: costs that land on the project get passed to the customers buying the compute.
The vendor side already shows the strain. Debt has gone from 9% of hyperscaler capital spending in FY2024 to 32% by midyear. Alphabet raised $84.75 billion in equity in June. Oracle closed its financial year with negative free cash flow of $23.7 billion and was downgraded in July to one notch above junk, on concentration risk from a single customer. Memory prices roughly doubled in the first quarter.
And the accounting points in two directions at once: Microsoft extended the assumed life of its data center shells from 15 to 25 years in the same quarter its chief financial officer noted that roughly two-thirds of capital expenditure went to short-lived chips. “Accounting statements may reflect orderly depreciation,” Goldman Sachs analysts wrote in May, “but operational obsolescence can impose a very different economic reality.”
A typical campus underwritten in 2024 assumed cheap interconnection, a twenty-year exemption and untested water consumption—and assumed all three would hold. None of them will.
The hidden costs of the data center boom were never really hidden. They were assigned to people who were not in the room, and who have now, county by county and statehouse by statehouse, let themselves in. The question is not whether the buildout continues. It is what these things are worth once the bill finally lands where it belongs.
