Your Water, Your Power, Your Money

The 2030 Date Was Set by Policy, Not by Steel

Almost every promise the tech industry has made about how it will power its AI build-out converges on the same year. 2030. That's the year Microsoft says Three Mile Island will be feeding its data centers. The year Google says its first small modular reactor will come online. The year corporate net-zero commitments come due. The year state energy roadmaps finish their work. The year federal grid-reliability planning horizons end.

The convergence is not an accident. 2030 is the date by which everyone — the utilities, the hyperscalers, the regulators, the politicians who voted for the tax breaks — has agreed to be measured. It is, by design, far enough away to let the build-out continue uninterrupted, and close enough to sound like a serious commitment. The framework I publish reads it directly: 2030 is a political-rhetorical anchor, not a construction schedule.

The single cleanest piece of evidence: Microsoft pledged in 2020 to become carbon negative by 2030. In its FY2024 sustainability report, Microsoft disclosed that its emissions have risen 23.4 percent above the 2020 baseline. The institution holding the most-cited 2030 commitment has, by its own published number, moved further from it every year since the commitment was made. Microsoft itself names the cause: the construction of more data centers, the embodied carbon in building materials, and the hardware required for AI workloads.

One Term You Need

Ghost Load™ is the framework's name for the gap between what an institution collects from you and what actually reaches the service you thought you were paying for. The formula is G = L − N: Ghost Load equals Total Load minus Necessary Load. Necessary Load is what it actually costs to deliver the thing. Ghost Load is everything above that line: administrative friction, financial extraction, capacity sold but not delivered, costs reassigned to people who didn't cause them. The data center build-out is, in measurable, documented terms, the largest single Ghost Load event in modern American grid history. The rest of this essay is the receipts.

The Three Showpiece Projects

1. Three Mile Island — One Reactor, Maybe by 2027

Announced September 20, 2024. $1.6 billion Constellation Energy commitment. $1 billion DOE loan. 20-year Microsoft Power Purchase Agreement. 835 megawatts. Original target: 2028. Accelerated in June 2025 to a 2027 target after PJM granted early interconnection approval.

2. Google + Kairos Power — Demo Reactors, Commercial by 2030

Google signed a Master Plant Development Agreement with Kairos Power in October 2024, contemplating six to seven Small Modular Reactors totaling 500 megawatts, with the first commercial deployment targeted for 2030 and the full fleet targeted for 2035. The MWt/MWe distinction matters: a 35 MWt thermal demonstration reactor would deliver only about 12 to 15 megawatts of electricity to the grid.

3. Stargate — Operational, But Burning Gas

Stargate is the joint venture between OpenAI, SoftBank, Oracle, and Abu Dhabi's MGX. $500 billion over four years. 10 gigawatts of compute capacity. The Abilene, Texas flagship: two buildings operational since September 2025, approximately 0.3 to 0.6 gigawatts online. Five additional U.S. sites, none operational. Total planned U.S. capacity: ~7 GW.

The decision that quietly undoes the "clean energy" narrative: at least three of the seven U.S. Stargate sites are using on-site natural gas plants. The reason is operational — the grid interconnection queue is too slow. The 2030 "clean energy" target is being met, on the ground, by burning fossil fuel directly behind the meter and reporting it as behind-the-meter generation rather than data center fossil emissions.

The Water-to-Electricity Trade-Off

At least six of the seven Stargate sites are publicly committed to closed-loop liquid cooling, which avoids the evaporative water draw that triggered most documented community conflicts. It is the only documented concession public pressure has won. Closed-loop cooling reduces water consumption but requires massive mechanical chillers, which directly spikes electricity demand. The extraction did not disappear. It changed its physical state.

Jevons Paradox

Hyperscalers argue that next-generation GPUs are more "energy efficient" per calculation. This is true at the chip level. Rack density has jumped from approximately 15 kilowatts per rack to over 100 kilowatts per rack in the AI-optimized generation. The efficiency gain is entirely absorbed into operator capability. The physical strain on the host community's grid increases exponentially.

What This Adds Up To

If every Stargate site delivers at full capacity by 2029, if Three Mile Island restarts on time, and if all announced nuclear deals deliver on schedule, the total dedicated low-carbon nuclear capacity coming online for U.S. data centers by 2030 is on the order of 1.5 to 2 gigawatts. The forecast new data center demand for that same window: 70 to 90 gigawatts in PJM and ERCOT combined. The promised nuclear substrate covers approximately 2 to 3 percent of the demand it has been marketed to cover. The 2030 narrative requires the substitution gap never to be summed.

The Grid Is Already Breaking

On May 4, 2026, the federal corporation that certifies the reliability of the North American electric grid issued an alert at the highest severity it has available. It was the second one in twelve months. Before that, the same level had been effectively dormant for more than a decade. The trigger was not weather. It was not a generator failure. It was data centers.

What a Level 3 Alert Actually Means

The North American Electric Reliability Corporation (NERC) was given mandatory enforcement authority by Congress in 2005, after the 2003 Northeast Blackout took out fifty million people because grid operators couldn't see what their own system was doing. Alert levels: Level 1 is informational. Level 2 is recommendation. Level 3 — "Essential Actions" — is mandatory, with deadlines and required reporting back to NERC. It is the highest severity NERC can issue without invoking a formal emergency order.

The May 4, 2026 alert was triggered by a documented pattern in which more than 1,000 megawatts of data center load dropped off the grid in seconds — repeatedly, across multiple regions, since 2022. The event NERC cited most prominently: the July 10, 2024 Eastern Interconnection fault that caused 1,500 megawatts of data center load to drop simultaneously. Fifteen hundred megawatts is the output of one and a half large nuclear reactors. Vanishing from the demand side. In under a second.

The Seven Essential Actions, Translated

Why the Timing Matters

The Computational Load Entity registration criteria opened for public comment on May 15, 2026 — the same date the NERC Inverter-Based Resource registration deadline took effect. The data center load class is being brought under federal-mandate reliability oversight as a new registered category for the first time in the history of the grid — three weeks before this essay was written. The institutional architecture is now formally acknowledging that data centers must be governed as grid actors, and that until now, they have not been. Simultaneously, NERC launched Project 2026-02, a mandatory reliability standards development project; the August 3 compliance responses will directly inform its scope and stringency.

The Oscillation Events at Named Frequencies

14.7 Hz oscillation, Summer 2024, Virginia. A two-hour periodic voltage sag at a cluster of data centers in northern Virginia, traced to the input stage of the data centers' uninterruptible power supply systems. Sub-synchronous resonance has a property engineers call self-excitation with negative damping: once it starts, it can grow on its own until something mechanical breaks.

July 10, 2024, Eastern Interconnection. A fault on a 340 kilovolt transmission line triggered six voltage violation events over 82 seconds. Voltage dropped to 24 to 40 percent of normal, in pulses 42 to 66 milliseconds long. Data centers transferred to backup power simultaneously. The resulting demand loss caused grid frequency to spike to 60.047 hertz and voltage to peak at 107 percent of normal. Operators had to manually pull capacitor banks out of service. This is the 1,500-megawatt loss event NERC cited.

The Power-Quality Signature in Your House

Whisker Labs and Bloomberg's joint investigation covering 770,000 home power-quality sensors found: more than three-quarters of sustained-high-distortion locations in the United States are within fifty miles of a large data center cluster. More than half are within twenty miles. Approximately 3.7 million Americans live in the most-affected areas. Wood Mackenzie's published assessment: "Near-instantaneous swings in AI power demand can damage reciprocating engines and gas turbines, while batteries may not respond fast enough to every spike."

Bad harmonics can damage residential appliances. Unaddressed, they can trigger electrical fires. It is Ghost Load™ in the most literal possible sense: the harm is being absorbed by every device, transformer, and rotating machine on the same circuit, unpriced.

Who's Actually Paying

The story being told: data centers pay their own way. The verified record shows the cost is being absorbed at three distinct layers. The household pays at all three.

Layer One: They Charge It to Your Bill

The 2027/2028 auction crossed a line no PJM auction had previously crossed. The cleared supply was 6,623 megawatts short of PJM's reliability requirement. The implied reserve margin fell to 14.8 percent against PJM's 20 percent target — the first time the entire PJM footprint failed PJM's one-event-in-ten-years reliability standard. The federal-state reliability line has been crossed, not merely approached.

A parallel cost-shift mechanism is under active FERC litigation. When a hyperscaler buys power directly from a nuclear plant behind the meter — as Microsoft did with Three Mile Island — the firm baseload capacity is removed from the public grid without new generation being built. Utilities replace the lost clean baseload with newly built natural gas peaker plants, funded through the rate base. The data center claims it is 100 percent nuclear powered. The household pays for the gas that fills the gap.

What the independent watchdogs have documented:

• Union of Concerned Scientists: $4.4 billion in 2024 PJM transmission costs assigned to residential and commercial customers in seven mid-Atlantic states, for projects principally serving data centers
• NRDC: $100 billion to $163 billion in cumulative PJM ratepayer costs projected through 2033
• PowerLines (April 2026): Residential ratepayers may bear approximately $700 billion of the $1.4 trillion U.S. utility AI infrastructure plan
• Harvard Electricity Law Initiative: Utilities may "profit from new data centers by making major investments and then shifting costs to their captive ratepayers"

Layer Two: They Don't Pay Taxes

14 additional states offer data center incentives without publicly reporting the cost. Good Jobs First conclusion: "Every state that has studied its return on investment for datacenter subsidies has found a sharply negative result." Total annual U.S. data center tax forfeit: conservatively $5 to $7 billion per year as of fiscal 2026, rising.

Layer Three: The Bill That Hasn't Arrived Yet

The four largest hyperscalers have announced combined 2026 capital expenditure of $610 to $725 billion, approximately 75 percent targeted at AI infrastructure. Hyperscalers raised $108 billion in debt in 2025. Cumulative AI-infrastructure debt issuance is projected at approximately $1.5 trillion. Capital intensity at the four hyperscalers has reached 45 to 57 percent of revenue — historically unprecedented for software companies.

The first quasi-public downside has already appeared: $1 billion DOE loan to Constellation, $303 million DOE grant to Kairos, $1 billion Defense Production Act Title III determinations in April 2026. The federal balance sheet is being positioned, quietly, to backstop a private commitment whose size exceeds any sector in the modern U.S. economy except residential mortgages.

What Happened to the Cities That Pushed Back

The clearest mistake an audit like this can make is to suggest that municipal denial of a data center solves the cost problem for the host community. The verified record shows the opposite. The municipalities that have successfully blocked a project still face cost increases driven by the same broader build-out — even when the local project never moved a shovel.

Tucson, Arizona — After Project Blue

The Tucson City Council unanimously rejected Project Blue's annexation on August 6, 2025. Tucson Water revoked the construction meter on April 24, 2026. The project is functionally foreclosed in its original form. The verified water-rate record after the denial: four consecutive annual 3.5 percent increases through FY 2029-2030, following a prior four-year cycle of 5.5 percent annual increases. Cumulative household water bill increase from 2023 to 2030 will exceed 35 percent — driven principally by drought-period Colorado River cuts and aging infrastructure. The reward for the public organizing is approximately the same rate increase it would have faced anyway.

Lake Tahoe — The Cost of Not Pushing Back Early Enough

NV Energy, which has supplied 75 percent of Liberty Utilities' electricity to the California-side Lake Tahoe community for decades, will end the full-requirements supply by December 31, 2027. The reason on the record: data center demand at the Tahoe-Reno Industrial Center — hosting Google, Apple, and Microsoft builds — and northern Nevada transmission constraints. Liberty proposed a 22.6 percent rate increase for summer 2026 residential bills. 49,000 ratepayers will now compete in the Western wholesale electricity market against the data center clusters that triggered the NV Energy exit.

Fayetteville, Georgia — After the Drainage

Fayetteville passed an ordinance banning new data center development after residents discovered that QTS had drawn 29 million gallons of water through two unmetered industrial connections during a severe drought. The retroactive bill: $147,474, paid in full. No fine imposed. The QTS campus — Project Excalibur — was not foreclosed. 615 acres, approximately 6.2 million square feet across 13 of 16 planned buildings. Construction continues through 2028-2030. The ban applies to new data centers. The campus that drained the water continues to build.

The Dalles, Oregon — The Disclosure Fight

Google's data centers in The Dalles consumed an estimated 355 million gallons of water in 2021 — roughly a quarter of the city's total water usage. When the local newspaper sought disclosure through Oregon's public records law, Google funded the city's litigation against the newspaper. The newspaper ultimately prevailed. The host city was funded by the customer to defend the customer's nondisclosure against the host city's own newspaper.

The Asymmetric Cost of Pushing Back

The institution facing pushback faces, at most, a delayed or relocated project. The household facing pushback faces, at minimum, the same rate increase it would have faced without pushback, often a larger one. The institution's cost is internalized into the price of the next project at the next site. The household's cost is internalized into the household.

Why We Cannot Rely on the 2030 Plan

The 2030 commitments depend on the same institutional architecture that has, repeatedly, in the public record, failed to deliver on commitments it made to the population on questions of life and death.

• Yucca Mountain. Designated as the permanent nuclear waste repository in 1987. Forty years later, still not operational. The same reactors now being expanded for data center loads continue to stockpile fuel in cooling pools with nowhere to go.
• Texas Grid Weatherization. Recommended after Winter Storm Sweetheart in 1989. Recommended again after the 2011 cold snap. The 2011 report explicitly warned the same failures would happen again. They did. Winter Storm Uri, February 2021. 246 dead officially, probably 700-plus in reality.
• Strategic National Stockpile. Promised since 1999 to maintain PPE and respirators for a pandemic. Casualties measured in hundreds of thousands.
• Flint, Michigan. EPA and Michigan DEQ both knew the water was poisoned in 2014. Lead pipe replacement not completed until 2024.
• Hurricane Katrina. Army Corps levees rated for Category 3. Failed at Category 3. 1,392 dead.
• PG&E. Criminally convicted on 84 counts of involuntary manslaughter for the 2018 Camp Fire. Bankruptcy. Emerged. Continued causing fires. Then December 11, 2025: the Hayward gas explosion in which the utility told fire crews "assistance not needed" ninety minutes before a home detonated.
• Ohio HB6. A $60 million bribery scheme to pass a $1.3 billion ratepayer-funded bailout for failing nuclear plants. Speaker Householder: 20 years, federal prison.
• DOE Loan Program. Solyndra. Fisker. Abound Solar. The same loan program now backstopping Three Mile Island and Kairos.

What the Alternative Costs and Builds

The full substrate deployment is approximately 3 percent of the cost of a single year's PJM ratepayer capacity recovery — and it is deployable in weeks, not in the 2028-to-2031-and-beyond window every other line item resolves to.

What Still Has to Be Built

The number is not small. It is approximately the same scale as what is already being committed on the current path. The dollars exist. The dollars are being spent. The dollars are being spent on a path that the published record has documented will fail to deliver against what it has promised. The alternative path is the same capital, redirected. The dollars do not change. The outcome does.

The Jobs the Current Path Won't Create

Conservative estimates of the workforce required to deliver a full physical grid rebuild on a fifteen-year horizon: 500,000 to 1 million new jobs across the trades, engineering, and manufacturing categories. Linemen currently face a national shortage of approximately 20,000 to 30,000 workers. Reshoring domestic large-transformer manufacturing — currently almost entirely absent from the United States — would create 10,000 to 20,000 manufacturing jobs not subject to displacement by AI. The universal feature of these jobs: they pay well, they are unionized in much of the country, and they come with health benefits, pensions, and a clear career path that does not require a four-year college degree. The current path is failing to convert the demand into the jobs because the dollars are being directed at the wrong layer.

Why This Matters More Than the Headlines

A nuclear strike, even from a state actor with multiple warheads, is geographically bounded. Devastating where it lands, manageable elsewhere. The institutional response architecture exists. Grid collapse is unbounded.

Within 72 hours of a sustained grid-down event across a major interconnection, the published modeling — FEMA, the EMP Commission, DOE resilience studies — converges on the same trajectory: water stops flowing, wastewater backs up, hospitals run generators for 72 hours then triage to who lives, refrigerated medication spoils, insulin-dependent diabetics die in the first week, dialysis patients die in the first ten days, grocery stores empty in 24 hours and do not restock, gas stations cannot pump, ATMs and payment systems fail, cell towers run on batteries for 4-8 hours then go dark. Within two weeks, casualty estimates from a national-scale grid-down event range from tens of millions to over two hundred million.

The grid is the only piece of infrastructure with no substitute. You can drink bottled water for a week. Every other modern system has a manual fallback. The grid does not. Everything modern runs on it, including the manual fallbacks for the other systems. Hospitals have generators, but generators need fuel that needs to be pumped from tanks that need electricity. The grid is the substrate of every other substrate. When it goes, nothing else holds.

A nuclear strike kills a city. A grid collapse unmakes a civilization. The first is a tragedy with a recovery path. The second is the recovery path itself failing. This is why substrate-first sequencing is not academic. It is the load-bearing question.

The Inspection Before the Road Trip

The grid the United States is currently operating has 1950s bones. Approximately 70 percent of transmission lines are over 25 years old. Large power transformers are mostly 1960s and 1970s vintage. The control systems were designed for a system that did not have to absorb 1,500-megawatt computational load drops in single seconds, and which did not produce 14.7-hertz oscillations and sub-second voltage violations at the customer end of the wire.

Nobody would drive a 1960s-architecture vehicle across the country without checking every gasket, every line, every fitting. Nobody would stake their family's safety on the assumption that a vehicle built around those vintages could carry a Ferrari engine bolted on this year, without first rebuilding the chassis to accommodate it. The institutional architecture has nevertheless decided it is acceptable to run the brain and heartbeat of the modern world on exactly that condition of vehicle.

The substrate correction is the inspection. The physical rebuild of the grid is the road trip. The sequence is not negotiable. You do not skip the inspection because the rebuild is coming. You do the inspection because the rebuild is coming, and the inspection tells you whether the rebuilt car is going to make it across the country.

There is no version of 2030 in which the legacy rails deliver the promise they have been marketed to deliver without the substrate first being corrected. The federal record, the state record, the utility filings, the academic engineering literature, the household water and electricity bills, and the cumulative institutional track record on every comparable commitment the architecture has made to the population in the last forty years all converge on the same conclusion.

That replacement is Ghost Load™ at civilizational scale. The Parallel Economy™ is the structural answer. The certified provider is the unit of rebuilding. The household that finds that provider, pays a fair price for honest work, and receives the service it paid for is the operational mesh. The lease is optional. The framework only insists the option be visible.