The Data Center Wars: Inside the Buildout Reshaping America's Land, Power, and Politics

§Understanding the Data Center Story

A data center used to be an unremarkable piece of infrastructure — a windowless building full of servers, invisible to almost everyone except the people who worked in it. That is no longer true. In the past two years, the AI boom has turned data centers into some of the largest, most expensive structures ever built in the United States, landing in places that had never previously hosted anything larger than a warehouse — and turning ordinary questions like how much water a building uses, or how a local government sets a tax rate, into national news.

This report is a primer, in two parts. The first explains what a modern AI-era data center actually is and how it works, from first principles: why a single server rack today can draw roughly thirteen times the power it did five years ago, who actually builds and owns these buildings, how they get financed, and why the industry is racing into rural counties instead of expanding the buildings it already has. The second part uses that grounding to explain the wave of local opposition that has followed — the real, documented reasons behind it, the misinformation running in both directions, and what happens next — because understanding why communities are pushing back requires first understanding what, exactly, they're pushing back against.

Three current data centers carry the technical throughline of Part One, because each makes an abstraction concrete: an unprecedented AI campus in rural Ohio, a converted bitcoin mine in West Texas, and a supercomputer complex outside Memphis now facing a federal lawsuit. A fourth project carries Part Two instead — Terafab, a chip fabrication plant in Grimes County, Texas that is not a data center at all, but was met within a day of its announcement by an 894-signature petition demanding protections its neighbors had no other way to extract. Together, the four show that the honest answer to almost every question this subject raises is not about data centers in general — it is about this one, in this county, built this way.

Once the rack has to be plumbed, the building changes, and once the building changes the economics change with it. Old halls cannot be retrofitted cheaply, so the industry builds new ones. New ones need power in gigawatt blocks, and power in gigawatt blocks does not exist on the shelf. So the buildings chase the power: to Pike County, Ohio, where the federal government spent fifty years enriching uranium and left behind a substation and a workforce; to Dickens County, Texas, where a failed bitcoin mine had already bought the interconnection — the legal right to draw a stated quantity of power from the grid, at a stated point, on a stated date; to Grimes County, Texas, where a decommissioned coal plant left a reservoir behind. And when a gigawatt of load lands next to four thousand people, the four thousand people notice.

That is the whole subject in one paragraph, and it is also why the public debate is so bad. Almost every claim being shouted across a county line — they are draining our water, they are raising our bills, they bring no jobs, they bring their own power so it costs you nothing — is true of some facilities and false of others, and the two sides are each generalising from the half of the evidence that suits them. Industry "mythbuster" pages and environmental-group "greenwashing" briefs describe the same cooling equipment and reach opposite conclusions. The honest answer is almost always facility-specific, and this brief tries to give it that way: not do data centers use water, but does this one, from what supply, and how would you know.

None of the three has behaved the way the public argument says it should. The biggest and most gas-intensive project has drawn the least organised opposition. The most community-friendly project sits in a state that has just frozen every new grid connection. And the project that generated a federal lawsuit did so not over water or rates but over air permits nobody thought to check.

§Executive takeaways

1. 1Rack power density, not AI demand in the abstract, is the mechanism driving everything else. A GB300 NVL72 rack consumes 135 kW of design power against a typical rack's ~10 kW, and rejects about 90% of its heat to liquid; air cooling simply fails there, which is why old buildings cannot be reused and new gigawatt sites must be found. Known Known
2. 2The chip vendor's guarantee in Ohio is a residual-value guarantee — a promise to cover a shortfall in what the asset is worth second-hand, not to pay the rent — with full recourse back to the tenant — which is worth little in the event it is designed to cover. NVIDIA's obligation is capped at $105bn and OpenAI has agreed to indemnify NVIDIA for every dollar paid; the trigger is OpenAI's insolvency or non-payment. Known Unknown on what OpenAI's indemnity is actually worth.
3. 3The duration mismatch runs the length of the stack — chips, contracts, leases, buildings — not just at the GPU. Microsoft depreciates servers over two to six years; CoreWeave's entire $103.7bn contracted book runs out by month 78; the leases underneath run 12 to 30 years and the buildings up to 39. Known Known
4. 4Residential rate impact is documented, quantified and regionally specific — it is not an advocacy talking point. PJM's independent market monitor attributes $6.3bn of the most recent capacity auction's $16.4bn in charges — a capacity auction being where the grid operator buys promises that generation will be available years ahead, and bills the cost onward to customers —, and $29.4bn of $63.6bn across four auctions, to data-center load. Estimated
5. 5"Closed-loop" describes the primary loop (the sealed circuit that carries heat off the chips) and tells you almost nothing about water use; the heat-rejection method on the secondary side — how that heat is finally dumped outdoors — is what decides it. Microsoft's disclosed fleet WUE — water usage effectiveness, the litres of water consumed per kilowatt-hour of computing — ranges from 0.03 L/kWh in EMEA to 0.34 in the Americas — a 11× spread inside one company's own numbers. Known Known
6. 6Opposition wins on procedure, not on merits. Virginia's largest data-center rezoning — over 20 million square feet, enough for as many as 37 data centers — was voided ab initio — treated as never having been passed at all — because a required newspaper advertisement never ran; QTS abandoned the appeal rather than start over. Known Known
7. 7The fiscal winner and the fiscal loser are different levels of government, which is why state and county politics diverge. Virginia's sales-tax exemption cost the Commonwealth $3.23bn across FY2024–25 against 2,807 net new direct jobs — about $1.15m of forgone state tax per job — while Loudoun County alone budgets roughly $1.3bn a year of data-center property tax. Known Known

1What the building is, and what AI changed

Walk into a large data hall and the first thing you notice is that it is mostly empty air and noise. Along the floor stand rows of steel cabinets — racks — each about the size of a tall refrigerator, each holding some number of servers stacked like pizza boxes. Overhead or underfoot run two things: fat copper or aluminium busbarsA solid metal bar that carries current instead of a cable. At data-center currents a cable would be impractically thick, so power is distributed along rigid bars that racks tap into — which is why a rack's capacity is quoted in amps of busbar as often as in kilowatts. carrying electricity, and fibre-optic cable carrying data. At the perimeter sit the two systems that make the rest possible: the electrical plant, which takes utility voltage down to something a server can eat and keeps it alive when the grid blinks, and the mechanical plant, which takes the heat the servers make and puts it outside.

Everything a server does converts electricity into heat. Essentially all of it — the useful work is a rounding error thermodynamically. So a data center's real specification is not floor area but power, quoted in megawatts of "IT load" (the electricity delivered to computing equipment, excluding what cooling and losses consume). This is why the industry talks in watts, not square feet, and why a "10-gigawatt campus" is a meaningful phrase and "a million square feet" is not.

For most of the cloud era the number that mattered — power per rack — barely moved. A conventional enterprise rack of web servers, databases and storage drew something on the order of ten kilowatts, and a raised floor with computer-room air handlers could cool it. Racks were spaced out. Hot aisles faced cold aisles. It was an air-conditioning problem, and as late as mid-2026 the Uptime Institute was still describing the typical rack as only then shifting toward that 10 kW mark.

Then training runs for large language models made a different demand: not many independent servers, but one very large machine. NVIDIA's NVL72 architecture wires 72 Blackwell-generation accelerators together with a copper backplane so they behave as a single accelerator, because the interconnectThe wiring between chips, and the speed at which they can pass data to each other. Training a large model means thousands of chips constantly sharing intermediate results, so the slowest link, not the fastest chip, sets the pace. Wiring 72 accelerators together tightly enough that they behave as one is an interconnect achievement, not a chip one. between chips is what limits the size of model you can train. Packing 72 of those into one standard 19-inch cabinet puts 135 kilowatts of design power — up to 155 at peak — into a footprint of about four square metres, fed by a single 1,400-amp busbar and six 33 kW power shelves. The CPUs, GPUs and NVSwitchNVIDIA's switching chip — the part that lets every accelerator in a cabinet talk to every other one at full speed, rather than through a slower general-purpose network. It is what turns 72 separate chips into something the software can treat as a single very large one. components are liquid-cooled through cold plates and a coolant distribution unitThe pumping and heat-exchange skid that serves a row or a rack: it circulates clean coolant through the cold plates on the chips and hands the collected heat over to the building's water system, keeping the two circuits separate. Every liquid-cooled hall needs one; they are the plumbing an air-cooled building was never built for.; only the optics, drives and power distribution board still take air, and the rack-level heat split is roughly 90% liquid to 10% air (Lenovo Press, GB300 NVL72 product guide).

At that density, air stops working — not as a matter of efficiency but of physics. To carry that heat away you would need a gale through the cabinet. So the coolant becomes water. In direct-to-chip cooling, small metal cold plates sit directly on the processors and a liquid loop carries the heat out of the rack; the memory, drives and power supplies still need some air, so most real deployments are hybrids. In immersion cooling, the whole server is submerged in a non-conductive fluid, which captures heat uniformly but makes maintenance awkward and, per the Uptime Institute, performs best below roughly 700–800 watts per chip — a ceiling the newest accelerators have already passed (Lake Powell Chronicle, Nov 2025).

Liquid cooling is not an upgrade that can be bolted onto an existing hall. It requires different floor loading, different pipework, different redundancy design, different service procedures and a different plant outside. Uptime puts the capital premium for supporting direct liquid cooling in a new build at roughly 5–10% — modest — but that number applies to new buildings. The installed base of air-cooled enterprise space, the tens of gigawatts already built around a 10-kilowatt rack, is largely not convertible at a sensible price. So the capital goes into new buildings, on new land, needing new power. The community fight is downstream of a thermodynamics problem.

2Why now, and why this fast

Two different computing jobs sit behind the demand, and conflating them is the most common error in the public conversation. Training is the one-off, enormously expensive process of building a model: tens of thousands of accelerators run flat out for weeks, tightly coupled, in one place, because every chip must exchange gradientsThe corrections a model computes during training — for each parameter, which way and how far to nudge it. Every chip works on a slice of the data and produces its own gradients, and they must all be pooled before the next step begins. That pooling, repeated millions of times, is why training wants one tightly wired building rather than many loosely connected ones. with every other chip. Inference is the ongoing job of answering a query with an already-built model: far cheaper per unit, far more numerous, and — crucially — latency-sensitiveSensitive to delay. A user waiting on an answer notices the round trip, and distance costs milliseconds, so inference wants to sit near the people using it. Training does not care: nobody is waiting on a job that runs for six weeks. and geographically distributable.

Training built the first wave of AI campuses and set the architecture: one huge coupled cluster, one huge power draw, one site. Inference is what makes the buildout keep going, because inference volume scales with users and with how much thinking each answer does. A reasoning model that deliberates before answering can consume many times the compute of a single-pass response for the same question. Nobody knows how fast AI usage will grow. What can be seen is that the compute cost per useful answer has been going up even as the cost per unit of compute goes down.

The capital response has no precedent in the sector's history. Amazon spent $98.4 billion on property and equipment in the first half of 2026 alone, against $57.2 billion in the same period of 2025 (Amazon 10-Q, Q2 2026). Alphabet spent $80.6 billion in the half, versus $39.6 billion a year earlier, and said it expects to increase further (Alphabet 10-Q, Q2 2026). Meta spent $50.9 billion (Meta 10-Q, Q2 2026). Microsoft's additions to property and equipment reached $115.9 billion for the fiscal year ended 30 June 2026, against $64.6 billion the year before and $44.5 billion the year before that (Microsoft 10-K, FY2026). Microsoft's property and equipment net of depreciation rose from $205.0 billion to $313.1 billion in twelve months.

Prior cloud cycles were fast; they were not this. The difference is that a cloud region could be added in increments of tens of megawatts, absorbed by an existing utility, and filled with rented capacity from many small customers. An AI campus arrives in increments of hundreds of megawatts to gigawatts, committed years before the revenue exists, and often for a single tenant. PJM — the grid operator for thirteen states and the District of Columbia — projects peak demand growth of 32 gigawatts between 2024 and 2030, of which all but 2 gigawatts is data centers (Canary Media, Dec 2025). Lawrence Berkeley National Laboratory's 2024 assessment put US data centers at 4.4% of national electricity in 2023 and projected 6.7% to 12.0% by 2028 (LBNL, Dec 2024).

The speed has a second cause that gets less attention: the capital is not waiting for the demand because the capital is being told not to. Contracts are signed for capacity years out — CoreWeave carried $103.7 billion of remaining performance obligations at 30 June 2026 — and those contracts, not observed usage, are what get financed. A remaining performance obligation is the money a customer has committed to pay under a contract the seller has not yet delivered on: signed revenue, not earned revenue. It is the most forward-looking figure in a set of accounts and the least tested, because it records what a customer has agreed to pay — not what that customer will still be able to pay in year five. The buildout is being underwritten against a promise book, and the promise book is the thing a sceptic should interrogate first.

3Who builds them, who owns them, and why each model exists

There is no single "data center industry." There are five distinct businesses that happen to produce similar buildings, and they exist because they are solving different problems with different balance sheets. Reading a news story about a data center is much easier once you can tell which one you are looking at.

HyperscalerThe handful of firms operating computing estates at global scale — the ones that buy servers by the hundred thousand and land by the square mile. The word is a size class, not a business model: what makes a hyperscaler is that everything it does, it does at a scale where it becomes cheaper to build the thing than to rent it. self-build. Microsoft, Amazon, Google, Meta and Oracle design and operate their own facilities because at their scale the cost of a landlord's margin exceeds the cost of learning to be one. They also want control of the physical layout — Meta's newest halls are purpose-built around specific cluster topologiesThe shape of the wiring — which machines are connected to which, and how many hops a message takes to cross the cluster. Different training methods want different shapes, and the shape has to be poured into the building: cable runs, hall geometry and power blocks all follow it. It is the reason a tenant may want to design the shell rather than rent one.. What has changed is that even they no longer self-build everything. Meta disclosed $278.99 billion of operating and finance leasesA lease long enough, or complete enough, that accounting treats the tenant as the effective owner: the asset and the matching debt sit on the tenant's balance sheet. An operating lease is treated as rent. The distinction matters here because it decides whose books a data center appears on — and a company can shift the building off its own capital budget without shifting the obligation. leases that had not yet commenced as at 30 June 2026, for data centers, colocations and network infrastructure, commencing through 2036 with terms up to 30 years — and added roughly $68 billion more in July 2026 alone, on 18- to 20-year terms. That is a company with $90 billion of cash choosing to rent.

Wholesale and retail colocation. Equinix, Digital Realty and Iron Mountain build shells and sell space, power and cooling to tenants who bring their own computers. Retail colocation (Equinix's core) sells by the cabinet and monetises interconnection — the value is being in a building where everyone else already is. Wholesale (Digital Realty's core) sells by the megawatt on long leases. Digital Realty disclosed 1,402 MW under construction at an estimated pre-tax stabilised cash yieldThe annual cash a building is expected to throw off once it is finished and fully let, divided by what it cost to build. "Stabilised" means the lease-up is over and the number has settled. It is the developer's headline return on a project, before financing — and its inverse is the multiple of cash flow the developer is effectively building at. of approximately 11.5%, and over 7 gigawatts of future development capacity (Digital Realty 10-Q, Q2 2026). Equinix's weighted-average remaining lease terms sit at 12–13 years at 5–6% discount rates (Equinix 10-Q, Q2 2026). Iron Mountain, a records-storage company that grew into data centers, entered a finance lease in February 2026 with an initial term of 31 years and undiscounted minimum payments of about $223 million (Iron Mountain 10-Q, Q2 2026).

Oracle sits at the boundary of that model — hyperscaler, tenant and landlord's counterparty at once — and discloses more frankly than anyone else in the sector. Remaining performance obligations stood at $638 billion at 31 May 2026, against $138 billion a year earlier, "primarily attributable to certain significant cloud contracts that were entered into during the period" (Oracle 10-K, FY2026). A 4.6× increase in one year is what a demand curve looks like when it arrives as a handful of signatures. Microsoft's equivalent figure rose 84% to $678 billion. These are the promise books the buildings are financed against.

NeocloudsThe new generation of specialist cloud providers that do one thing: rent out AI accelerators. They carry none of a general cloud's breadth — no databases, no storage tiers, no sprawling service catalogue — and all of its capital intensity, concentrated in a single fast-wasting asset.. CoreWeave and Nebius rent accelerators by the hour or the year. They exist because a start-up that needs 5,000 GPUs for eighteen months cannot buy them and a hyperscaler will not prioritise it. Their economics are brutal in a specific way: the asset is the chip, the chip depreciates fast, and the contract has to pay for it before it does. Nebius reported Q2 2026 revenue of $582.3 million against depreciation and amortisation of $259.7 million — 45% of revenue — while spending $8.1 billion on property and equipment in the half (Nebius Q2 2026 results, 12 Aug 2026). CoreWeave's cost of capital tells the same story from another angle: a sale-and-financing arrangement on critical infrastructure carries an imputed interest rateThe interest rate hidden inside a deal that is not called a loan. Sell an asset and lease it back, or agree to pay in instalments, and accountants work backwards from the payments to the rate you are effectively being charged. It is the cleanest read on what the market thinks of a borrower's credit, because it is arithmetic rather than marketing. of 15%, and a related data-center finance lease 13%.

Power-first developers. Helios is the cleanest example of the newest of these models, and the least understood. Galaxy Digital did not set out to be a landlord; it bought a bitcoin mine. Argo Blockchain had built roughly 160 acres in Dickens County in 2022 with 180 MW of interconnection; when crypto turned, Galaxy bought the site for about $65 million in 2023 (Texas Real Estate Research Center, Apr 2026). What it was really buying was a completed interconnection study. By 2026 the campus covers over 1,500 acres and ERCOTThe Electric Reliability Council of Texas — the operator of the Texas grid, which runs almost entirely separate from the rest of the country's. It decides who may connect, in what order, and it is the body that grants or withholds the interconnection a developer is really shopping for. Its counterpart in the mid-Atlantic and Midwest is PJM. has approved more than 1.6 GW of gross capacity. Galaxy builds the shell, the power and the cooling, and leases to a tenant — CoreWeave — that brings the chips. Blackstone-owned QTS, Vantage and Switch run variants of the same trade at larger scale.

The power-first model is also where this asset class stops being industrial and turns financial. Blackstone disclosed in its Q2 2026 accounts that its Core+ real estate strategy now includes "a NYSE-listed REITA real estate investment trust: a company that owns income-producing property and, in exchange for passing nearly all its profit through to shareholders as dividends, pays little or no corporate tax on it. REITs are how ordinary savers come to own warehouses, hospitals and — now — data centers, and how a building's rent becomes a traded yield. focused on acquiring and owning data centers, Blackstone Digital Infrastructure Trust" (Blackstone 10-Q, Q2 2026). A decade ago a data center was a cost line inside a technology company. It is now a listed income vehicle sold to institutional and individual investors on a yield — which means the capital funding the buildout has a redemption profileThe pattern of when investors are entitled to ask for their money back. A fund open to redemptions holds assets that take years to sell against claims that can be made in months; when enough holders ask at once, the manager must sell into whatever market exists that quarter. The mismatch is dormant in good conditions and decisive in bad ones., and redemption profiles behave badly in exactly the conditions that would strand a campus.

Chip-vendor-financed. The Ohio project is the first at this scale, and it inverts the usual arrangement. Normally the tenant's credit supports the landlord's construction debt. Here, the tenant is a company without a credit rating, so the chip supplier stands behind the lease instead — and takes exclusivity over the silicon in return. The press coverage was not precise about what NVIDIA actually agreed to. The filing is read closely in §5.

4How a Data Center Actually Gets Built

A developer looking for a site is not looking for land. Land is abundant and cheap almost everywhere a data center might plausibly go. The developer is looking for an electrical interconnection — a legal and physical right to draw a very large amount of power from a specific point on the grid at a specific date — and everything else is negotiable around it.

This is why the sequence looks strange from the outside. Sites get chosen because a coal plant closed there (Gibbons Creek, Grimes County), or a uranium enrichment complex left a substation behind (Portsmouth, Pike County), or a bitcoin miner already paid for the studies (Helios, Dickens County). The buildings follow the wires.

1. 01
   Find the power

   An existing or retiring interconnection, a transmission corridor, or a place where new generation can be built alongside. Land selection is downstream of this.
2. 02
   Enter the queue

   Apply to the ISOIndependent system operator — the referee of a regional grid. It does not own wires or power stations; it dispatches whoever does, keeps supply and demand matched second by second, and controls the queue of everyone waiting to connect. ERCOT and PJM are the two that run through this report. or utility for a large-load study. Interconnection process duration rose about 70% on average between 2010 and 2023, and 78% of generation capacity proposed between 2000 and 2018 withdrew before coming online.
3. 03
   Secure water and discharge

   Either a municipal supply agreement, a groundwater right, a surface-water right, or a design that needs almost none — the cooling architecture is chosen here, and it is effectively irreversible.
4. 04
   Zoning, permits, incentives

   Rezoning or special-use permit at county level; air permits if there is on-site generation; and the tax negotiation — abatement, PILOT, or exemption — which is usually settled before the public hearing.
5. 05
   Build, energise, fit out

   Shell and electrical plant first, then the tenant's racks. Long-lead electrical gear and skilled electricians, not concrete, set the schedule.

Outsiders underestimate the queue. Berkeley Lab's analysis of US interconnection queues, published in Joule, found that 78% of the generation capacity proposed between 2000 and 2018 withdrew before coming online — only 14% of proposed solar and 32% of proposed gas plants were completed — while the duration of the interconnection process rose about 70% on average between 2010 and 2023 (PV Tech reporting Berkeley Lab, 14 Jan 2025). Berkeley Lab itself notes this can indicate a healthy competitive market for sites, but that completion rates which are too low become "a drain." Either way the arithmetic for a utility planner is the same: most announced capacity never happens, and a forecast that assumes otherwise builds a grid nobody needs.

Texas has just made that concrete. On 3 August 2026 Governor Greg Abbott wrote to the Public Utility Commission of Texas and ERCOT directing a "comprehensive verification and audit of all data centers advancing through ERCOT's interconnection process" before any additional project is approved, on the basis that ERCOT was holding roughly 474 gigawatts of interconnection requests — more than five times the grid's record peak — of which about 90% were data centers. Two days later ERCOT suspended the Batch Zero large-load classification notifications that had been scheduled for 7 August, and said it would seek a good-cause exception at the PUCT's 20 August open meeting (ERCOT Market Notice M-A080326-01).

The questionnaire reads like a summary of this entire report. Developers must disclose their tax incentives, grants and abatements; how much power they will use and how much they will generate themselves; how much water they will consume, from where, and what cooling technology they will use; steps taken to mitigate noise and traffic; and their ownership and controlling interests. Materially false information or non-response makes a project ineligible (Gibson Dunn, Aug 2026). ERCOT told the PUCT the exercise would cover roughly 250 to 300 projects and take several months. Galaxy Digital, whose Helios expansion depends on ERCOT approvals, disclosed the risk in its own 10-Q: "PUCT and ERCOT must complete this audit before any data center project moves forward. This review may impact the power approval timelines for certain of our data center sites."

The audit is not an environmental review, a zoning decision, or a rate caseThe formal proceeding in which a regulated utility asks a state commission for permission to charge more, and anyone with standing may contest it under oath. Bills are not set by the market; they are set here, on an evidentiary record, and the record is open to whoever bothers to file.. It is a solvency and honesty test applied by a grid operator, on the theory that the queue is full of speculative requests — the same project counted four times by four developers shopping the same site, or a request filed to establish a place in line before financing exists. Whether that theory is right is the single largest open question in the sector, and Texas is about to produce the first serious dataset on it.

5How Data Center financing actually works

Almost everything in this industry runs through three contracts, and a reader who knows them can follow any deal announcement.

The colocation lease. A landlord builds a shell and delivers a contracted quantity of power and cooling; the tenant installs its own computers and pays rent per kilowatt per month, typically on a 10- to 15-year term with escalatorsContractual clauses that step the rent up every year, by a fixed percentage or by an inflation index. They are why a lease signed today produces a rising cash flow without anyone renegotiating — and why rent climbs on a schedule that pays no attention to what the tenant's own revenue is doing.. The landlord's return depends on filling the building and on the tenant's credit. Digital Realty's disclosed 11.5% pre-tax stabilised cash yield on capacity under construction is the honest headline number for this business: build for roughly nine times stabilised cash flow, finance a chunk of it, and the spread is the business.

The GPU-cloud contract. A neocloud commits capacity to a customer for a term — usually one to six years — at a fixed rate, and books it as remaining performance obligation. CoreWeave's disclosure is unusually informative about duration: of $103.7 billion of unsatisfied RPO at 30 June 2026, 41% is expected to be recognised within 24 months, 39% between months 25 and 48, and the balance between months 49 and 78 (CoreWeave 10-Q, Q2 2026). That is a book with a hard horizon: by month 78 — six and a half years — every contract currently signed has run off.

The credit wrap. The credit wrap is the new one, and Ohio is where to learn it. The reporting called it NVIDIA "backing $105 billion in financing." The filing is more specific than that.

On 17 August 2026 NVIDIA entered multiple residual value guaranties with SB Energy relating to leases for approximately 4.25 gigawatts of IT load at the Portsmouth site, with aggregate payment obligations "cumulatively capped at $105 billion for its initial commitment." An OpenAI affiliate is the tenant on a twenty-year lease; SB Energy builds, owns and operates. NVIDIA can, at its sole discretion, provide credit support for approximately a further 3.8 GW. If OpenAI becomes insolvent or fails to pay, NVIDIA pays the gap between the guaranteed minimum lease value and what SB Energy recovers through a replacement lease or a sale, and may then assume the lease, force a re-letting, start a sale process, let the lease terminate, or defer for up to a year while covering project costs. The guarantee dies on the twentieth anniversary of each lease, on OpenAI terminating, or — the clause worth pausing on — on OpenAI achieving a satisfactory credit rating. And: "OpenAI has agreed to reimburse and indemnify NVIDIA for any and all amounts actually paid by NVIDIA to the Lessor." NVIDIA filed the whole arrangement under Item 2.03 — creation of an off-balance-sheet obligationA commitment that binds a company but does not appear among its liabilities, because it only becomes payable if something else happens first. A guarantee is the classic case: nothing is owed until the party being guaranteed fails. The obligation is real, disclosed in the notes rather than the totals, and invisible to anyone reading only the balance sheet. (NVIDIA Form 8-K, Item 1.01/2.03).

1. 01
   SB Energy builds and owns

   Develops land, power and shell on private and remediated DOE land; raises construction capital against the lease.
2. 02
   OpenAI leases, 20 years

   Tenant on approximately 8 IT-GW; deploys NVIDIA's DSX platform, subject to limited exceptions. No public credit rating.
3. 03
   NVIDIA guarantees residual value

   Capped at $105bn for the initial 4.25 GW. Effective on lease commencement and ready-for-service, expected from 2028.
4. 04
   OpenAI indemnifies NVIDIA

   Reimburses "any and all amounts actually paid" — recourse runs back to the tenant whose failure is the trigger.
5. 05
   NVIDIA takes exclusivity

   Sole AI compute infrastructure provider at the campus; also invests $1.5bn in SB Energy alongside SoftBank and OpenAI.

That loop answers the mandate's hardest question — if AI demand slows, who is actually on the hook? — better than any commentary has. The obvious reading is that NVIDIA has taken on $105 billion of risk to prop up its largest customer. That reading is not wrong. It is incomplete. NVIDIA's exposure is the shortfall after re-letting or sale, not the gross lease. If AI demand merely cools, an 8-gigawatt powered campus in 2032 is still one of the most valuable pieces of industrial real estate in North America and re-lets at some price; NVIDIA pays a gap, not a lease. The scenario in which NVIDIA pays close to the cap is one in which OpenAI fails and there is no replacement tenant for gigawatts of finished, powered shell — that is, a scenario in which demand for AI compute has collapsed generally. In precisely that scenario, NVIDIA is also the company whose own revenue has collapsed, and the indemnity from OpenAI is worthless.

The guarantee is therefore not a hedge. It is a correlated bet stacked on top of an existing correlated position — which is a reasonable thing for NVIDIA to do if it believes the demand is real, and a very expensive thing if it is wrong. And the order of loss is simple: the tenant's equity goes first, then the developer's equity and the lenders behind it, then the guarantor, and only then — via stranded generation and transmission — the ratepayer. Whether the ratepayer sits fourth or first in that queue is exactly what the tariff fights in §11 and §14 are about.

Look again at what the hyperscalers have signed. Microsoft's operating and finance lease commitments including imputed interest totalled $443.5 billion at 30 June 2026, of which $411.1 billion — 93% — falls after fiscal 2027. Meta had $278.99 billion of leases not yet commenced, plus about $68 billion signed in July 2026, plus $349.31 billion of non-cancellable contractual commitments. These are not the balance sheets of companies that will walk away lightly. Whatever happens to AI demand, an enormous quantity of 20- and 30-year rent has already been promised by counterparties with investment-grade credit — and that, more than any forecast, is what is actually financing the buildings.

6Is this normal, or unprecedented?

Three prior cycles get invoked and only one of them fits well.

Dark fibre, 1998–2001. Carriers laid enormous quantities of optical fibre against forecasts of internet traffic doubling every hundred days. Traffic grew fast, just not that fast, and much of the fibre sat unlit for a decade. The analogy is used to argue that AI capex will strand. It has one deep flaw and one deep insight. The flaw: the marginal cost of adding fibre capacity, once a conduit is dug, is close to zero — you light another strand or upgrade the optics. Overbuilding fibre was cheap and the glut therefore enormous. Adding a gigawatt of interconnected, permitted, water-served, cooled electrical capacity costs billions and takes five years. Physical scarcity puts a floor under this cycle that fibre never had. The insight, which survives: what stranded in the fibre bust was not the asset, it was the capital structure. The fibre kept working; the companies that borrowed to lay it did not survive to operate it. The right question is not "will the buildings be useful" but "will the entities that financed them still own them."

Retail and malls. Overbuilt against a demand assumption that a substitute technology then destroyed. Poor fit: nothing is currently substituting for compute. Worth keeping only as a reminder that the physical asset can be fine while its use case migrates elsewhere — an inference-heavy world may want many small facilities near users, not a few enormous ones in Appalachia.

Nineteenth-century railway overbuild. The closest fit, and the least cited. Rail was built ahead of demand, with land grants and public subsidy, by promoters whose incentives were to build rather than to operate. Multiple parallel lines were laid to the same towns. Bankruptcies were routine and yet the track remained, was consolidated cheaply, and became the backbone of the economy. The rhyme is exact: capital committed ahead of proven demand, public inducements steering siting, duplicated routes into the same "town" (see: 474 GW of ERCOT requests against 91 GW of peak demand), and a physical asset that survives its first owner.

The buildings will probably be fine. The distinctive risk of this cycle is that the public commitments made to support them — new transmission, new generation, tax abatements granted for thirty years — outlive the private demand that justified them. Georgia's regulators have thought harder about this than most; see §14.

7The obsolescence wrinkle

A GPU is superseded roughly every eighteen months to two years and is generally reckoned to have an economic life of three to five. The building around it is financed on a twenty-to-thirty-year assumption. Those two clocks are attached to the same asset, and nobody has fully reconciled them.

Depreciation is an estimate, not a measurement. A company decides how many years an asset will earn its keep and spreads the purchase price across them, so choosing six years rather than three halves the annual charge and lifts reported profit without one thing changing inside the building. That is why the length of these schedules gets argued over instead of looked up. Estimates outnumber facts on this subject, so take the issuers' own disclosures first. Microsoft depreciates "servers and network equipment" over two to six years and "buildings and improvements" over five to fifteen (Microsoft 10-K, FY2026, significant accounting policies). Digital Realty depreciates buildings over a remaining useful life of 5 to 39 years, amortises in-place lease value over a weighted-average 13-year term, and customer relationship value over 10 to 15 (Digital Realty 10-Q, Q2 2026). Iron Mountain signed a 31-year finance lease in February 2026. Meta's newest leases run 18 to 20 years, with the not-yet-commenced book running to 30. The Ohio lease is 20.

There is a hard step in that chart between the sixth year and the twelfth: everything that touches a chip is priced inside six or seven years, and everything that touches concrete is priced beyond twelve. Nobody is contracting across the gap. A landlord signing a fifteen-year lease against a tenant whose own revenue book runs out in six is taking, whether or not it is described that way, a view on the second decade of AI demand.

Short sellers accuse the hyperscalers of flattering earnings by depreciating AI hardware over five or six years when its real frontier life is two or three. The industry replies that a chip cascades: it trains frontier models for two years, serves inference for two more, and runs smaller workloads after that, so six years of economic benefit is genuine even if two years of frontier relevance is not. The dispute is usually conducted with estimates. It does not have to be.

Amazon settled the empirical question against itself. In its annual report for 2025 the company disclosed that, effective 1 January 2025, it "changed our estimate of the useful lives of a subset of our servers and networking equipment from six years to five years," and stated the reason plainly: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." The cost of that honesty was $1.4 billion of additional depreciation and $1.0 billion of net income in a single year, or $0.10 per share, falling mostly on AWS (Amazon 10-K, FY2025, Note 1). A company with every incentive to lengthen shortened instead.

If a four-year-old accelerator earns its keep only by being redeployed to cheaper work, then revenue per rack declines across the asset's life while the rent underneath it escalates on a contractual schedule. The cascade argument saves the depreciation charge and indicts the lease.

The clearest live evidence sits in Nebius's income statement. Depreciation and amortisation was $259.7 million against revenue of $582.3 million in Q2 2026 — 45 cents of every revenue dollar consumed by the wasting of the asset, in a quarter when revenue grew 454% year on year. Growth is currently outrunning the depreciation. The sector's real question is what that ratio looks like when it stops.

An issuer has now written the mismatch into its own risk factors, in language a critic could not improve on. Oracle warns that if it overestimates demand or a key customer cannot pay, "we could be locked into multi-year commitments for excess data center space and related capital expenditures, as well as associated financings, without receiving corresponding revenue" — and, more precisely: "the terms, renewal options and pricing adjustments in our long-term data center leases typically do not align with the duration and pricing of customer contracts, and if customers do not renew their contracts, we may be unable to re-lease, repurpose or assign such capacity on acceptable terms, if at all." That is the year-six-to-year-twelve gap, disclosed by a company with $638 billion of contracted revenue behind it.

Not all of this is an accounting problem. NVIDIA's own framing at Ohio is that the campus is "long-lived infrastructure for NVIDIA compute" that can be "upgraded repeatedly with each new generation." That is a real answer — the building is a socket, the chips are light bulbs. But the socket has to fit. A hall built for 135 kW liquid-cooled racks in 2028 must accommodate whatever density 2038 demands, and NVIDIA is already preparing the industry for megawatt racks and 800 VDC distribution. The wrinkle is not that buildings outlive chips. It is that buildings have to outlive three or four complete rebuilds of the electrical and cooling plant inside them, and no one has yet had to do that twice.

8Power, water, and what the clean-energy claims are worth

Two claims dominate the public argument — they bring their own power so your bills are safe and they use closed-loop cooling so they don't use water — and both are technically defensible sentences deployed to support conclusions the evidence does not reach.

Power: what "behind the meter" really covers

Start with the meter itself. "Behind the meter" means the generation sits on the customer's side of it: the electricity is made and consumed on the site, and never travels the shared grid or registers as demand on it. In front of the meter, the same building is simply a very large customer, drawing from the pool everyone else draws from and paying the rates everyone else pays. Which side of one piece of metering equipment a project sits on is what most of the argument in this section turns on.

There is a genuine and rapid shift toward dedicated generation. The reason is not principally environmental; it is that the interconnection queue is five years long and a dedicated plant is not. Ohio is the largest example on record. SB Energy and SoftBank will build "at least 10 GW of new energy generation" to yield 8 IT-GW of capacity, and invest at least $4.2 billion in regional grid infrastructure through a partnership with AEP Ohio "designed to protect ratepayers" (NVIDIA/SB Energy/OpenAI release, 17 Aug 2026). The project's own site discloses the composition: of the roughly 10 GW, 9.2 GW is natural gas, backed by about $33 billion of gas generation investment and 765 kV transmissionExtra-high-voltage power lines. The higher the voltage, the more power a line can carry and the less of it is lost as heat along the way, so bulk transfers over long distances are built at this class rather than at the far lower voltages that run down a street. It is the most expensive and most contested kind of wire to build. (PORTS-Pike Technology Campus). A separate entity, Portsmouth Gas DevCo, has an initial phase of up to 2 GW of simple- and combined-cycleTwo ways of burning gas for power. A simple-cycle turbine is a jet engine on the ground: cheap, quick to build, quick to start, and it throws its exhaust heat away. A combined-cycle plant captures that exhaust heat to raise steam and drive a second turbine, which makes it markedly more efficient but slower and dearer to build. Developers in a hurry reach for the first; developers running a plant for decades want the second. gas plant before the Ohio Power Siting Board, with construction from Q1 2027 and commercial operation from Q3 2028 (PORTS-Pike Energy Center).

Ohio will not add load to the shared grid, and it will add roughly nine gigawatts of new gas-fired combustion to the Ohio River valley. Those are both true and the second is rarely said by anyone quoting the first. "Protecting ratepayers" and "decarbonising" are different objectives, and this project chose the first.

Helios is the opposite configuration: a grid-connected load, buying from ERCOT, with over 1.6 GW of approved gross capacity — and now inside a statewide interconnection pause it did not cause. Colossus/Southaven is a third: on-site generation built first and permitted afterwards, which is what produced the litigation.

What fraction of the buildout is genuinely behind the meter? No authoritative published figure exists, and this report will not invent one. What can be said from the queue data is the opposite of the industry's implication: ERCOT was holding roughly 474 GW of interconnection requests, about 90% of them data centers, which is by definition grid-seeking load — the audit exists precisely because so much of the pipeline is not self-supplied. PJM projects 32 GW of peak growth to 2030, all but 2 GW from data centers (PJM, 2026 Load Forecast Report, 14 Jan 2026), and PJM has no equivalent of the Ohio arrangement at scale. Dedicated generation is real, growing, and a minority of the pipeline. Both the sibling grid report and the rate cases in §11 document real residential impacts from grid-tied facilities; that a handful of flagship projects bring their own power does not touch those findings.

The additionality question

Every large operator claims to match its consumption with carbon-free energy. The mechanism matters more than the claim. Additionality is the test the whole thing turns on: did the buyer's money cause clean generation to exist that would not otherwise have been built, or did it buy a share of generation that was going to be there anyway? Two contracts can read identically on a sustainability page and answer that question in opposite directions. An unbundled renewable energy certificate is a tradeable attribute severed from the electron: buying one changes an accounting line, not a grid. A power purchase agreement that provides the revenue certainty a new wind farm needed to reach financial close is additional — it caused generation that would not otherwise exist. Annual matching lets a buyer offset a January night in Ohio with a July afternoon in Texas. Hourly matching does not.

Meta's filing is more revealing than its marketing. The 10-Q says the company has "entered into agreements ranging from nine to 25 years to purchase clean and renewable energy that do not specify a fixed or minimum volume commitment," and that "the ultimate spend under these agreements may vary and will be based on actual volume purchased." A contract with no minimum volume gives a developer materially less financeable certainty than one with a floor — which is the whole additionality mechanism. That is not an accusation of greenwashing; it is a note that the disclosed contract terms are weaker than the headline suggests, and readers should distinguish the two. WattTime's research position — that hourly matching without additionality has little to no emissions effect — is the sharpest statement of the underlying point (WattTime), and 2026 modelling work from NYU and MIT finds that how an hourly standard is written can produce large differences in reported emissions with no change in actual grid decarbonisation (TechXplore, Aug 2026).

Water: the distinction almost everyone gets wrong

"Closed-loop cooling" describes the primary loop — the sealed circuit that carries heat from the chips to a heat exchanger. It says nothing at all about how that heat then leaves the building, and the secondary side is where all the water goes.

Inside the sealed primary loop, direct-to-chip cooling needs very little fluid at all — a technical brief cited by the Lake Powell Chronicle, Nov 2025 puts it at sometimes under four gallons per rack, circulating a water-glycol mix or a dielectric fluidA liquid that does not conduct electricity, so live electronics can be bathed in it without shorting. It is what makes immersion cooling possible at all, and it is considerably more expensive than water.. What happens at the fork is a blowdownWater deliberately drained out of a cooling tower and replaced with fresh. As water evaporates it leaves its dissolved salts behind, so the remaining water grows steadily saltier until it would scale or corrode the plant. Blowdown is the purge that resets it — and it is why an evaporative site both consumes water and discharges it. tax: an evaporative tower deliberately drains and replaces water to control mineral concentration as it evaporates, which is why it both consumes water and discharges it; a dry radiator does neither, at the cost of more electricity in hot climates.

Once you hold that distinction, the competing claims resolve. Oracle's own explainer is careful and, read closely, agrees: it describes deploying "a variety of cooling methods like closed-loop cooling that do not rely on continuous consumption of potable water" at upcoming sites in New Mexico, Michigan, Texas and Wisconsin — the operative words are non-evaporative, not closed-loop (Oracle, Feb 2026). Vantage is equally precise about the lifecycle: piping is flushed with water at commissioning and that flush water is disposed of as wastewater; the loop is then filled, sealed and treated with anti-corrosives and pH adjusters; "routine topping off should not be necessary" and make-up waterFresh water added to a cooling system to replace what it has lost. In an evaporative design the losses are continuous and by design, so make-up runs constantly; in a sealed loop there is nothing to replace unless something has leaked or been opened up for service. How much make-up a facility needs is therefore a plain-language test of what kind of system it really has. is "generally only needed in the event of a leak or maintenance event" (Vantage, Apr 2026).

And the trade-off is conceded by the operator with the most to lose from conceding it. Microsoft, announcing designs that avoid "more than 125 million litres of water per year per datacenter," states that the shift from evaporative to mechanical cooling will produce "a nominal increase in our annual energy usage compared to our evaporative datacenter designs across the global fleet," mitigated by warmer cooling temperatures and high-efficiency chillers. Pilots are at Phoenix and Mount Pleasant, Wisconsin in 2026, operational from late 2027 (Microsoft Cloud blog). Dry cooling is not a free efficiency gain. It is a decision to spend electricity instead of water — defensible in Phoenix, pointless in Dublin.

Microsoft's FY2025 fleet WUE was 0.27 L/kWh globally — but 0.34 in the Americas and 0.03 in Europe, the Middle East and Africa, an elevenfold spread inside one company's own portfolio, driven by climate and design rather than corporate policy (Microsoft datacenter efficiency disclosures). Its fleet PUE was 1.17. Google reported a 2025 fleet PUE of 1.09 and that 87% of its freshwater withdrawal came from sources at low or medium risk of scarcity (Google Data Centers). Aggregate volume rose anyway: Google's reported 2025 water consumption was about 10.9 billion gallons, up roughly a third year on year, with 7.7 billion gallons replenished.

Efficiency per unit improved and total consumption rose sharply, because the denominator grew faster. Any argument that cites only WUE, or only gallons, is choosing its number.

Helios is the concrete example on the good side of this ledger, with an important caveat about what is actually verified. Galaxy states that Helios uses a closed-loop water system to minimise consumption. Its SEC filings — the 10-Q covering Q2 2026 — do not discuss water, cooling architecture, or water-usage effectiveness at all. The frequently repeated figure of roughly 10 million gallons recirculated on site could not be independently confirmed against a Galaxy filing, a Texas water-rights record or a utility disclosure, and it is treated here as unverified. What is verified is the structural point: Dickens County sits in a semi-arid part of West Texas where an evaporative design would have been a live community issue, and the developer chose a design that avoids it.

9The rest of the world, and the wall the buildout will actually hit

The buildout is not only American. The foreign projects are useful mainly as a calibration on how much announced capacity is real. The UAE–US AI Campus in Abu Dhabi is a headline 5 GW; Stargate UAE, developed by G42's Khazna alongside OpenAI, Oracle, NVIDIA, Cisco and SoftBank, is a 1 GW facility within it, and its first phase of 200 MW was due for completion in the third quarter of 2026, with more than 5,000 workers and over 100,000 cubic metres of concrete on site (The National, 5 Dec 2025). Saudi Arabia's PIF-backed HUMAIN targets roughly 6 GW by 2034 with first Riyadh and Dammam facilities around 100 MW each.

Two hundred megawatts against a five-gigawatt announcement is 4%. That ratio is the most useful discipline a reader can apply to any data-center headline, foreign or domestic: announced capacity and energised capacity differ by an order of magnitude, and the press release almost always quotes the former.

The domestic count runs the same direction. Cleanview's US tracker — a live, independently maintained map of named projects, not an industry survey — puts operating US data-center capacity at 61,212 MW across 1,285 facilities against 378,579 MW of planned capacity across 1,991 more: roughly six planned megawatts for every one actually running, a gap in the same order of magnitude as the UAE calibration above and a separate corroboration of §9's central caution, from a dataset built independently of Data Center Watch or LBNL's queue figures (Cleanview, US Data Center Map, accessed Aug 2026).

The binding physical constraint on all of it may turn out to be neither chips nor power but people. In June 2026 Google.org committed $50 million to skilled construction trades training, roughly $20 million of it to the Electrical Training Alliance — the IBEW's official training partner — and the National Electrical Contractors Association, with further funding to sheet metal workers, plumbers, pipefitters and welders. The union's own account is unusually blunt: IBEW International President Kenneth W. Cooper describes "unprecedented demand for skilled electrical workers," and the article calls meeting it "one of the biggest challenges the IBEW and NECA have faced in their history" (IBEW, 17 Jun 2026).

That is an encouraging development and also an admission. When the largest technology companies in the world start funding electrical apprenticeships, it is because they have discovered that the constraint on their strategy is a trade you cannot import, cannot automate, and cannot accelerate below the years it takes to train someone. The Ohio project's 35,000 construction jobs over a six-year buildout are not a public-relations number; they are a statement of how many people the developer thinks it must find, in a region that has not run a construction programme at that scale since the enrichment plant was built.

10What is actually happening

In the first quarter of 2026, by the count of Data Center Watch, at least 75 US projects worth approximately $130 billion were blocked or delayed — roughly matching the whole of 2025 in three months. Active opposition groups more than doubled from the end of 2025, to 833 across 49 states. Over 300 state data-center bills were filed in the first six weeks of the year, statewide moratorium proposals appeared in fourteen states from both parties, and Maine came within a single House vote of enacting the first statewide ban (Data Center Watch, Q1 2026).

Two caveats belong immediately next to those numbers. Data Center Watch does not publish its methodology beyond the research window, does not define "blocked or delayed," and does not disclose its funding; the $130 billion is a sum of announced project values, which — as §9 established — routinely overstate what would have been built. Treat the direction as solid and the magnitudes as indicative. A second, independent tracker maintained by Will Manidis counts ten active bans or moratoria, eight advancing bills, sixteen jurisdictions in discussion, and eighteen of fifty states restricting or considering restrictions (datacenterbans.com). The two counts are not contradictory — one counts local actions against projects, the other counts jurisdictional restrictions — but the gap between them is a reminder that no authoritative public census exists.

The outcomes in specific places are not in doubt; they sit in court records and board minutes.

Virginia. The Prince William Digital Gateway — adjacent to Manassas National Battlefield Park, and the largest data-center rezoning ever approved in the Commonwealth — was voided. On 31 March 2026, in Board of County Supervisors of Prince William County v. Oak Valley Homeowners Association, the Court of Appeals of Virginia held the three rezonings void ab initio for failure to comply with the advertising requirements of Va. Code § 15.2-2204(A). Judge Raphael wrote that "a public body must advertise proposed zoning ordinances in compliance with Code § 15.2-2204(A)," and that citizens may enforce that requirement in a timely challenge whether or not they took part in the public hearings. The ordinances would have permitted more than 20 million square feet of gross floor area — as many as 37 data centers running around the clock (Virginia Lawyers Weekly, 22 Apr 2026). QTS then dropped its appeal to the Virginia Supreme Court, ending the project. On 9 July 2026 the same board rejected the Dulles Cloud South proposal.

Ohio. A proposed constitutional amendment to ban data centers over 25 MW failed to qualify for the 2026 ballot, gathering roughly 70,000 of the 413,488 signatures required by the 1 July deadline — a useful corrective to the impression that opposition is uniformly ascendant. Local moratoria have nevertheless proliferated, including a three-month pause enacted by Cleveland on 15 July 2026.

Texas. The most consequential action in the country is not local at all. Governor Abbott's 3 August directive and ERCOT's consequent suspension of Batch Zero has effectively frozen new large-load approvals for a grid carrying roughly 90% of national data-center interconnection interest, pending an audit of some 250 to 300 projects.

Elsewhere. Arizona paused its data-center tax incentives on 13 June 2026. New York's moratorium passed 4 June 2026; New Jersey's Fair Share Act was signed 7 July; local moratoria took effect in Seattle (9 June), Spokane (23 June) and Broomfield, Colorado (7 July). Chandler, Arizona — the earliest of the modern fights, dating to 2014 — amended its zoning code in 2022 to require pre-construction sound studies, mitigation and post-construction noise verification, and unanimously rejected a proposed facility in 2025.

What the three case studies actually drew

The three anchors do not behave the way that aggregate would predict.

The Ohio PORTS project — the largest of them all, with nine gigawatts of new gas attached — has drawn no organised opposition. An independent tracker of Ohio projects records that "as of April 2026, no formal moratorium or local opposition group has been announced specific to the PORTS project" (stopohiodatacenters.org). Inside Climate News, reporting from the developer's open house at the Pike County YMCA, found more than 100 residents examining posterboards — including previously unreported detail on the three turbine types in the $33 billion project — and "no signs of an opposition campaign, which was notable, considering how often data centers have inspired organized resistance across Ohio and the rest of the country" (Dan Gearino, Inside Climate News, 6 Aug 2026). One retired educational aide from Jackson said "We do not need all these data centers." Hers was not the majority view. The explanation offered locally is historical and economic in equal measure: Pike County spent half a century dependent on the Portsmouth Gaseous Diffusion Plant, ranks near the bottom of Ohio counties for median household income and near the top for unemployment, and its county seat, Waverly, has about 4,100 people. A community that has already accepted uranium enrichment as its industrial identity, and needs the work, has a different baseline than a wealthy exurb that has neither. Concerns raised are specific rather than existential — groundwater management near contaminated aquifers, legacy contamination exposure for construction crews, and whether SoftBank's ratepayer commitment survives thirty years of operation.

The Memphis-area SpaceXAI complex is the opposite in every respect, and its chronology is the best-documented account of how one of these fights actually unfolds.

The Terafab project in Grimes County, Texas, is a third pattern again, and a reminder that the mandate's subject bleeds into adjacent industry. Terafab is a chip fabrication plant, not a data center: SpaceX and Tesla, with Intel providing silicon expertise, confirmed on 6 August 2026 a $16.8 billion first phase at the Gibbons Creek Reservoir site, roughly 9.5 miles southeast of Anderson, with a footprint exceeding 100 million square feet and at least 3,000 jobs; Governor Abbott awarded a $30 million Texas Enterprise Fund grant, and the project also qualifies under the Texas JETI programme (Office of the Governor, 6 Aug 2026). Austin's Giga Texas North campus hosts only an R&D pilot line; conflating the two is a common error.

The local reaction ran to type. The day before the announcement, 894 Grimes County residents presented the Commissioners Court with a petition seeking twelve specific protections before any tax breaks are approved: independent legal counsel paid for by the county, water and environmental safeguards, road-repair requirements, emergency-services funding, and clawbackA clause that takes an incentive back if the promise it was granted for is not kept. Tax breaks are awarded years before the jobs or the investment arrive; a clawback is what turns the developer's projection into an enforceable term instead of a press release. provisions if a developer misses its commitments. "If they damage our roads, they're the ones paying to fix our roads," said Bobbie Jo Swafford of Grimes County Citizens for Responsible Development. "Basically, you break it, you buy it." Commissioners accepted the petition and took no action, though Judge Joe Fauth said a group would fold its ideas into a resolution calling on the Governor and the Legislature to set statewide standards (KBTX, 5 Aug 2026). Land clearing began within days; an Iola resident a mile away reported charred wood and ash across her hundred acres (KBTX, 14 Aug 2026). SpaceX has committed publicly that the facility will draw from the reservoir's stored rainwater rather than groundwater and will not deplete the Navasota River. Both school districts endorsed the project through tax-incentive agreements. That combination — endorsement by the taxing authorities, petition by the neighbours — is the shape of nearly every one of these disputes.

11The documented grievances

Strip out the noise and six grievances have real evidentiary support. They do not all apply to every project, and saying which apply where is the whole job.

Electricity rates

The rate grievance is the best documented of the six, and none of it comes from advocacy groups. PJM's independent market monitor, Monitoring Analytics, attributes $6.3 billion — 38% — of the $16.4 billion in charges from PJM's most recent capacity auction to data centers, and $29.4 billion of $63.6 billion, or 46%, across the last four base residual auctionsPJM's main capacity auction, held once a year for a delivery period three years ahead. Generators bid to be available on the worst day of that future year; the clearing price is paid to everyone who clears, and the total lands on customers' bills. "Residual" because it buys what is left to cover after everything already contracted is counted. (Utility Dive, 20 Jul 2026). An earlier Monitoring Analytics analysis of the December 2024 auction attributed $6.5 billion of $16.4 billion — 40% — to data centers, of which $6.2 billion derived from forecasts of data centers not yet built, and observed that "the extreme uncertainty in the load forecasts based on uncertainty about the addition of large data center loads is also unique and unprecedented."

A capacity auction is not a market for electricity. It is a market for the promise that generation will be standing ready on the worst day of the year: the grid operator forecasts how much demand it must be able to serve, buys commitments to cover that number, and passes the cost along to everyone who takes power from the system. The forecast is the whole mechanism. Buy against a number that is too big and households pay, in cash, for power stations held ready to serve load that never turns up.

The cost is being driven substantially by forecast load, not delivered load. If the forecast is inflated by speculative interconnection requests, ratepayers pay real money for imaginary demand. This is the precise mechanism connecting §4's 80% queue-withdrawal rate to a household's bill, and it is why Texas's audit — a test of whether queued projects are real — is a consumer-protection measure as much as a reliability one.

Virginia's Joint Legislative Audit and Review Commission estimated in December 2024 that a typical Dominion residential customer could face generation- and transmission-related cost increases of $14 to $37 a month in real terms by 2040. The State Corporation Commission approved a new rate class for large load in November 2025, effective January 2027, requiring such customers to pay for at least 85% of contracted distribution and transmission demand and 60% of generation demand. In its 2026 rate order the SCC allowed Dominion base-rate increases of $565.7 million for 2026 and $209.9 million for 2027 against requests of $822 million and $345 million, with a typical residential bill rising $11.24 then $2.36.

Ohio produced the template. A tariff, in utility language, is not a tax. It is the published schedule of terms on which a utility must serve a class of customer — rates, minimums, contract length — approved by the regulator and binding on both sides. A minimum-take tariff writes a floor into that schedule: the customer pays for a set share of the capacity it reserved whether or not it ever draws the power. In case 24-508-EL-ATA the Public Utilities Commission of Ohio approved a settlement on 9 July 2025 creating a data-center-specific tariff: qualifying customers — new or expanded projects above 25 MW, and mobile data centers above 1 MW — must pay for at least 85% of contracted capacity whether or not they use it, on contracts of at least twelve years, with an exit fee of three years of minimum charges and financial assurances against cancellation. PUCO Chair Jenifer French called it "a well-balanced package that safeguards non-data center customers on an industrial and residential level," and the order simultaneously directed AEP Ohio to lift its moratorium on connecting new data centers — which matters, because the utility had already stopped connecting them (PUCO, 9 Jul 2025). The Ohio Manufacturers' Association challenged it; the case reached the Supreme Court of Ohio. The design principle is now near-universal: make the load pay for the infrastructure it causes, whether or not it shows up.

Georgia went further and made the forecast itself the object of regulation. Its Public Service Commission's own fact sheet records the revision: Georgia Power estimated in 2022 that it needed 400 MW of new generation over seven years; by 2023 the estimate was 6,600 MW; by mid-2025, 8,500 MW; and on 19 December 2025 the Commission certified 9,985 MW, roughly 80% of it expected to serve data centers, against a statewide fleet of about 22,000 MW (Georgia PSC Data Center Fact Sheet, March 2026).

A twenty-five-fold revision in three years is not a forecast; it is a queue readout. §14 returns to what Georgia did about it.

Water

Facility-level water disclosure is where the evidentiary record is genuinely thin, and it is important to say so rather than to fill the gap. Corporate fleet averages exist — Microsoft's 0.27 L/kWh, Meta's 0.18–0.20, Google's 10.9 billion gallons in 2025 — but a fleet average tells a specific county nothing. The number a community needs is this facility's annual withdrawal from this aquifer or this municipal system, and it is very rarely published. Where a project makes a water commitment it is usually in the form of a design claim, as at Helios and Terafab, rather than a metered figure with a reporting obligation attached.

The Texas audit is now asking every project in the ERCOT queue how much water it will consume, where it is coming from, and what cooling technology it will use. That is the first mandatory, project-level water disclosure regime of any scale in the United States, and it arrived through a grid operator rather than an environmental agency.

Noise

Noise is the grievance that most reliably converts neighbours into opponents, because it is continuous, nocturnal, and hard to prove. Documented figures: data hall interiors reach around 96 dB; diesel standby generators up to 105 dB; cooling towers on gas plant up to about 70 dBADecibels weighted the way human hearing works, discounting the very low and very high frequencies the ear is less sensitive to. The scale is logarithmic — every 10 dB is roughly a doubling of loudness — and the weighting is the reason a low-frequency hum can dominate a complaint while barely moving the meter. within 400 feet; Prince William residents have reported levels routinely exceeding 60 dB at homes (EESI). Loudoun County's ordinance limits industrial noise to 55 dB at a residential property line, and the county received a dozen complaints about one Vantage facility in a year, none of which was found to violate the ordinance — because the ordinances measure broadband sound levels while the complaint is about a tonal, low-frequency hum. Roughly a third of Northern Virginia's data centers sit within 200 feet of residential zoning.

The Southaven class action is the sharpest instance, and it is unusually specific about measurement. Filed 8 June 2026 in federal court in Mississippi (No. 3:26-cv-00148) against X.AI Corp., Space Exploration Technologies Corp. and MZX Tech LLC, it alleges residents measured levels exceeding 70 dBA at their property lines, in violation of local ordinance, from 57 temporary turbines running continuously since August 2025 — describing "high-pitched squealing, whining, roaring, low-frequency rumbling and tonal humming," and claiming anxiety, sleep loss, headaches, nausea and tinnitus. The class is defined geographically, from Stateline Road West down to Nail Road (ClassAction.org). Two further suits were filed in May 2026 over facilities in Dowagiac, Michigan and Vineland, New Jersey. These are allegations, not findings.

Air quality

Distinct from noise, and only relevant where there is on-site combustion — which is exactly why the behind-the-meter trend deserves scrutiny rather than applause. The Earthjustice/NAACP complaint alleges potential annual emissions from the Southaven turbines of more than 1,700 tons of nitrogen oxides, up to 180 tons of fine particulate matterAirborne particles small enough — under 2.5 microns — to pass the body's defences and lodge deep in the lungs. It is the pollutant with the best-established link to respiratory and cardiovascular harm, which is why it appears in nearly every air-permit fight., 500 tons of carbon monoxide and 19 tons of formaldehyde (Earthjustice case page). Those are plaintiff figures from a pending action, not findings. A study cited by TIME using NASA data reported nitrogen dioxide rising 3% in the post-xAI period around Memphis with peak episodes in late 2024 — a small effect, honestly reported, and stated precisely here because the larger claims circulating are not.

Land value — where the evidence genuinely runs out

The claim that a data center destroys neighbouring property values is made constantly and is, as of this writing, almost entirely unsupported by systematic evidence in either direction. The only formal study located here is "Data Centers and 2023 Home Sales in Northern Virginia," by Terry Clower and Keith Waters of George Mason University's Center for Regional Analysis, which mapped facility locations against 2023 sales across single-family homes, townhouses and condominiums and found that the farther a house was from a data center, the lower the sale price (GMU Schar School, Nov 2025).

That is the industry's favourite finding, and its own authors will not let it be used that way. Waters attributes the pattern "at least partially" to the amenities, infrastructure and employment that happen to cluster in the same places, and notes that many data centers sit where residents do not notice them. Trade press examining the study adds the harder objections: it covers only "data center alley," a region with two decades of siting experience, unusual wealth, and — a real community benefit — a 38% reduction in Loudoun County's property tax rate since 2010 (Data Center Knowledge). No study evaluates the effect across a wide area, and none of the paired-comparison figures circulating on either side has a published methodology.

So the honest position is uncomfortable for everyone. A resident who says a facility 200 feet from a bedroom window will hurt their sale price is making a plausible claim that nobody has tested. A developer who cites the Virginia study is quoting research whose authors disclaim the causal reading. Where the evidence does exist — noise, measured at a property line — it is a better proxy for the harm being described than any price series, which is probably why the litigation in §11 is pleaded in decibels rather than dollars.

Tax, and who actually nets what

The public debate is least well informed here, and the primary record is unusually good. Virginia's Department of Taxation, reporting to the legislature on 2 January 2026, disclosed that data-center operators claiming the retail sales-and-use tax exemption invested $32.0 billion in FY2024 and $48.6 billion in FY2025 — $80.6 billion over the biennium — of which $54.8 billion was exempt equipment and software, producing a reported tax benefit of $3.234 billion. Over the same two years those operators reported 2,807 net new direct jobs (Va. Dept. of Taxation, RD40, 2 Jan 2026).

Figure 23 · Evidence: Virginia Department of Taxation / VEDP, Biennial Data Center Retail Sales and Use Tax Exemption Report, 2 January 2026, Tables 1 and 2 (FY2024–FY2025); Loudoun County budget as reported by WUSA9 (Apr 2026). Method: $80,568,048,000 ÷ 2,807 = $28.70m; $3,233,847,000 ÷ 2,807 = $1.152m. Jobs are net new direct jobs reported by operators; including the 22,107 "additional jobs supported" modelled by VEDP would cut both ratios roughly ninefold. The Department states the underlying data "was self-reported by data center operators and was not independently validated." Synthesis: Both the industry's and the critics' headline numbers are in this table; which one is quoted determines which side of the argument the reader lands on.

The state's own return-on-incentive model finds cumulative tax revenues exceeding the exemption's value by $2.1 billion over five years, an average annual return of 11% and a revenue-to-cost ratio of 1.7 — a positive result on the state's own assumptions, and the strongest version of the industry's case — from a regulator rather than a trade group. Set against it: Good Jobs First estimates roughly $267 million of forgone revenue would otherwise have supported K-12 education statewide.

The structure matters more than the net. The Commonwealth forgoes about a billion dollars a year; Loudoun County alone budgets roughly $1.3 billion of data-center personal property tax, about 40% of its general fund. The state pays and the host county collects. That single asymmetry explains a great deal of otherwise puzzling politics — why state legislators from non-host districts file restriction bills while host-county supervisors approve projects, and why the fiercest local fights happen in counties that are adjacent to the tax base rather than inside it.

At smaller scale the mechanism is a PILOT — a payment in lieu of taxes. Dickens County commissioners signed a $20.5 million, ten-year agreement with Galaxy: $18.5 million of PILOT payments plus $2 million earmarked for housing, with an initial $500,000 and the $2 million housing tranche due 1 October 2026, administered by a separate board (The Texas Spur, 19 Aug 2026). That is what a community benefit agreement looks like in a county of a few thousand people, and it is a real transfer of value to a county that would otherwise see none of the project's tax base.

Jobs

Every project is announced with two numbers and the public hears only the larger one. Ohio: 35,000 construction jobs over a six-year buildout through 2032, and 2,500 permanent operating jobs. The ratio is fourteen to one. Helios: Galaxy expects over 100 permanent positions and CoreWeave perhaps 50 more, against roughly 1,350 craft professionals on site at construction peak.

Per megawatt, the permanent employment is thin and consistent: about 0.31 jobs per MW of IT load promised at PORTS-Pike, about 0.19 per MW of gross power at Helios. Virginia's whole exempt fleet — 62 reporting operators — added 1,610 net new direct jobs in FY2025 on $48.6 billion of investment. These are real jobs, often well paid, and they are not a jobs programme. The construction phase is the jobs programme, and it ends.

Neither side says the fair thing: data centers are an unusually good deal for a county's tax base and an unusually poor one for its labour market, and communities that evaluate them as employment projects are evaluating the wrong thing.

12Who organises, and how the fights are actually won

The people who show up are not who a national political frame would predict. It is overwhelmingly local, frequently bipartisan, and organised around proximity rather than ideology — the reliable predictor of whether someone opposes a project is how far they live from it. Data Center Watch describes statewide moratorium proposals introduced from both sides of the aisle in fourteen states.

The work is split between three kinds of organisation, and they are not interchangeable.

Neighbour groups form after an announcement, are geographically tight, and specialise in turnout and in the public record — the Grimes County petition of nearly 900 signatures with twelve named protections is the archetype. They rarely win outright and they are extremely effective at raising a project's cost and timeline.

Established environmental and civil-rights organisations supply what neighbour groups lack: standing, counsel, and evidentiary capacity. The Memphis case is the clearest demonstration. What broke the story open was not a protest but a public-records request, an aeroplane chartered with Southwings, and a thermal camera — SELC's senior attorney describes it as providing "the visual and technical evidence that forced public accountability." The NAACP supplied the plaintiff. Earthjustice and SELC supplied the Clean Air Act expertise. That combination — local standing plus national litigation capacity — is the only one that has produced federal court orders.

Formal intervenorsParties admitted to a regulatory case with the right to file evidence, cross-examine witnesses and appeal the outcome — consumer advocates, industrial customers, environmental groups. They are not the public commenting; they are litigants inside the proceeding, and the record they build is what a commission is obliged to decide on. in regulatory proceedings. The least visible and probably the most consequential. The Ohio Consumers' Counsel was a party to the AEP Ohio settlement that produced the 85% minimum-take tariff. The Ohio Manufacturers' Association challenged it. Georgia's PSC ran 39 hours of hearings on 1,200 pages of sworn testimony with 35 cross-examined witnesses in one phase, and 64 hours on 2,444 pages with 55 witnesses in another. Nobody marched. The outcome — that Georgia Power must backstop the cost of 9,985 MW of new generation through 2031 if the data-center contracts do not materialise — protects more households than any moratorium yet passed.

Procedure beats merits — the decisive tactical fact of the last two years. Digital Gateway was not defeated on water, noise, traffic or viewshed, all of which were argued at length. It was defeated because a statutorily required advertisement did not run in the newspaper, making three rezonings void from inception regardless of their substance. Prince William County's staff had done everything correctly. The developer, facing a restart of a multi-year entitlementThe full set of public approvals a site needs before anything can be built — rezoning, special-use permits, site plans, conditions. Winning them takes years and is the single most valuable thing a developer holds, because the land is worthless without them and they do not transfer to a project that has to start over. process it had already won once, walked away from 2,100 acres. Sophisticated opponents have learned the general lesson: notice defects, open-meeting violations, incomplete applications and permit-scope errors are faster, cheaper and more reliable than persuasion.

On funded opposition

The mandate asks for documented cases of funded opposition or funded support tied to competing commercial interests. This report found none that can be responsibly asserted, and says so explicitly. Data Center Watch's own funding and methodology are undisclosed, which is a reason to be careful with its numbers, not evidence of anything about its motives. Trade-aligned "mythbuster" material and advocacy-aligned "greenwashing" briefs both exist in volume and both have obvious institutional interests, but an interest is not a payment and this report will not treat it as one. Where a campaign's funding could not be verified, that is stated rather than inferred. Anyone alleging astroturf in either direction should be asked for the filing.

13The fact-check

Both sides of this argument are running on claims that do not survive contact with a primary source. The table below adjudicates seventeen of the most widely circulated, in both directions, against the strongest document available. The rule applied throughout: an industry "mythbuster" and an advocacy brief are each treated as a claim to be tested, never as a source to be adopted.

Two of those rows deserve a paragraph rather than a cell, because they are the flashpoints the mandate singles out.

On the EPA turbine rule. The obvious reading is that one of the two sides must be spinning. Read the rule and neither is. The EPA declined to bring stationary temporary turbines within the "nonroad engine" definition at 40 CFR 1068.30 — which is precisely what advocacy groups had argued and what the Shelby County Health Department had denied, so their claim is correct. In the same document the EPA created an optional temporary subcategory with genuinely lighter compliance and separately exempted non-major-source turbines from Title V, which is what the law firms told their clients, so theirs is correct too. The rule tightened the principle and loosened the paperwork. Anyone quoting one half of it is quoting accurately and describing the rule falsely.

On water. What to take from this section is a question, not a number. When a developer says "closed loop," ask what happens on the secondary side — cooling tower or dry cooler. The first answer means the facility consumes water continuously and the community should ask from what supply. The second means it does not, and the community should instead ask what the higher electricity draw does to the local load. There is no configuration in which the facility is free.

14The policy response, and where this goes

Policy has moved further in eighteen months than the public argument has, and it has converged on a small number of instruments that will decide the next decade.

Minimum-take tariffs. The dominant instrument. Ohio's 85% of contracted capacity over at least twelve years with a three-year exit fee; Virginia's 85% of contracted transmission and distribution demand and 60% of generation demand from January 2027; Georgia's requirement since early 2025 that customers above 100 MW sign bespoke contracts covering upstream generation, transmission and distribution costs, filed with the Commission thirty days before execution. The design intent is identical everywhere: convert a forecast into an obligation, so that a project that does not materialise costs its developer rather than the ratepayer.

Backstops and clawbacks. Georgia's is the most aggressive and the least covered. Alongside certifying 9,985 MW, the Commission secured Georgia Power's agreement to financially backstop the associated generation costs through 2031, froze base rates through 2028, and reserved four remedies if contracts do not appear: suspend unstarted generation projects, retire other plants early, terminate out-of-state purchase power agreements, or sell surplus into the market. That is a regulator pricing the option that the AI buildout disappoints — and putting the premium on the utility.

Verification. Texas's audit is the newest instrument and potentially the most transformative, because it attacks the input rather than the output. If a grid operator can force every queued project to disclose its incentives, its self-generation, its water source and cooling technology, its noise mitigation and its ultimate ownership — on pain of ineligibility for false or absent answers — then the speculative duplication that inflates load forecasts becomes expensive. Every other jurisdiction is watching, and the result will be public within months.

PreemptionA higher level of government overriding a lower one's authority to decide a question. A state that preempts its counties can strip them of the power to say no to a project — or, written the other way, can strip them of the power to say yes on their own terms. Counties hold only the powers the state has lent them, which is why this fight is settled in state capitols rather than council chambers., in both directions. Over 300 state data-center bills were filed in the first six weeks of 2026 across more than thirty states. The narrower NCSL tracker, updated 1 July 2026, counted data-center moratorium or banning bills in fifteen states, with New York's one-year moratorium passed by the legislature, Virginia continued, and Georgia among the failed or stalled measures (NCSL, updated 1 Jul 2026). The fault line is no longer incentives versus no incentives; it is state authority versus local authority. West Virginia limited local control over data centers and microgrids, and some legislators are now moving to restore it. Pennsylvania's Senate Bill 991 would have the Department of Environmental Protection designate at least fifteen pre-approved sites with an accelerated two-phase permitting process — a state offering developers a route around county politics. Meanwhile Minnesota established a separate water-permitting requirement for data centers in 2025 and more states are following, and South Carolina and Maryland enacted rate-negotiation laws.

Consumption taxes. Virginia has enacted a first-of-its-kind energy consumption tax of $0.011 per kWh on data centers — a mechanism that neither restricts siting nor subsidises it, but prices the externality directly.

How developers are adapting

The case studies show three adaptations, each a rational response to the constraints above.

Go where the industrial consent already exists. The largest AI campus ever announced went to a county with a fifty-year history of federal nuclear industry and drew no organised opposition. A 2,100-acre project next to a Civil War battlefield in a wealthy Virginia exurb was fought for years and killed. Developers have noticed. Legacy federal and industrial sites — remediated DOE land, retired coal plants, failed bitcoin mines — carry pre-existing interconnection, pre-existing zoning, and a population that has already made its peace with heavy industry. Expect the next wave to concentrate there.

Bring the power. Behind-the-meter generation started as a schedule optimisation — a way around a five-year queue — and has become a political instrument. "We are not taking electricity from the grid" is now the first line of the Ohio project's community messaging, and it is the most effective answer to the rate-impact argument. The unresolved question is whether communities will accept nine gigawatts of new gas combustion as the price of not affecting the grid; Pike County, so far, has.

Pay, visibly and locally. The community-benefit agreement has gone from afterthought to standard term. PORTS-Pike carries an $80 million community benefits fund — $40 million from SB Energy, matched by OpenAI — directed at affordable energy, workforce development and economic development. Memphis City Council allocated 25% of xAI's tax revenue to the nearest neighbourhoods. Dickens County negotiated $2 million specifically for housing, plus a workforce hub, because the constraint on a county of a few thousand people hosting 1,350 construction workers is beds, not jobs. Galaxy has also gone further than the county line: a 15-year, $75 million agreement announced in July 2026 makes it Texas Tech Athletics' official data-center and digital-assets partner — renaming the football stadium Galaxy Stadium — with the university describing it as the first step in a broader relationship on AI research, workforce development and economic investment across West Texas (Texas Tech Athletics, 17 Jul 2026).

§Fifteen findings worth repeating

Each stands alone, and each rests on evidence established above.

1. 1A single AI rack draws roughly thirteen times what a typical rack does. NVIDIA's GB300 NVL72 sits at 135 kW of design power — up to 155 kW at peak — against a typical rack's ~10 kW. That one number is why old data centers cannot be reused and new gigawatt sites must be found.
2. 2NVIDIA's $105 billion Ohio guarantee is reimbursed by the company whose failure triggers it. The 8-K discloses residual value guaranties capped at $105bn, triggered by OpenAI's insolvency or non-payment — and states that "OpenAI has agreed to reimburse and indemnify NVIDIA for any and all amounts actually paid." The guarantee also terminates if OpenAI obtains a satisfactory credit rating.
3. 3Nothing in this industry is contracted between year six and year twelve. CoreWeave's entire $103.7bn contracted book runs off by month 78; the colocation leases underneath run 12–15 years, hyperscaler leases up to 30, and Digital Realty depreciates buildings over as much as 39. The gap is where the buildout's economics will be settled.
4. 4Virginia's largest data-center rezoning was voided because a newspaper advertisement never ran. The Court of Appeals held the 2,100-acre Digital Gateway rezonings void ab initio for a notice defect under Va. Code § 15.2-2204(A); QTS dropped its appeal and abandoned the project. Opposition wins on procedure, not on merits.
5. 5Georgia Power's forecast of new generation needed went from 400 MW to 9,985 MW in three years — a twenty-five-fold revision. The certified figure is roughly 45% of the utility's entire existing 22,000 MW statewide fleet, and about 80% of it is expected to serve data centers.
6. 6Virginia forgave $3.23 billion of sales tax over two years for 2,807 net new direct jobs — about $1.15 million per job. The same operators invested $80.6 billion over the period, or $28.7 million of capital per direct job. The state's own model still finds a 1.7 revenue-to-cost ratio, on self-reported and unvalidated inputs.
7. 7The state pays and the county collects. Virginia forgoes roughly a billion dollars a year statewide while Loudoun County alone budgets about $1.3 billion of data-center property tax — some 40% of its general fund. That asymmetry, not ideology, explains why state legislators restrict and county supervisors approve.
8. 8Data-center load accounts for 46% of PJM's capacity charges across four auctions — $29.4 billion of $63.6 billion. In an earlier auction the independent market monitor attributed $6.2bn of the $6.5bn charged to data centers to forecasts of facilities not yet built. Households are paying today for load that may never arrive.
9. 9"Closed loop" is a claim about the wrong loop. It describes the sealed circuit at the chips and says nothing about heat rejection outside, which is where all the water goes. A closed-loop facility with an evaporative cooling tower consumes water continuously; the same facility with a dry cooler consumes almost none and more electricity.
10. 10Microsoft's own water intensity varies elevenfold across its own fleet — 0.03 L/kWh in EMEA against 0.34 in the Americas. Any argument that treats "data center water use" as a single industry number is wrong before it starts; climate and cooling design dominate.
11. 11Nine of the ten gigawatts of "new energy generation" at the largest AI campus in America is natural gas. The PORTS-Pike project discloses 9.2 GW of gas backed by about $33 billion of investment — genuinely protective of the grid and of ratepayers, and not what most readers hear in the phrase "new energy generation."
12. 12Announced capacity and delivered capacity differ by roughly an order of magnitude. Stargate UAE's first phase of 200 MW is 4% of its announced 5 GW campus; Berkeley Lab finds about 80% of projects entering US interconnection queues eventually withdraw.
13. 13The largest AI campus ever announced has drawn no organised local opposition, and it sits on a former uranium enrichment site. Pike County, Ohio has hosted federal nuclear industry for fifty years. Consent, not resources, is what varies between the projects that get built and the ones that do not.
14. 14Permanent employment runs about 0.2 to 0.3 jobs per megawatt. PORTS-Pike promises 35,000 construction jobs and 2,500 permanent ones — fourteen to one. Helios expects roughly 150 permanent positions across landlord and tenant for 800 MW of contracted gross power.
15. 15A grid operator, not an environmental agency, has imposed America's first project-level water disclosure regime at scale. Texas's August 2026 audit requires every data center in the ERCOT queue to state its water consumption, source and cooling technology — or be denied connection.

§Evidence register

Every source below was opened and read in the course of preparing this brief. Tier 1 = issuer filings, regulator, government and primary technical documents. Tier 2 = wire, trade, peer-reviewed and reputable local reporting, and named-party litigation summaries. Tier 3 = single-source or aggregated trackers whose methodology or funding is not disclosed; these carry no load-bearing conclusion on their own and are labelled as such in the text.

§Not verified — and carrying no conclusion

The following could not be confirmed against a primary source in the course of this work. None of them supports any claim, figure or verdict above; they are listed so a reader can see exactly where the record stops.

Evidence cutoff 19 August 2026. Figures are stated as of the dates given; several threads in this subject — the Texas audit, the Ohio Power Siting Board proceeding, the Clean Air Act litigation in Mississippi, and the PUCT's treatment of ERCOT's good-cause request — were unresolved at that date and will have moved. This document is an explanatory brief prepared by AIR Research. It is not investment advice, not a recommendation, not a solicitation, and contains no per-company investment verdicts.