Counting the Gigawatts: Real Estate in the Decade Ahead

You can see the future first in the interconnection queue.

The rest of the industry is talking about the maturity wall, the pace of cuts, and whether office has bottomed. Those are real questions and they are the wrong ones, in the same way that debating freight rates in 1955 was real and wrong while the Interstate Highway Act was being drafted. The composition of American real estate is being reset right now by a variable that does not appear on a single page of a standard offering memorandum, is not taught in any real estate program, and is not tracked by any of the four major brokerages as a primary metric.

The variable is interconnected electrical capacity. Not power in the abstract. Contracted, energized, deliverable megawatts at a specific parcel on a specific date.

Here is the compression. In 2025, four companies spent roughly $410 billion of capital expenditure. In 2026 those same four companies have guided to approximately $725 billion, a 77 percent increase, of which something like three quarters is directed at physical AI infrastructure. For scale, total annual construction spending on all multifamily housing in the United States runs on the order of $110 to $130 billion. Four corporate capital budgets now exceed the entire American apartment construction industry by a factor of five or six, and the gap is widening at 70 percent a year.

That is not a technology story that happens to touch real estate. That is a real estate story that the real estate industry is not in the room for.

This piece is the full argument. Part I counts the gigawatts and converts them into the units this industry actually understands: acres, square feet, dollars, and tradespeople. Part II shows why power became the binding constraint and why constraints of this type propagate. Part III works through the four second-order consequences, one of which is a direct and largely unremarked collision with the housing shortage. Part IV is what specifically gets repriced, with the arithmetic. Part V gives three dated scenarios with probabilities. Part VI is what would prove this wrong, which is the section most essays of this genre omit and the one that makes the rest worth reading.

A note on the numbers. Figures here are drawn from Lawrence Berkeley National Laboratory's Queued Up 2026 edition (published June 2026, data through year-end 2025), Goldman Sachs Research and S&P Global 451 Research load forecasts, company capital expenditure guidance, Berkeley Lab and industry cost benchmarks, NAHB and Census construction data, Gallup and Good Jobs First survey work, and OEM backlog disclosures. Forecasts in this sector diverge wildly by methodology, and I flag the ranges where they matter rather than picking the most dramatic number. Nothing in Part V is a fact. It is labeled as a scenario because that is what it is.

Part I: Counting the Gigawatts

The line that went flat and then did not

American electricity demand was essentially flat from roughly 2005 to 2021. Retail sales hovered near 3,800 terawatt-hours through a period in which the population grew by 40 million and real GDP grew by more than a third. Efficiency gains in lighting, motors, and appliances absorbed the growth. An entire generation of utility planners, regulators, transmission engineers, and, critically, real estate developers built their careers inside a regime where load growth was zero and power was a solved problem you called the utility about after you had already bought the land.

That regime ended.

Data centers consumed roughly 4.1 percent of US peak summer power demand in 2025. Goldman Sachs Research puts 2026 at 5.3 percent, with peak demand rising from about 31 GW to about 41 GW. S&P Global's 451 Research, using a broader definition that captures all utility power supplied to hyperscale, leased, and crypto facilities, puts 2026 at 75.8 GW. The definitions differ enough that you should not average them, but they agree on the shape: year-over-year capacity additions scheduled at 13.6 GW in 2026, an acceleration rather than a plateau.

Now the queue. As of year-end 2025, more than 2,060 GW of generation and storage capacity was actively seeking transmission interconnection, and by mid-2026 the figure had swelled past 2,600 GW. The entire installed generating capacity of the United States is roughly 1,300 GW.

Read that twice. The queue to connect to the American grid is now twice the size of the American grid. Median time from queue entry to commercial operation is approaching five years, and the majority of projects that enter never get built.

Converting gigawatts into real estate

A gigawatt is an abstraction to most of this industry. Here is the translation table. These are ranges, because campus design, cooling architecture, and power density vary enormously, and anyone quoting you a single number is selling something.

One gigawatt of AI data center capacity requires	Quantity
Construction capital, shell and infrastructure	$11.3B at the standard benchmark, $15B to $25B for AI-optimized
All-in including GPU fit-out	$30B to $45B
Land, campus plus buffer and substation	1,000 to 2,000 acres
Building shell	5 to 7 million square feet
Annual electricity consumption at full load	8.76 terawatt-hours
Equivalent US residential consumption	Roughly 830,000 homes
Peak construction workforce	3,000 to 5,000
Skilled electricians at peak	800 to 1,500

Now multiply by the 13.6 GW of additions scheduled for 2026 alone.

$154 billion of construction at the standard benchmark. $204 billion at the AI-optimized shell benchmark. Somewhere between 13,600 and 27,200 acres, which is 21 to 43 square miles of new campus. Between 68 and 95 million square feet of shell, which is roughly the entire office inventory of Philadelphia. And residential-equivalent electricity consumption of 11.3 million American homes, which is more than the total housing stock of the state of Florida.

In one year. Of additions.

That last figure is the one to sit with. Every year from here, the incremental data center load being added to the American grid consumes more electricity than an entire large state's worth of houses, and it is being added to a grid whose planners spent two decades assuming load growth of zero.

The order-of-magnitude framing

The useful way to hold this is in orders of magnitude, and there are three of them stacked.

The first OOM is scale. Data center load goes from a rounding error to a low single-digit percentage of national electricity in about five years, and credible forecasts put it in the high single digits to low teens by 2030. That is one order of magnitude of share growth, and share growth in a system with fixed capacity is the definition of a scarcity event.

The second OOM is capital density. A hyperscale AI campus runs $11M to $25M per megawatt on the shell alone, and $30M to $45M per megawatt all-in. Compare a garden apartment at roughly $290,000 per unit, or an industrial warehouse at $150 to $200 per square foot. Per acre of land consumed, a 1 GW campus deploys somewhere near $30 billion across 1,200 acres, or $25 million per acre. That is Manhattan land economics arriving in rural Ohio, Louisiana, and West Texas, and the local land market has no framework for pricing it.

The third OOM is time compression. These campuses are financed, designed, and delivered on 24 to 36 month schedules by counterparties with investment-grade balance sheets and no patience. Against that, a large power transformer is now a 60-month lead item and a heavy-duty gas turbine is five to seven years out. The buildout is attempting to run one order of magnitude faster than its own supply chain.

Those three facts, stacked, produce every consequence in the rest of this article.

Part II: The Entitlement Is Now the Interconnect

Why power became the constraint and why it will stay one

Capital is not the constraint. That is the first thing to internalize, because the entire instinct of this industry is to assume that whatever is scarce is scarce because money is scarce. There is more capital chasing AI infrastructure than there are places to put it. The constraint is upstream and it is physical.

Large power transformers, the multi-hundred-ton units that step transmission voltage down to distribution, had 24 to 30 month lead times before 2020. They now exceed five years, with Tier 1 manufacturers quoting 60 months or more. Heavy-duty gas turbines are worse. GE Vernova's gas turbine backlog reached 100 GW in the first quarter of 2026 and the company expects 110 GW of orders and slot reservations by year end. Across GE Vernova, Mitsubishi Power, and Siemens Energy the combined backlog exceeds 170 GW, with some frames sold into the next decade. Wait times on large units like the M501JAC have reached seven years.

Behind those, the interconnection process itself: five years median, most applicants withdraw, and study queues that were designed for a world where a handful of merchant generators applied each year are now processing thousands of requests.

None of these bottlenecks can be solved with money on a real estate timeline. You cannot pay a transformer factory into existence in eighteen months. Industry estimates suggest nearly half of planned US data center developments could be delayed or canceled on power infrastructure and component shortages alone.

This is the definition of a durable constraint: a physical input with a multi-year replacement cycle, no substitute, and demand growing faster than the supply of the machines that make it.

What that does to land

When the scarce input is not the dirt, the dirt stops being the asset.

The market has already begun to reprice this and the industry has largely not noticed, because it is happening in transaction types that do not show up in the standard indices. Reported data center land pricing averaged roughly $244,000 per acre in 2024 and has kept climbing. Parcels of 50 acres and up are running 23 percent higher year over year, with average transaction size around 224 acres. And the pricing driver has explicitly inverted: the deciding variable in these transactions is an executed power commitment, not location, not acreage, not even entitlement in the traditional sense.

Two adjacent parcels, identical soils, identical zoning, identical highway access. One has an executed large-load interconnection agreement with a defined energization date. The other does not. These are not comparable properties. They are different asset classes, and the spread between them is the single largest valuation discontinuity in American land right now.

Here is what that discontinuity is worth, computed rather than asserted.

The powered-site arbitrage, worked. Consider a shuttered industrial site: an old aluminum smelter, paper mill, or steel finishing plant with 300 MW of existing, still-energized service. There are dozens of these across the Ohio Valley, the Southeast, and the Pacific Northwest. As industrial land, 400 acres at $30,000 per acre plus a functionally obsolete building is worth perhaps $12 million.

Now value the power. A stabilized data center is worth roughly $11.3M per MW at the standard benchmark. Assume a 20 percent development margin, so the developer profit is $2.26M per MW. The site with existing service can be delivered now. An equivalent greenfield site must wait five years for interconnection, assuming it gets there at all. Discounting that profit five years at a 12 percent cost of capital:

$2.26M × (1 − 1 / 1.12⁵) = $2.26M × 0.4326 = $978,000 per MW of pure acceleration value

Across 300 MW, that is $293 million, on a site whose dirt is worth $12 million. A 24x, and it exists entirely in the electrical service, not the real property.

The honest caveats matter here and they are exactly where the edge lives. Existing service does not mean available capacity, because the utility may have reallocated it after the plant closed. The transmission behind the substation may be constrained even if the substation is not. The interconnection agreement may not survive a change of use or a change of ownership. Every one of those is a diligence question that a competent real estate developer is structurally equipped to answer and a software company is not. That asymmetry is the opportunity.

Real estate people are the right people and are not in the room

Strip the technology away and look at what this business actually is.

It is a process of acquiring land subject to a discretionary approval, navigating a multi-year regulatory queue with an inflexible public counterparty, managing a community that does not want the project, financing a long-dated asset against an uncertain residual, and delivering vertical construction on a schedule. That is real estate development. That is precisely, exhaustively, the job.

The interconnection queue is an entitlement process. It has an application, a study period, a set of technical conditions, a cost allocation fight, a negotiation, and a political dimension. Developers have been running exactly this playbook against planning commissions for eighty years. The counterparty is now a regional transmission organization instead of a city council, the study is a system impact study instead of a traffic study, and the mitigation is network upgrades instead of a roundabout, but the skill is identical.

The people currently running this process are largely infrastructure funds, energy developers, and corporate real estate teams inside technology companies. They are competent at power and thin on land assembly, community relations, and the specific craft of making a hostile local approval come out your way. Meanwhile the industry that has spent a century mastering exactly that craft is on the sidelines, mostly because it has decided that data centers are somebody else's asset class.

They are not somebody else's asset class. They are the largest land use event since the interstate highway system, and they are being executed by people who have never lost a zoning hearing and are about to lose a great many.

Part III: Four Consequences Nobody Is Connecting

IIIa. The labor wall, and why the AI buildout raises your rent

The construction industry entered 2026 roughly 439,000 workers short, with some projections putting the 2026 shortfall near 499,000. The shortage is concentrated precisely where data centers consume labor: electricians, high-voltage field technicians, mechanical contractors, controls specialists, and commissioning teams.

The price signal is not subtle. Data center construction work pays up to 30 percent above typical construction wages. Specialized electricians in Northern Virginia and Texas are commanding compensation up to $280,000, against a national median electrician wage near $62,000. Contractors taking data center work report backlogs approaching a year.

Now connect that to a fact from an entirely different conversation.

Multifamily starts fell 40.2 percent in May 2026 to a 295,000 annualized pace. Overall housing starts fell 15.4 percent to 1.18 million. NAHB projects multifamily starts down 5 percent in 2026 and another 6 percent in 2027. New apartment deliveries are forecast to fall 28 percent in 2026 to roughly 382,000 units. Single-family starts are at their lowest level since 2019.

The consensus explanation is interest rates and affordability, and that explanation is substantially correct. It is also incomplete, and the missing piece is a straightforward crowd-out.

There is one pool of electricians. Every journeyman pulling wire on a hyperscale campus in Loudoun County or Abilene is a journeyman not wiring an apartment building. When one buyer can pay $280,000 for the same license that another buyer budgeted $95,000 for, the labor does not split the difference. It moves. And it moves fastest in exactly the metros where data center development and multifamily development overlap: Phoenix, Dallas, Atlanta, Columbus, Northern Virginia, Salt Lake City.

The magnitude is estimable. At 800 to 1,500 electricians per gigawatt at peak, 13.6 GW of 2026 additions absorbs somewhere between 11,000 and 20,000 electricians into a trade already short by more than that. Against a national electrician workforce in the neighborhood of 800,000, that is 1.4 to 2.5 percent of the entire trade redirected in a single year, concentrated in perhaps fifteen counties, at a 30 percent wage premium, on multi-year contracts.

The consequence, stated plainly: the AI infrastructure buildout is inflationary for housing construction, and the mechanism is the labor market, not the capital market. Every marginal gigawatt raises the cost and lengthens the schedule of every apartment building competing for the same trades in the same region. This is not a talking point either side of the housing debate is currently making, and it is arithmetic.

There is a real counterargument and it deserves space. Wage premia of this size are the most powerful recruiting mechanism the trades have had in forty years. Commercial electrical apprenticeship applications rose 70 percent between 2022 and 2024, and roughly 60 percent of Gen Z workers report intent to pursue skilled trade work. If a decade of $200,000 electrician jobs pulls 200,000 people into the trade, the long-run effect on housing construction cost is deflationary, because the constraint that has throttled homebuilding since 2009 finally clears. The apprenticeship pipeline runs four to five years, which means the crowd-out is a 2026 to 2030 problem and the relief is a 2030s benefit. Both things are true. The sequencing is the whole story, and it is brutal for anyone trying to build housing in the interval.

IIIb. The ratepayer revolt, which is a land use fight

The political reaction has arrived faster than almost anyone in the industry modeled, and it is the constraint most likely to bind before the physical ones do.

Residential electricity rates rose 7.3 percent between April 2025 and April 2026 by federal analysis. Voters have connected that to data centers, correctly or not, and the polling is remarkable: a Gallup survey found 71 percent of Americans oppose a data center near them, including 48 percent strongly, which is higher opposition than to a nearby nuclear plant. A February 2026 Wisconsin poll found 70 percent of voters believe the costs of new data centers outweigh the benefits, up from 55 percent six months earlier. That is a 15 point swing in half a year, which is the signature of an issue that is still moving.

The institutional response followed. New York enacted the first statewide moratorium on new large-scale data center permits pending up to a year of ratepayer and environmental standard-setting, and roughly a dozen states are considering similar action. Of at least 63 local data center moratorium actions tracked by Good Jobs First, 54 had passed. Governors in Montana, Wyoming, and Missouri joined a federally brokered Ratepayer Protection Pledge. And the cost is already measurable: 75 major projects worth more than $130 billion were delayed or canceled amid organized local opposition in the first quarter of 2026 alone, including a $12 billion Wisconsin campus killed outright.

$130 billion of projects stopped in ninety days by local land use politics. That number should be the most-quoted figure in this entire industry and almost nobody in real estate has seen it.

Two implications.

The first is that market selection now runs through political geography. The winning data center markets over the next five years will not be the ones with the cheapest power or the best fiber. They will be the ones where the community benefit structure, the tax abatement arithmetic, and the ratepayer cost allocation have been negotiated in a way local voters will tolerate. That is a land use competency, and states and counties that develop it will capture tens of billions of dollars of investment from neighbors that do not.

The second is the arbitrage. This is a fight over zoning, entitlement, community benefits agreements, host agreements, and tax increment structures. The technology industry is spending $725 billion a year and is, institutionally, terrible at this. It sends site selection consultants and lawyers to hearings where the winning move is eighteen months of relationship-building, a genuinely negotiated community benefit, and a local partner with standing. Real estate developers do this for a living. The single highest-return skill in this cycle may turn out to be the ability to get a controversial project approved in a hostile county, and that skill is currently sitting in multifamily and industrial development shops that have not realized what it is now worth.

IIIc. The demand-side question, honestly

Here is the question the industry avoids because the answer is genuinely unknown: if machines do a growing share of cognitive work, what happens to the five billion square feet of American office?

Start with what is measurable. Office-using employment is forecast to grow 0.3 percent over the 2026 to 2030 period. Essentially flat. Office-using employment has never been flat or negative across a five-year window without an accompanying recession. Not once, through every technology transition since the sector has been measured. National office vacancy sits somewhere between 14 and 21 percent depending on whose universe you use, with Moody's at a record 21 percent in the first quarter of 2026 and CoStar's broader universe at 14 percent. Office-to-residential conversions hit a record, with 11.8 million square feet completed or under construction and roughly 90,300 apartments in the conversion pipeline at the start of 2026, up 28 percent year over year.

Now the two theories.

The substitution theory says AI removes cognitive labor, cognitive labor occupies offices, therefore offices empty. The supporting evidence is real: Big Tech hiring of new graduates is down roughly 50 percent from pre-pandemic levels, and early-career candidates fell to 7 percent of Big Tech hires in 2024, down 25 percent year over year.

The Jevons theory says that when the cost of a capability collapses, total consumption of it rises rather than falls, because uses that were previously uneconomic become economic. Cheaper cognition means more analysis, more products, more firms, more coordination, and coordination is what offices are for.

Both theories are too coarse, and the actual answer is a third thing: recomposition, which produces much less square footage change than either camp expects.

Work it. A 200-person professional services firm with 60 junior staff in open plan at 110 square feet each and 140 mid-level and senior staff at 220 square feet each occupies 37,400 square feet of assigned space, or roughly 50,500 square feet with a 35 percent common area factor. Now cut the junior cohort in half, which is an aggressive assumption about AI substitution.

New assigned space: 30 × 110 + 140 × 220 = 34,100 square feet, or 46,000 with common area.

Headcount falls 15 percent. Footprint falls 8.8 percent.

The reason is that the roles most exposed to automation are the roles with the smallest spatial footprint. Cutting the bullpen does not empty the building. If AI instead increases the value of senior judgment, and senior people occupy more space per head, the footprint effect could be close to zero even with meaningful headcount reduction.

So the office collapse thesis is weaker than its advocates think on a per-firm basis. The real signal is the aggregate one, and it is more consequential than the per-firm math: office-using employment going flat for five years outside a recession would be the first time in the history of the sector that the growth engine simply stopped. Office does not crater. It stops growing, structurally and permanently, for the first time since the invention of the elevator.

For an asset class that has always been underwritten with an implicit assumption of secular demand growth, permanent flatness is a more dangerous regime than a crash. A crash reprices and clears. Flatness means every building must now win its tenants from another building, forever, and the only reliable winner in a zero-sum leasing market is the newest, best-located asset. The middle of the office stock does not get a recovery. It gets converted, or it gets demolished, or it sits.

IIId. The capital stack inverts, and the residual problem nobody prices

For thirty years the defining feature of the technology industry was that it was asset-light. That was the entire investment thesis: infinite marginal returns, no factories, no inventory, no real estate.

That industry no longer exists. Four companies are deploying $725 billion of capital expenditure in a single year, roughly three quarters of it into physical infrastructure. Microsoft disclosed that $37.5 billion of a single quarter's capex went to short-lived assets, primarily GPUs and CPUs with three to five year useful lives. The most valuable companies in the world have become, in the most literal accounting sense, industrial real estate developers who also write software.

The consequence for this industry is that the best credit tenants in the history of commercial real estate have arrived simultaneously, wanting fifteen and twenty year leases, on purpose-built assets, in enormous size. On the surface that is the best leasing environment anyone has ever seen.

Underneath it is a residual value problem of a kind real estate has not previously encountered, and it is being systematically underpriced.

The mismatch. An apartment building has a 60-year physical life and an economic life limited mainly by location and maintenance. A warehouse built in 1995 is still a warehouse. A data center built for 8 kW per rack air cooling in 2020 is functionally obsolete for 130 kW per rack liquid-cooled AI training in 2026. The shell has a 30-year physical life and an unknown, possibly 10 to 15 year, functional life, driven by a compute architecture that changes every 18 months. Nobody knows what a 2026-vintage AI hall is worth in 2041 because nobody knows what compute looks like in 2041.

What that is worth in basis points. Take a 15-year NNN lease and ask what going-in yield the residual assumption implies. If you underwrite the building at 100 percent of cost in real terms at year 15, you can accept a low going-in yield. If the honest residual is 40 percent of cost, you must amortize the other 60 percent over the lease term. At an 8 percent discount rate, the 15-year annuity factor is 8.56, so recovering 0.60 of cost requires:

0.60 / 8.56 = 0.0701, or 701 basis points of additional annual yield

Even at a generous 70 percent residual, the shortfall of 0.30 requires 0.30 / 8.56 = 350 basis points.

That is the range. The residual assumption on a data center is worth somewhere between 350 and 700 basis points of required going-in yield, and it is the least examined number in the sector. Deals are being priced today off tenant credit, which is genuinely excellent, and off comparable transaction cap rates, which encode the same unexamined residual assumption as each other. The credit tells you that you will collect rent for fifteen years. It tells you nothing whatsoever about what you own in year sixteen, and in a purpose-built asset with a technology-driven obsolescence cycle, year sixteen is most of the value.

The bear case, stated fairly. Power infrastructure built for 2024 and 2025 demand levels risks becoming stranded by 2027 and 2028 if AI demand disappoints, because the lag between groundbreaking and energization means today's commitments are bets on demand three years out. An AI infrastructure company with a weak business model has perhaps 36 months before its GPU fleet is simultaneously obsolete and financially impaired. And this is landing on top of a commercial real estate sector already working through a debt maturity wall in the $1 trillion-plus range.

The bull case, also stated fairly, and it is the stronger one on the specific question of overbuild. The limiting factor on data center construction is not capital, it is power. Interconnection queues, transformer lead times, turbine backlogs, and siting difficulty make an unconstrained overbuild physically impossible. You cannot build the 2006 housing bubble's worth of excess capacity when the critical component has a 60-month lead time and the local county just voted you down. The same constraints that make this buildout so hard are the constraints that cap how badly it can be overdone. That is a genuinely important asymmetry and it is the best argument against the crash scenario.

Part IV: What Actually Gets Repriced

Strategy, not commentary. Here is the list, with the reasoning.

Repricing up

Land with executed interconnection. Already moving, still early outside the top eight markets. The spread between powered and unpowered land is the defining valuation fact of the decade, and in tertiary geographies it has not been discovered.

Obsolete heavy industrial with live electrical service. Closed smelters, mills, refineries, and fabrication plants. The worked example above produces a 24x on the dirt. There are dozens of these sites and most are owned by industrial companies with no idea what the switchyard is worth. This is the single best identifiable arbitrage on the board and it has a closing window, probably 24 to 36 months, before it is fully discovered.

Grid-adjacent rural land in ERCOT, PJM, MISO, and the Southeast. Not because the land is good but because the transmission is there and the politics are, for now, more permissive than in the coastal markets.

Substation-adjacent industrial and flex, generally. Even for non-data-center uses, because the power-constrained tenant list is growing: electrification of fleets, industrial reshoring, cold storage, vertical farming, and manufacturing all now compete for the same interconnection.

Skilled trades businesses. Stated bluntly: the highest risk-adjusted return available to a real estate investor in this cycle may not be a building. It may be an electrical contractor. A firm with 200 licensed electricians and high-voltage certification in a data center market has pricing power that no landlord in America currently has, a backlog approaching a year, and it trades at a small multiple of earnings because it is classified as a services business rather than as the scarce infrastructure it has become.

Workforce housing in data center counties, on a specific and time-limited basis. Covered below, because the trade has a back half that most people are not underwriting.

Repricing down

Land in a data center market without a power path. This is the mirror image and it is worse than neutral. When adjacent parcels trade at power-driven prices, the unpowered parcel gets marked to a comp set it cannot access, and the owner spends three years discovering that the bid was never for them.

Middle-tier office, permanently. Not the good stuff and not the conversion candidates. The 1985-vintage, four-story, suburban, surface-parked, structurally sound and architecturally unremarkable building in a market with flat office employment. In a zero-growth demand regime, this asset never has a good year again.

Multifamily deliveries in 2027 and 2028, and therefore rents in those years. This is the most actionable and most consensus-adjacent call here. Starts collapsed in 2025 and 2026. Deliveries follow starts by 18 to 24 months. Multifamily deliveries are forecast down 28 percent in 2026 and starts down again in 2027. The supply cliff arrives in 2028 into a market where household formation did not stop. Rent growth in 2028 to 2030 will surprise to the upside in exactly the markets where 2024 and 2025 oversupply is currently causing pain. The developer who can start a project in 2026 and 2027, against every instinct and every capital market signal, delivers into the tightest market of the decade.

The trade with a back half: construction workforce housing

This one is worth working, because it is the most predictable rent spike and the most predictable crash in the country right now, and both halves are visible from here.

A 1 GW campus carries a peak construction workforce of 3,000 to 5,000 for roughly three years. Call it 4,000. In the rural and exurban counties where these campuses site, perhaps 60 percent of that workforce comes from outside commuting range, so 2,400 workers need housing. At an average of 1.5 workers per unit, given that trades crews double up, that is roughly 1,600 units of new demand.

A county of 30,000 people has on the order of 1,200 to 1,500 total rental units, running at 92 to 94 percent occupancy, which means roughly 100 units of available inventory.

Sixteen hundred units of demand against one hundred units of supply is not a 10 percent rent event. It is a 50 to 70 percent rent event, plus hotels running at rates that make no sense, plus RV parks, plus man-camps, plus a housing affordability crisis for the schoolteachers and nurses who were already there. That part is happening now, in something like forty counties, and it is being underwritten enthusiastically.

The back half is the part that is not being underwritten. The construction workforce leaves. A 1 GW campus that employed 4,000 people to build it employs perhaps 50 to 150 people to operate it. That is a 96 percent reduction in local employment on a known date, roughly 36 months from groundbreaking, published in the project schedule.

So the honest underwriting of a workforce housing deal in a data center county is a three-year hold with spectacular yield and a terminal value that has to assume something close to the pre-project rent level, discounted for the fact that you and four other developers all added supply into the spike. Build it, harvest the spike, exit before energization. Underwrite the exit cap at the county's pre-boom fundamentals, not at trailing NOI. Anyone underwriting a ten-year hold on trailing rents in these counties is going to learn a very old lesson about company towns.

Part V: Three Scenarios, Dated

Everything above is roughly knowable. What follows is not. It is labeled as scenarios with probabilities so that it can be scored later, which is the only thing that makes a forecast worth writing down.

Base case: The Grind. 2027 to 2030. Roughly 55 percent.

The buildout continues but delivers at 60 to 70 percent of announced pace. Turbines, transformers, and moratoria throttle it. Announced-to-energized gigawatt ratios stay poor and the gap becomes a running industry scandal by 2028. Data center load reaches 8 to 10 percent of US electricity by 2030 rather than the 12 percent-plus the aggressive forecasts carry.

Real estate consequences: powered land stays scarce and expensive through the period. Electrician wages plateau at high levels by 2029 as the apprenticeship pipeline finally delivers. Multifamily starts trough in 2026 and 2027, producing a genuine supply shortage and above-trend rent growth from 2028 through 2030. Office stays flat and bifurcated. Two or three states establish themselves as the permissive jurisdictions and capture a disproportionate share of the investment, and their neighbors spend the 2030s writing reports about why they lost it.

Bull case: The Buildout Holds. Roughly 20 percent.

AI demand validates commercially, behind-the-meter generation scales faster than expected, and federal action preempts or overrides the worst of the local opposition. The nuclear pipeline delivers: the Three Mile Island Unit 1 restart energizes in the second half of 2027 as accelerated, the roughly 9.8 GW of announced nuclear commitments to AI infrastructure largely materializes, and the first SMR deployments land in the early 2030s rather than never.

Real estate consequences: power becomes the permanent organizing principle of American land value, full stop. Location is repriced as a function of interconnection for a generation. The industry reorganizes around it, with a dedicated power capability inside every major developer by 2029. Housing construction stays labor-constrained and expensive until roughly 2031, at which point the largest cohort of new tradespeople in fifty years arrives and homebuilding productivity improves structurally.

Bear case: The Air Pocket. 2027 to 2028. Roughly 25 percent.

A demand disappointment, a model efficiency breakthrough that collapses inference cost per unit of capability, or a credit event at a leveraged neocloud. Capex guidance gets cut. Half-built campuses stop.

The interesting part of this scenario is that the stranded asset is not primarily the data center. Hyperscaler-owned campuses with investment-grade balance sheets behind them get finished slowly or mothballed cheaply. The stranded assets are the towns: the county that rezoned 2,000 acres and issued the bonds, the utility that built the substation and the transmission spur into a rate base with no load behind it, the developer who delivered 800 apartments for construction workers who are now in another state, the school district that hired against a projected tax base. The 7.3 percent rate increases stay, because the infrastructure was built and someone pays for it, and now there is no data center to blame or to tax.

For real estate specifically: powered land reprices down hard but not to zero, because the underlying scarcity is real even at half the demand. Workforce housing in the forty counties gets crushed. Electricians return to the general construction market, which is actively good for housing. And the multifamily supply cliff still arrives on schedule in 2028, because those starts were already not happening, which makes this the one scenario where a housing developer wins on both sides.

What to watch, quarterly

Falsifiable indicators, in rough order of information value:

1. Announced GW versus energized GW. The gap is the whole story and almost nobody publishes the ratio.
2. Interconnection queue withdrawal rates, not queue size. Queue size measures enthusiasm. Withdrawals measure reality.
3. Turbine and transformer slot pricing and lead times at GE Vernova, Mitsubishi, and Siemens. This is the physical clock.
4. Hyperscaler capex guidance revisions. Not the level, the revision. The first down-revision by a major is the bear case's opening bell.
5. State moratorium count and the polling trend. The Wisconsin poll moved 15 points in six months. Watch the second derivative.
6. Residential electricity rate CPI by state, specifically in Virginia, Ohio, Georgia, and Texas. This is the political fuel gauge.
7. Electrician wage prints and apprenticeship completions in the top ten data center counties.
8. Multifamily permits in data center counties versus national. If the crowd-out thesis is right, the gap widens through 2027.

Part VI: What Would Prove This Wrong

Every essay in this genre projects a trendline and then treats the projection as a fact. The honest version states its own kill conditions, so here are mine.

Efficiency outruns demand. The cost per unit of AI capability has fallen by orders of magnitude repeatedly, and there is no law requiring compute demand to grow faster than compute efficiency. If algorithmic and hardware efficiency gains outpace demand growth for two consecutive years, the load curve flattens and most of this article is about a two-year event rather than a decade. This is the most likely way the thesis fails and it is the one I hold with the least confidence.

The load goes offshore. Training in particular is not location-sensitive. If the marginal training cluster moves to the Gulf states, the Nordics, or anywhere with abundant stranded power and permissive politics, the American land consequence shrinks dramatically even if global AI growth continues exactly as forecast. Inference stays domestic because latency matters, but inference is a fraction of the load and a fraction of the power density.

Behind-the-meter generation decouples the whole problem. If on-site generation at scale becomes standard, data centers stop competing for interconnection and stop showing up on residential bills. That removes the ratepayer politics and the queue constraint simultaneously, which would collapse the powered-land premium and neutralize the moratorium risk in one move. It would also make the labor crowd-out worse, not better.

Load growth turns out to be a definitional artifact. The forecast spread here is genuinely embarrassing. Goldman at 41 GW and 451 Research at 75.8 GW for the same year is not a rounding difference, it is a disagreement about what is being counted. If the truth is near the low end and the high-end forecasts are double-counting announced-but-never-built capacity, the scale of everything above shrinks by half.

The political reaction wins outright. 71 percent opposition is a very large number. If the moratorium count keeps compounding and a few states make it permanent, the buildout relocates to five permissive states and becomes a regional rather than national real estate phenomenon.

I hold Parts I and II with high confidence, because they describe measured conditions rather than forecasts. I hold Part III at moderate confidence, with the labor crowd-out argument the one I would most want to see tested against better data. Part V is scenarios and should be treated as such.

Parting Thoughts

The industry's operating assumption is that this is a sector story. Data centers are an alternative property type, they belong in a specialist allocation, and the rest of real estate continues on its own logic. That framing is comfortable and it is wrong, because power is not a sector. It is an input to everything.

The apartment developer who cannot get 8 MW of service for a mid-rise in a data center county is in this story. The industrial developer whose tenant needs 15 MW for automated fulfillment is in this story. The homebuilder whose electrician left for $280,000 is in this story. The office owner underwriting a five-year lease into permanently flat office employment is in this story. The land banker holding 600 acres in Central Ohio has either the best trade of their career or a very expensive misunderstanding, depending entirely on a question about transmission capacity they have probably never asked.

What has actually changed is the identity of the scarce thing. Real estate value has always accrued to whoever controlled the constrained input. For a long time that was capital, which is why the banks won. Then it became entitlement, which is why the developers who mastered local politics won. It is now interconnected electrical capacity, and the people who will win are the ones who figure out, before it is obvious, that a substation is a real estate asset and a queue position is a development right.

The repricing has already started. It is happening in counties most of the industry could not find on a map, in transactions that do not clear through the channels the industry watches, at prices that would not make sense to anyone still valuing the dirt.

You can see it first in the queue. It is public, it is free, and almost nobody in this business has read it.