ARUON Insights

Four U.S. States Are Winning the Data Center Race - and Big Tech Now Outspends Most Nations' Militaries

rbernard@aruon.ai (Richard Bernard Jr) — Sat, 12 Sep 2026 22:14:30 GMT

A note on this one: an earlier version of this piece ran in August 2025. The buildout has moved fast enough since that most of the figures needed replacing. This is the 2026 update.

The AI revolution is reshaping the global economy, and at its core lies an unprecedented demand for data centers — the facilities housing the servers, GPUs and power infrastructure needed to train and deploy AI models.

U.S. data center construction starts reached roughly $78 billion in 2025, up about 190% since the launch of ChatGPT in late 2022. Data centers now account for a large and rising share of U.S. non-residential construction spend. The demand curve behind it is steeper still: hyperscalers have committed $300+ billion through 2028.

But this growth isn't uniform. A handful of states are capturing most of it — and the reasons are specific, structural, and worth understanding whether you're siting a facility or buying land near one.

The states winning, and why

Virginia: still the incumbent

Northern Virginia — "Data Center Alley" — remains the world's largest data center market, with more than 300 facilities and gigawatts more planned. Low energy costs, dense fiber, proximity to federal demand, and long-standing sales tax exemptions on equipment built an advantage that took two decades to accumulate.

The caveat matters as much as the headline: that density is now a constraint. Parts of Northern Virginia and several PJM zones have turned "power available" into a multi-year negotiation. Virginia wins on existing capacity and loses on new greenfield speed — which is why the growth is moving.

Texas: energy optionality and rural headroom

Texas is on pace to pass Virginia as the nation's #1 hub, with roughly 140 data centers under construction. ERCOT's scale, gas and renewable surplus, competitive retail options, and the ability to pursue behind-the-meter generation create optionality few other markets match. Dallas, Houston and Austin are the established hubs, but the real expansion is rural West, East and South Texas — cheaper land, less organized opposition, and property tax abatements through state economic development programs.

Arizona: cooling, solar and permitting speed

Phoenix is among the fastest-growing capacity markets in the country, built on abundant solar, a dry climate that favors efficient cooling, low-cost land, and the Qualified Facilities Tax Incentive Program. Its exposure is water — a line item that has moved from footnote to underwriting assumption across the Southwest.

Georgia, Tennessee and the Carolinas: the economic-development play

Atlanta's fiber density and low energy costs attract hyperscalers with sales tax exemptions. Tennessee's TVA ecosystem of local power companies offers faster local decision-making than many investor-owned utility rate cases. North Carolina has landed nearly $11 billion in life-sciences manufacturing investment in a single year, and the Carolinas are now pulling movers at a faster rate than Texas or Florida. These states combine competitive pricing with active economic-development machinery and secondary-county opportunities away from saturated metro nodes.

The number that reframes the stakes

Here is the comparison that puts the buildout in perspective.

In 2025, Amazon, Alphabet, Microsoft and Meta were projected to spend a combined $364 billion in capital expenditure, with well over half of it allocated to AI data centers and computing infrastructure. Set that against national defense budgets:

United States: roughly $997 billion
China: roughly $314 billion — less than four U.S. companies' combined capex
Russia: roughly $149 billion
Germany: roughly $89 billion
India: roughly $80 billion

Four American companies now out-invest the combined defense budgets of Russia, Germany and India. Global military expenditure reached $2.718 trillion in 2024, up 9.4% — and U.S. private-sector AI capital deployment still dwarfs the entire defense outlay of most nations.

That is not a curiosity. It is a statement about where strategic capability is being built, and how fast.

Why leadership here compounds

Economically, the buildout is driving well over $1 trillion in investment across a five-year horizon, concentrated in construction, engineering and skilled trades — and the follow-on wave of healthcare, logistics, schools and services that arrives with every permanent job.

Strategically, compute underwrites everything from cyber defense to autonomous systems. Capability built domestically is capability not dependent on a foreign supply chain — a lesson the semiconductor shortage taught at considerable expense.

Technically, leading the buildout means setting the efficiency and sustainability standards rather than inheriting them. Liquid cooling, grid-interactive load management and on-site generation are all being proven out in these facilities right now.

The part that doesn't make headlines

Every state ranking above is a first-pass filter, not an answer. A state average tells you nothing about whether a specific parcel can reach an existing circuit with credible headroom, whether zoning permits the use by right, or whether the fiber on the map is physically diverse. Those are parcel-level questions, and they decide deals.

That gap is what we built REIR to close — owner of record from the county's own roll, zoning district with the code section cited, substation distance and voltage, fiber carriers on route, flood, wetlands and soils, every figure traced to an official source and gaps flagged rather than guessed. Our coverage map is published live, so you can check your own market.

The states are winning. The sites still have to be earned.

Screening land in one of these markets? Check the coverage map or call the ARUON research desk at (833) 545-7058.

Power Is the New Real Estate - But Only If You Know Which Game You're Playing

rbernard@aruon.ai (Richard Bernard Jr) — Sat, 12 Sep 2026 21:57:33 GMT

In the AI data-center boom of 2025–2026, a quiet consensus has formed: power is the scarce resource. Land is relatively easy. Fiber can be engineered. But three-phase power at meaningful scale has become the binding constraint. Interconnection queues stretch years. Large power transformers sit at roughly 128 weeks average lead time. Generator step-up units push past 140 weeks. Utilities and developers routinely disagree on energization timelines by 18 months or more.

That story is true and incomplete.

Most of the public narrative is shaped by the hyperscale fight: 100 MW, 300 MW, even multi-gigawatt campuses that must enter regional transmission queues, trigger expensive network upgrades, and often fall under new large-load tariffs written to protect other ratepayers. Those projects face genuine multi-year paths in many markets. Virginia remains the densest existing cluster and also one of the most constrained for greenfield capacity. Parts of Northern Virginia, certain PJM zones, and a handful of mature hubs have turned "power available" into a multi-year negotiation.

Proximity and edge compute live in a different regime.

Facilities in the 0.5–20 MW range — especially the 1–5 MW sweet spot common for regional inference, content, or low-latency enterprise workloads — typically connect at distribution or sub-transmission voltages. They often sit underneath the tariff thresholds that states and utilities have written for the biggest loads: 25 MW here, 50 MW or 75 MW there, 100 MW in several Southern and Western utility territories. When the fight is over gigawatt-scale customers, a 5 MW or 12 MW edge campus can still move on a standard commercial process — provided the local feeder or substation actually has headroom.

This is the structural opportunity the market has not fully priced.

What the data shows when you screen for proximity rather than hyperscale

Across independent assessments focused on smaller three-phase loads, a consistent set of states rises to the top when you weight industrial power price, generation surplus or rapid expansion potential, utility willingness to serve, and practical connection speed — rather than pure existing density:

Texas repeatedly leads. ERCOT's scale, gas and renewable surplus, competitive retail options in many areas, and the ability to pursue behind-the-meter generation create optionality that few other markets match. Node congestion is real and the large-load queue has drawn regulatory scrutiny, yet distribution-level and sub-75 MW projects can still find workable paths — especially outside the most contested East Texas corridors.

Oklahoma, Louisiana and several Entergy-territory markets follow closely on cost and industrial infrastructure, with industrial rates in the mid-to-high 5¢ to low 7¢/kWh range. Public-power and cooperative models in Nebraska, parts of Iowa, North Dakota and Wyoming offer both low rates and faster local decision-making than many investor-owned utility rate cases. Georgia, Tennessee (via the TVA ecosystem of local power companies), the Carolinas, Mississippi and Alabama combine competitive pricing with active economic-development machinery and secondary-county opportunities away from the most saturated metro nodes.

Ohio and Indiana appear higher on some lists because of population and fiber reach plus deliberate grid investment; they appear lower on pure cost rankings. The same pattern holds for parts of the Mountain West.

Cheapest average retail price does not automatically equal best proximity site. New Mexico's very low industrial average is real, yet water, fiber diversity and deliverability can dominate the economics. And Virginia's unmatched existing density is a liability for many new greenfield proximity projects precisely because the grid is already under pressure.

The practical ranking is therefore not a price list. It is a filter for places where a modest campus can reach an existing 12.47 / 13.2 / 24.9 / 34.5 kV circuit with credible headroom, secure written utility capacity confirmation, and still meet latency objectives to demand or cloud hubs.

The real critical path is rarely the RTO queue

For these smaller loads the binding constraints are more local and more mechanical:

Feeder and substation headroom at the specific parcel. A three-phase line at the frontage is not the same as available megawatts.
Transformer and switchgear procurement. Pad-mounted three-phase units remain among the tightest categories; new factory slots can stretch many months even when the utility study is relatively clean. Remanufactured inventory can compress that window dramatically when available.
Customer contribution, easement and construction responsibility matrices that only become clear after a formal load letter and utility response.
Fiber physical diversity and hazard exposure — flood, wind, wildfire, water stress — any of which can erase a power-price advantage.

Time-to-power planning allowances that treat a simple ≤1 MW service drop as 6–12 months, and a 5–20 MW feeder or modest substation job as 18–36 months, are more realistic underwriting assumptions than optimistic developer schedules.

A utility statement that "we can serve" without naming megawatts, voltage, point of delivery, upgrade scope, cost responsibility and target energization date is not an availability commitment.

The diligence standard has risen

State averages and even county-level lists are only first-pass filters. The sites that advance are those that survive a standardized utility package: load letter with ramp profile, one-line and voltage preference, site exhibits, request for firm available capacity at the parcel, upgrade cost and schedule ownership, and a written capacity letter suitable for land-acquisition conditions. Parallel work on diverse fiber routes and hazard screening is non-negotiable.

Behind-the-meter gas, temporary generation, or customer-funded substation work is increasingly used to de-risk schedules in Texas, parts of the Midwest, and certain public-power territories. These options are not free, but they can turn a multi-year grid path into a manageable critical path when the alternative is waiting for regional transmission upgrades.

What this means for the next 18–36 months

The hyperscale land-and-power war is creating a protected corridor for exactly the loads that proximity compute needs. Most of the new restrictive tariffs and study requirements were written for the biggest customers. Sub-threshold projects that treat power as a feeder- and parcel-level problem rather than a state ranking exercise can still move.

The developers and investors who capture that window are the ones who stop treating "power-friendly state" as a marketing label and start treating it as a set of verifiable engineering and commercial conditions. They will screen for utility territory first, issue identical load letters, level the responses on cost responsibility and schedule, and only then put non-refundable capital at risk.

That level of granularity is what separates a ranking table from an executable site.

It is also why we built REIR, the ARUON REI Researcher, and the Proximity Compute Scout service. Both are designed to move beyond state averages and marketing maps: map utility territories at the parcel level, surface the tariff thresholds that protect or constrain sub-20 MW loads, flag known headroom or congestion signals, and structure the diligence package so that written capacity positions — not assumptions — drive land control decisions.

Power is the new real estate. For proximity compute, the real estate that matters is measured in available megawatts on a specific feeder — not in acres, and not in statewide average cents per kilowatt-hour. The teams that treat it that way will move while others are still waiting in the wrong queue.

Screening a site for proximity compute? Check our coverage map, see the Proximity Compute Scout, or call the research desk at (833) 545-7058.

The Data Center Land War Is Ending Two Ways: Up Into Orbit, and Out to the Edge

rbernard@aruon.ai (Richard Bernard Jr) — Sat, 12 Sep 2026 21:57:30 GMT

Something broke in the American land market this year, and it wasn't subtle.

In Ashburn, Virginia, a data center developer started offering homeowners roughly $4.4 million per acre to assemble a 130-acre site — in a county where median land ran about $125,000 an acre a year earlier. The National Association of Home Builders is now publicly warning that AI data centers are outbidding homebuilders for land, poaching their electricians and driving up utility bills. Farmland that sold at $40,000 an acre commands $300,000 or more the moment it's labeled "data center ready."

Homebuilders can't compete, and it's not because they're slow. Their land price is capped by what a family can pay for a house. A hyperscaler's land price is capped by long-term compute demand — which right now looks like no cap at all.

But here's what most of the commentary misses: the gigawatt campus itself is getting squeezed too. From above and from below.

The squeeze from above: compute is leaving the planet

I wrote last fall about SpaceX's plan to scale Starlink V3 satellites into distributed orbital data centers — 1 Tbps downlink per satellite, 25 Gbps laser links, solar power around the clock, vacuum cooling for free. I said then that it could change how we handle AI workloads, edge computing and global connectivity.

Since then: SpaceX filed with the FCC for up to one million orbital data center satellites — a filing the Commission accepted in January 2026 — combined with xAI into a $1.25 trillion entity, revealed its first orbital data center node ("much simpler than a Starlink satellite," in Musk's words), and began pilot testing on-orbit compute on V3 hardware, with 60 satellites per Starship flight adding 60 Tbps of capacity at a time.

Think about what orbit actually solves: no land acquisition, no rezoning fight, no interconnection queue, no community opposition, no water for cooling, no $4.4 million acres. Every one of the terrestrial hyperscale bottlenecks — power, permitting and public consent — simply doesn't exist at 350 kilometers.

Orbit won't take every workload. But the massive, latency-tolerant AI training runs that justify gigawatt campuses? Those are exactly the workloads that can leave the ground first.

The squeeze from below: physics keeps the edge on Earth

What orbit can't solve is latency to the user standing on Main Street. Autonomous vehicles, real-time video analytics, industrial IoT, AI inference answering a customer's question — these need compute within 5 to 50 milliseconds of the person or machine generating the data. That math keeps them on the ground, close to population.

And the ground already has the perfect real estate for it: cell tower sites. Roughly 160,000 to 185,000 towers across the country, each one already holding grid power, backhaul fiber, an access road and a hardened, fenced compound. A modular edge facility of 50 kW to a few megawatts at a tower base doesn't need 300 acres, doesn't need a new substation, and doesn't trigger the community firestorm that's now canceling hyperscale projects across the country. American Tower has already identified over 1,000 of its own parcels suitable for multi-megawatt edge facilities and is building in Raleigh, Austin, Tampa, Oklahoma City and beyond.

So the picture for the next decade looks like this: training migrates toward orbit and remote gigawatt campuses; inference migrates to the edge; and the traditional mid-size metro data center gets squeezed between them. The winning terrestrial asset isn't the biggest building — it's the distributed parcel with power, fiber and zoning that already works.

The bottleneck nobody prices in: knowing which parcels qualify

Here's the hard part. Not every tower parcel converts to edge compute — most compounds carry kilowatt loads, and a viable edge facility needs a megawatt-plus, acreage for expansion, real fiber and cooperative zoning. And not every cheap rural parcel is a data center site — the cheapest land is usually cheap for a reason.

That screening problem is why we built REIR, the ARUON REI Researcher. Give it an address — or just a requirement, "data center use, five-plus acres, Dallas–Fort Worth" — and it returns an investor-grade site report: owner of record from the county's own roll, zoning district with the code section cited, by-right versus special permit, substation distance and voltage, fiber carriers on route, FAA and FCC filings, flood, wetlands, slope and soils.

Every figure is traced to an official source, gaps are flagged instead of guessed, and a human analyst signs off before anything ships. We score every U.S. ZIP code on market fundamentals, with parcel-level screening live across a growing set of states and counties — the same corridors where towers and edge capacity are going in. Our coverage map is published live and updates as we build, so you can check your own market rather than take my word for it.

In a market where the difference between a $40,000 acre and a $300,000 acre is one screening report, that's not overhead. That's the trade.

Where this leaves us

The land war headlines are real, but they're the closing chapter of the centralized era, not the opening chapter of the next one. Compute is splitting: up to orbit for scale, out to the edge for speed. The parcels that matter most for the next build cycle aren't the trophy megasites — they're the thousands of quiet, powered, fibered sites spread across the corridors where people actually live.

The players who can find and qualify those sites fastest will own the decade.

If your market is on our map — or you want it to be — check the coverage map, reach the ARUON research desk, or call (833) 545-7058.

The Map Moved. Most Developers Are Still Building on the Old One.

rbernard@aruon.ai (Richard Bernard Jr) — Sat, 12 Sep 2026 21:57:26 GMT

Somewhere on an interstate tonight, a family is following a 26-foot moving truck south. They aren't reading Census reports or cap-rate surveys. They're chasing something simpler: a better job, a bigger yard, a paycheck that stretches further. Multiply that one truck by hundreds of thousands, and you get the most powerful force in American real estate — and the most ignored.

Here's the number that stopped me cold: last year the South absorbed essentially all of America's net domestic migration, roughly 358,000 people, while the Northeast and West each lost about 200,000. Celina, Texas, a town north of Dallas, grew 24.6% in a single year, per Census Vintage 2025. Not a decade. A year.

So was the developer who bought land in Celina five years ago a genius? Or did they just refuse to argue with the moving trucks?

That's the whole thesis: it is easier to thrive when you follow population and industry. Development is hard everywhere, but a rising market forgives your mistakes and a flat one punishes your perfection. As an estimator, I'll tell you a secret — the same apartment building costs roughly the same to build in a shrinking metro as in a booming one. The concrete doesn't care. The demand on the other side of delivery is what makes or breaks the deal.

The map has moved

The 2025 Census estimates tell a clear story, and a more interesting one than the usual "everyone's moving to Austin" take. National population growth slowed to just 0.5%, the slowest since 2021, as international migration fell sharply. That makes domestic migration the cleanest signal we have of where Americans actually choose to live. And the signal is loud:

South Carolina is now the fastest-growing state (1.5%), with the Carolinas pulling in movers at a pace that has them beating Texas and Florida on migration rate.
Texas added more people than any state — roughly 391,000 residents in one year. Houston and Dallas–Fort Worth each absorbed a mid-sized city's worth of new neighbors.
Charlotte gained more residents than any city in America (+20,731). Not New York. Not L.A. Charlotte.
The growth isn't downtown — it's the outer ring. The five fastest-growing cities in the country are all Texas exurbs: Celina, Fulshear, Princeton, Melissa and Anna. Thirty to forty-five minutes out, where the land is.

The U-Haul Growth Index 2025 — the ultimate revealed-preference data — ranks Texas #1, followed by Florida, North Carolina, Tennessee and South Carolina (South Carolina held the top spot in 2024). People vote with a 26-foot truck, and the polls have closed.

Jobs go first. People follow. Housing compounds in the middle.

Population doesn't move in a vacuum. It follows paychecks — and the paychecks are being planted in the ground right now, in the form of the largest industrial construction wave of my lifetime:

AI and data centers: roughly $78 billion in U.S. data center construction starts in 2025, up 190% in one year, with hyperscalers committing $300+ billion through 2028. Texas alone has 140 data centers under construction and is on pace to pass Virginia as the nation's #1 hub.
Semiconductors: TSMC's $65 billion Phoenix campus. Samsung's $40 billion bet on Taylor, Texas. These aren't press releases anymore — they're job sites full of construction workers who all need somewhere to live tonight.
Life sciences and advanced manufacturing: North Carolina landed nearly $11 billion in life-sciences manufacturing investment in a single year. BMW's Spartanburg plant remains the nation's top automotive exporter by value, with an EV battery ecosystem now growing around it.

And then comes the follow-on wave: healthcare, logistics, schools, services — and construction itself. Economic multiplier data pegs every new manufacturing job at roughly 2.5 additional local service jobs, and every one of those workers is a renter before they're a buyer.

This is the flywheel: cheap land and business-friendly policy attract industry; industry creates jobs; jobs pull people; people need housing; housing demand creates the development opportunity. Population and industry chase each other. The developer just has to stand where they're both headed.

The timing kicker nobody's talking about

Here's where it gets interesting for anyone in commercial real estate. The apartment building boom peaked at just over 1 million units under construction in late 2023, and it bruised a lot of markets. Austin, Phoenix and Nashville added 20–30% to their apartment stock in a few short years. Rents went flat or negative. Plenty of smart people got humbled.

But look at what's happening now: multifamily starts have fallen more than 40% from the peak, and the pipeline has drained to roughly 690,000 units under construction — the lowest since the mid-2010s. Meanwhile absorption has started outpacing new deliveries again and vacancy is coming off its peak. Deliveries fall off a cliff in 2027–2028 in the exact markets people keep moving to. Whoever entitles and builds now delivers into a shortage.

And policy is quietly clearing the path. Florida's Live Local Act unlocks density and tax incentives for workforce housing. Texas now allows apartments on commercially zoned land without a rezoning fight. The states winning the population race are also making it easier to house the people they're winning.

Where I'd be looking

If I were placing chips on the map for the next 36 months, here's how I'd sort it:

Build into the dip: Dallas–Fort Worth (ULI's #1 real estate market two years running), Houston, Raleigh–Durham, Charlotte's secondary counties, Tampa, Orlando. Demand intact, pipeline collapsing. The window is open.
The exurban frontier: the Celina–Princeton–Anna ring north of DFW, west Houston, Fort Mill outside Charlotte. Hypergrowth off a small base, faster entitlements, land you can still afford.
Follow the plant, not the skyline: workforce housing near the new fabs, battery plants, ports and hospital campuses — Greenville–Spartanburg, Charleston, Huntsville, the Austin–Taylor corridor. These lease up before the amenity-heavy towers do.
Patience plays: Austin and Phoenix urban cores. Great decade ahead; painful next 18 months while the oversupply burns off.

The honest caveats

I negotiate for a living, so let me be straight with you, because credibility is the only currency that compounds. Sun Belt migration is decelerating from its pandemic peak — the trend is intact, but the slope is flatter. Insurance costs in Florida and the Gulf are a real underwriting line item, not a footnote; Florida coverage now averages north of $6,000 a year. Property taxes in Texas bite too: an effective rate near 1.8% against a roughly 1.1% national average. And a $10 billion data center does not create the same housing demand as a $10 billion factory — count the permanent jobs, not the press release.

Following growth reduces risk. It does not replace underwriting.

The bottom line

You can't control interest rates. You can't control construction costs — believe me, I price them every day. But you can control which direction the wind blows on your project. Fighting demographic gravity in a shrinking market takes brilliance and luck. Building where people and paychecks are already headed just takes discipline and a map.

The map is right there. The trucks are already rolling.

I'm pricing a 300-unit garden-style deal in the Celina–Princeton corridor right now. If you're screening land in a corridor like it, our team publishes a live map of where we hold county-level parcel records and zoning — see where ARUON covers today, or talk to the research desk.

Data sources: U.S. Census Bureau Vintage 2025 estimates; ULI/PwC Emerging Trends in Real Estate 2026; U-Haul Growth Index 2025; ConstructConnect (Feb. 2026); CBRE and NAA 2026 outlooks.