The unit of AI infrastructure is no longer the building. It is the campus.
A single NVIDIA GB200 NVL72 rack pulls 120 kW. Forty of them fill a row. A training run for a frontier model now demands 100 to 200 MW of flat, sustained, uninterruptible load, sometimes more, for months at a time. The 2015 colocation industry was built around a 1 to 2 MW per acre planning assumption. The 2026 hyperscale build is sized at 5 to 10 MW per acre. The AI campus coming online in 2028 and beyond will sit at 15 to 25 MW per acre, an order of magnitude denser than the facility class it is replacing. That density progression is not a refinement of the old form factor. It is the obsolescence of it.
The industry is still pricing AI infrastructure as if the unit of competition were a 30 to 100 MW single-tenant box on 20 acres. It is not. The unit is the campus: 500 to 3,000 contiguous acres, 500 MW to multi-gigawatt aggregate load, behind-the-meter or on-site generation, dedicated 230 kV or 500 kV interconnection, water rights measured in millions of gallons per day, diverse long-haul fiber, owned substation infrastructure, controlled security perimeter, and increasingly an anchor tenant from the sovereign or defense complex. The campus is to the building what the building was to the rack. Everything downstream of that recognition, including land pricing, broker assumptions, utility planning, county entitlement timelines, gets repriced.
The Six-Stack Site Requirement (the adjacency math nobody quantifies)
A campus-grade site is not scarce because of one constraint. It is scarce because six independent constraints must co-locate on the same parcel. Each constraint, taken alone, narrows the US map. Stacked, they collapse it.
The six are: large contiguous acreage (500 to 3,000 acres, single ownership or assemblable), 500 kV-class transmission within economic reach (typically within 5 to 15 miles, with available capacity confirmed by an interconnection study rather than hoped for), industrial-volume water access (5 to 50 million gallons per day of permitted withdrawal or recycled supply, with downstream discharge rights), gas pipeline lateral or mainline capacity sufficient for behind-the-meter generation (200 MMcf/day class for a multi-gigawatt campus), a jurisdiction with workable zoning, CEQA or NEPA pathway, and political appetite for industrial use at scale, and finally a labor shed deep enough to staff construction surges of 3,000 to 8,000 workers and long-term operations of several hundred technical roles.
Any one of these knocks out 95 percent of the US land mass. Stacked, they reduce the candidate set to the low hundreds. Run the filter honestly across PJM, MISO, ERCOT, and the Western Interconnection, and the count of sites where all six are simultaneously available, in 2026, with realistic optioning, sits between roughly 150 and 250. By 2028, after the current optioning wave from the named hyperscalers and a handful of sovereign and defense-adjacent buyers, fewer than 50 of those will remain unencumbered at scale.
The conventional brokerage map treats power as a constraint and treats the other five as solvable downstream. That assumption was correct in 2015. It is incorrect in 2026. Water is the binding constraint in much of ERCOT and the desert Southwest. Transmission queue position is the binding constraint in PJM. Jurisdiction is the binding constraint across most of Virginia, the Phoenix corridor, and increasingly Atlanta. Gas lateral access is the binding constraint anywhere a campus wants behind-the-meter generation in under 36 months. The bottleneck is not power. The bottleneck is six-way adjacency.
What The Campus Replaces (the colo era, why it ends)
The colocation era was an economic equilibrium between two facts: network density was scarce, and per-rack power was modest. Equinix and Digital Realty solved network density. The standard product was a 10 to 50 MW facility, 8 to 15 kW per rack, sited near major fiber intersections, in jurisdictions that did not yet have an opinion about industrial load. The unit economics worked because retail colocation customers paid for cross-connects, not megawatts.
Hyperscale broke that equilibrium by inverting it. Microsoft, Google, Meta, and AWS pushed past 100 MW per site, drove rack density to 15 to 25 kW, and moved siting decisions from network proximity to power cost. Central Oregon, northern Iowa, Loudoun County, and central Ohio were the geographic expression of that shift. The single-tenant warehouse, sized for cost per watt, became the standard.
AI breaks hyperscale by inverting it again. Rack density is now 60 to 120 kW and rising. Training requires sustained load profiles that resemble industrial smelting more than colocation. Existing hyperscale shells cannot be retrofitted past 25 kW per rack without gutting the power and cooling spine. The cost of retrofitting a 100 MW shell to AI density approaches the cost of building from scratch. The asset that was state-of-the-art in 2020 is, by 2027, undersized and overcapitalized at the same time.
The political defect compounds the technical one. A 200 MW single-tenant warehouse employing forty people, drawing visible water and visible power, contributing modest property tax, generates moratoria. Loudoun County imposed restrictions. Chesterfield County tightened theirs. Prince William County followed. The pattern is durable because it is rational. Communities are running arithmetic, not politics. A facility that consumes resources visibly and contributes invisibly is a stranded political asset before it is a stranded technical asset.
The campus is the form factor that solves the technical density problem and the political durability problem at the same time. That is why it will replace the standalone facility, not augment it.
The Compute Density Curve (1-2 to 5-10 to 15-25 MW per acre)
The density progression is the single most underdiscussed number in infrastructure. The 2015 colocation standard placed roughly 1 to 2 MW of IT load per acre of developed site, inclusive of building footprint, parking, substation, cooling yard, and setback. The 2020 to 2026 hyperscale wave pushed that to 5 to 10 MW per acre, driven by taller buildings, denser racks, and consolidated cooling. The AI campus class, designed natively for 60 to 120 kW racks, liquid cooling at scale, and shared district infrastructure, will operate at 15 to 25 MW per acre when fully built out.
That is a 10 to 20x density increase against the colo baseline in roughly a decade. It also reframes the land question. A 1 GW AI campus at 20 MW per acre needs 50 acres of densely built compute, not 500. The remaining 950 to 2,950 acres on a campus-grade parcel are not wasted. They are the buffer that absorbs substation expansion, on-site generation, water treatment, hydrogen or battery storage, future phase optionality, security perimeter, logistics staging, and the multi-use industrial footprint that anchors the political constituency. The acreage is not for the compute. It is for the everything-around-the-compute that makes a gigawatt of compute physically and politically deliverable.
Brokers and underwriters still price land per acre against an implied MW per acre assumption that is at least one generation stale. That mispricing is the asymmetry the next wave of campus developers will compound against.
Anchor Deals Already Reshaping The Map (Meta-Oklo, AWS-Talen, MSFT-3MI, AEP large-load filings)
The named hyperscaler campus moves of the past 18 months are not a series of disconnected procurement announcements. They are the early visible expression of the same underlying recognition: the unit is the campus, the constraint is the six-stack, and the surviving sites are being optioned now.
Meta's 1.2 GW Oklo small modular reactor agreement in Ohio is a behind-the-meter generation play that bypasses the PJM interconnection queue and the AEP transmission constraint at the same time. AWS's 1.92 GW direct interconnect with Talen Energy at the Susquehanna nuclear facility is a colocated baseload play that solves transmission and generation in one transaction. Microsoft's 837 MW Three Mile Island restart with Constellation is a generation revival play that converts a stranded nuclear asset into a campus-grade power supply. AEP, Dominion, Duke, Entergy, and Oncor have collectively disclosed mega-load PPA filings and pending large-load tariffs that, taken together, imply tens of gigawatts of contracted AI-class load over the next five years.
Each of these is a campus transaction in substance even when it is described in press releases as a power deal or a tenant deal. The buyer is securing the integrated six-stack: power, generation, transmission, water proximity, jurisdiction, and acreage. The seller, whether a utility, a nuclear operator, or a landowner, is providing one or two stack elements that the buyer combines with the rest. The deals look heterogeneous from outside. From the inside, they are the same deal.
The implication for the remaining inventory is straightforward. Every named hyperscaler announcement subtracts a campus-grade site from the candidate pool. The optioning wave that began in 2023 will continue accelerating through 2027. Forward-looking buyers who have not already secured campus-grade ground are bidding into a market with a closing window.
The Scarcity Window (how many campus-grade parcels remain in 2026, in 2028)
The honest count of US sites that satisfy the six-stack adjacency requirement, in 2026, with realistic optioning availability, sits in the low hundreds. PJM contributes perhaps 40 to 60 candidates concentrated in central and western Pennsylvania, eastern Ohio, and southern Virginia. MISO contributes 30 to 50 candidates across Indiana, southern Illinois, and central Iowa. ERCOT contributes 50 to 80 candidates concentrated in the Permian, west Texas, and a narrow ring outside the Dallas and Houston load pockets. The Western Interconnection contributes 30 to 50 candidates in Wyoming, eastern Montana, parts of New Mexico, and the Pacific Northwest. The total ceiling, generously counted, is in the range of 200 to 250 sites.
By 2028, after another 24 to 30 months of optioning by the hyperscalers, the sovereign buyers, and the small set of independent campus developers, the count of campus-grade sites that remain unencumbered at scale falls below 50. By 2030, it falls below 25. That residual inventory is what the top 25 AI campuses by power will occupy.
Forward projections that follow from this scarcity:
By 2027, the median transaction price for a fully entitled, power-confirmed, water-secured campus-grade parcel exceeds $500,000 per acre on the strongest sites, with outlier transactions clearing $1 million per acre where transmission and water are already locked. Brokers still pricing against agricultural or light-industrial comps will be replaced by structured transaction teams pricing against megawatt-equivalent value.
By 2029, behind-the-meter generation, primarily gas-fired in the near term and small modular nuclear in the mid term, will provide more than 40 percent of incremental AI campus power, because the front-of-meter interconnection queue cannot clear fast enough to serve the load. Utility planning will reorient around campus-scale customers rather than aggregated residential and commercial load growth.
By 2030, the top 25 AI campuses ranked by aggregate power will host more total compute than the top 250 legacy colocation facilities combined. The colo era will not end with a public collapse. It will end with the slow migration of training and frontier inference workloads to the campus class, leaving the legacy facilities to serve latency-sensitive enterprise and edge workloads at progressively thinner margins.
By 2032, campus-grade ground will trade more like permitted spectrum than like industrial real estate. The asset is not the dirt. The asset is the unrepeatable adjacency of six independently scarce inputs on a single parcel, with the legal and political work already complete.
The builders who priced this correctly in 2025 and 2026 will own the physical layer of the AI economy by the end of the decade. The ones who priced it as data center real estate will be looking at a market that no longer exists as they understood it.
Autonomous Industries is building in the AI industrial campus category.