The Top Tier of the Asset Class Is Not a Data Center Anymore. The Language Has Not Caught Up.
In 1996, the United States hosted roughly 50,000 small and mid-sized precision machine shops feeding the defense industrial base. By 2024, fewer than 18,000 remained, with the median operator age above 58 and roughly 70 percent of remaining shops still using paper travelers and CNC controllers older than the engineers who run them. Over the same window, Hadrian raised more than $260 million to build a single autonomous precision factory ("Factory 3") sized at 270,000 square feet in Mesa, Arizona. Anduril's Arsenal-1 in Pickaway County, Ohio is permitted at five million square feet on roughly 500 acres, targeting tens of thousands of autonomous systems per year. Those two facts describe the same shift from different angles.
The asset class that the market still calls "data centers" has split in half. The bottom half is fading commodity colocation: enterprise lift-and-shift, retail interconnection, secondary-market wholesale. The top half is something the language has not caught up to. It is strategic industrial capacity: frontier model training campuses, sovereign compute enclaves, AI-coupled precision manufacturing, autonomous fabrication cells, and increasingly the physical defense industrial base itself, stacked on the same parcel under the same fence line. The reason the same developers keep showing up on both compute deals and defense-prime production deals is not coincidence. It is that the underlying site requirements have converged.
The Term "Data Center" Is About To Become Obsolete For The Top Tier
The phrase "data center" was coined in an era when the asset was a room. It survived through the era when the asset was a building. It is now actively misleading in the era where the asset is a 1,000-acre campus with a dedicated 500 kV substation, a 1.2 GW generation tie, a cleared workforce pipeline, an ITAR-compliant production wing, and a five-year backlog of customers who are not enterprise IT buyers.
Categorical language lags physical reality by roughly a decade in industrial real estate. "Industrial" used to mean steel mills. Then it meant logistics. Then it meant cold storage. Each transition was preceded by a period where the new use case was misclassified into the old vocabulary and therefore mispriced by capital markets. The same pattern is unfolding now. The top quartile of large-format powered sites is no longer being underwritten by colocation REIT comparables. It is being underwritten by buyers who already know the comparable set is closer to a defense industrial REIT crossed with a sovereign infrastructure fund. The vocabulary will follow the capital. It always does.
By 2028, the trade press will draw a hard line between two categories that today share a name. Commodity colo will remain "data center." The top tier will get a new label, likely some variant of "strategic compute campus," "sovereign industrial site," or simply "industrial AI campus." The relabeling itself is worth 100 to 200 basis points of cap rate compression for the assets that qualify, because the buyer universe expands from digital-infrastructure funds to the much larger pool of sovereign wealth, defense-aligned infrastructure capital, and pension allocators who have explicit mandates to fund onshore industrial reconstitution.
What Strategic Industrial Capacity Actually Hosts
The tenant stack on these campuses is not a list of cloud providers. It is a stack of functions that until recently could not occupy the same address.
The anchor is frontier model training. A single training cluster for a next-generation foundation model now consumes 300 to 800 MW of contiguous load, with individual training runs lasting 60 to 120 days and aggregate annual draw exceeding the consumption of mid-sized US cities. Two of the three labs with frontier-scale ambitions have publicly disclosed plans for gigawatt-class single-campus footprints by 2027. The second floor of the stack is sovereign compute: dedicated, jurisdictionally-segmented inference and training capacity for federal customers, allied governments, and federally-funded research consortia. Sovereign capacity carries different security, audit, and personnel-clearance requirements than commercial cloud, and it does not coexist with hyperscale multi-tenant operations on the same security domain.
The third floor is AI-coupled precision manufacturing. This is the Hadrian, Anduril, Saronic, Apex Space, and Castelion class of operator. Software-defined production cells, robotic CNC and additive cells, autonomous inspection and test stations, all networked to a control plane that schedules across machines the way Kubernetes schedules across servers. These cells produce flight-critical hardware for missile programs, autonomous surface vessels, satellites, and unmanned aircraft. They require ITAR-compliant facility security, cleared personnel, redundant power at industrial-process tolerances, and the ability to receive raw stock and ship finished assemblies under chain-of-custody control.
The fourth floor is cleared production for traditional primes. Lockheed, RTX, Northrop, General Dynamics, and Boeing Defense have all publicly committed to reshoring components of their supply chain that were offshored between 1995 and 2015. The capacity to do that reshoring at scale does not exist in the brownfield industrial parks that hosted the original supply chain. It is being built on greenfield sites that were originally optioned for compute.
The same parcel that hosts a 500 MW GPU cluster on its east half will host a 200,000 to 400,000 square foot autonomous machine shop on its west half within the same security perimeter, sharing substation, fiber, water treatment, and cleared workforce housing. That is the operating model. It is already being built.
Why The Site Profile Converges
A frontier training campus and an autonomous defense production cell appear to have nothing in common until you list what each one actually requires.
Both require contiguous power at industrial scale, with redundant feeds and on-site firm generation. Frontier training pulls 500 MW to 1.2 GW. A fully built-out autonomous production complex of the Arsenal-1 class pulls 80 to 200 MW of process load with similar redundancy requirements. The grid interconnection studies, transmission upgrades, and generation procurement are the same exercise. Both require fiber redundancy at carrier-grade diversity. Both require water for cooling and process. Both require federal-grade physical security: anti-vehicle perimeter, mantraps, badge-and-biometric access control, continuous monitoring. Both increasingly require a cleared workforce, which in turn requires proximity to a population center large enough to recruit from but with site-control discipline tight enough to vet.
Then come the variables that traditional colocation underwriters do not model. Proximity to a deepwater or inland port matters for industrial tenants because finished assemblies and inbound stock move on flatbed and container, not on fiber. Proximity to a DoD prime's existing footprint matters because cleared production is networked production, and the second-tier suppliers must be within drive time of the integrator. Jurisdictional posture matters because counties and states differ by an order of magnitude in their willingness to permit ITAR facilities, host federal security review, and execute community benefit agreements that survive a change of board. CFIUS-cleanliness of the cap stack matters because the customer will not sign without it.
The result is that the site selection map for top-tier strategic industrial capacity is much narrower than the map for commodity colocation. There may be 200 to 400 US sites that can credibly host 500 MW of compute. There are fewer than 60 that can credibly host that compute plus an ITAR production wing plus a cleared workforce plus the jurisdictional and CFIUS posture to close a sovereign or defense-prime anchor. That scarcity is not in the published comp set yet. It will be.
The Hadrian Signal
Hadrian is not interesting because it is a machine shop. It is interesting because it is the first widely-capitalized expression of a category: factory-as-a-service, where the production capacity is the product and the customer rents output rather than buying hardware. Hadrian's Factory 3, at roughly 270,000 square feet, is sized to absorb the throughput of dozens of legacy precision shops with a fraction of the headcount. The economic claim is that an autonomous cell running 22 hours a day with two-shift remote supervision produces parts at 30 to 60 percent lower fully-loaded cost than a traditional shop running 8 to 10 hours with full machinist staffing, and does so with 5 to 10 times the throughput per square foot.
Whether Hadrian specifically captures the category is not the relevant question. The category exists. Anduril Arsenal-1, Saronic's Port of Franklin facility, Apex Space's bus production line, and Castelion's missile production footprint are all expressions of the same underlying shift toward software-defined industrial cells operating at hyperscale-comparable capital intensity. The capital is not flowing to legacy contract manufacturers. It is flowing to greenfield, software-native, autonomous production at sites that were unimaginable in this use case five years ago.
For the developer of strategic industrial capacity, the Hadrian signal means the demand stack is durable. Compute alone could be over-supplied by 2029 in a way that compresses returns. Compute plus cleared autonomous production plus sovereign capacity cannot be over-supplied on any visible horizon, because the gating constraint is not capital and not even power. It is sited, secured, cleared, and jurisdictionally-permitted real estate. That constraint takes a decade to build and cannot be rapidly arbitraged by incumbents.
How The Capital Structure Changes
The 2015-vintage colocation REIT was structured for a tenant base of enterprises and cloud providers, a financing base of public bond markets and IG-rated unsecured debt, and an equity base of REIT-dedicated public investors. The strategic industrial capacity vehicle of 2028 will be structured differently in every layer.
The tenant base will include sovereign customers, defense primes, federally-funded research consortia, and AI labs whose contractual posture looks closer to a long-cycle defense supplier than a colocation customer. Take-or-pay structures will lengthen from 7 to 10 years toward 15 to 25 years. Inflation-linked escalators will be standard. Termination-for-convenience clauses will be replaced by milestone-based step-ups tied to facility delivery and clearance attainment.
The financing base will shift toward project finance with sovereign-friendly LP participation, federally-guaranteed debt where the tenant has a federal nexus, and infrastructure bond issuance under dedicated digital-infrastructure or strategic-asset sub-indices that the rating agencies are already drafting frameworks for. Tenor will extend from 7 to 10 years toward 20 to 30 years. Spreads on the highest-quality strategic industrial paper will compress 80 to 150 basis points below comparable commodity colocation paper within 24 to 36 months of the first dedicated sub-index launch.
The equity base will look more like a defense industrial REIT crossed with a sovereign infrastructure fund than a colocation REIT. Allocators with explicit mandates to fund onshore industrial reconstitution, allied sovereign wealth, defense-aligned strategic capital, and large pensions with industrial-policy alignment will dominate the LP base. The pricing implication is straightforward: required returns drop, hold periods extend, and the buyer universe widens.
The underwriting checklist will add three variables that did not exist in the 2015 colo playbook. CFIUS posture of the cap stack will be diligenced upfront, not at exit. FOCI mitigation plans will be standard documentation for any campus hosting cleared production or sovereign workloads. Federal security review timing will sit on the same gantt chart as utility interconnection and environmental review. Sites that cannot pass all three become uninvestable for the top tier of the buyer pool. Sites that can pass all three trade at premiums that the current comp set does not capture.
Closing
By 2030 the largest industrial campuses in the United States will not be classified as data centers in any document that matters. They will be classified as strategic industrial assets, underwritten by capital that does not exist on today's colocation cap tables, and tenanted by customers whose names do not appear in the current REIT disclosure. The developers who are still building to a 2015 colo spec at that point will be selling into a commoditized bottom tier. The developers who recognized the split early and built to the converged site profile, ITAR-grade security, cleared workforce, sovereign-friendly cap stack, and contiguous industrial-process power, will own the only asset class in American real estate that is simultaneously scarce, federally important, and structurally underpriced. The relabeling is the smallest part of what is about to happen.
Autonomous Industries develops powered, permitted industrial campuses engineered for the strategic industrial capacity stack. For institutional inquiries, contact us at conatus.com.