In January 2025, OpenAI announced Stargate with a stated ambition to bring its first Abilene, Texas campus from ground to power in roughly eighteen months. The legacy industrial benchmark for a hyperscale campus of comparable scale is thirty-six to sixty months, and many active programs run longer. Eighteen months is not a marginal improvement on a sixty-month process. It is a different process.
This is the same shape of compression the defense industrial base ran from roughly 2018 to the present. A new class of software-defined autonomous primes (Anduril, Saronic, Shield AI, Apex, Castelion, Hadrian) collapsed the cost and timeline of producing defense systems by an order of magnitude against the legacy primes (Lockheed Martin, Raytheon, General Dynamics, Northrop Grumman, Boeing Defense). The transition is not finished, but the structural direction is settled. Hyperscale physical infrastructure is on the same trajectory and roughly two to four years behind. The legacy infrastructure development stack (Bechtel, Burns & McDonnell, Black & Veatch, Fluor, Sargent & Lundy, Quanta Services on the construction side) is the Lockheed of this transition. The AI-native developers operating at the edges are the Anduril.
The argument of this paper is structural, not speculative. The compression has already begun. By 2028 it will be the dominant pattern for new-build announcements above 100 megawatts.
What Lockheed Actually Is
The legacy prime model is not defined by the products it builds. It is defined by the development process those products force. Cost-plus contracting is the financial signature, but the operational signature is sequential phase-gate program management, deep reliance on tier-one engineering consultancies, multi-decade program timelines, and an institutional posture that treats schedule and budget overruns as normal rather than anomalous. An F-35 program runs across decades. A Virginia-class submarine block takes years per hull. A Columbia-class boat is a fifteen-year endeavor. A new munition often requires a five to seven year qualification cycle before first article delivery.
The consultants and the primes coordinate the work in serial. Requirements precede design. Design precedes prototyping. Prototyping precedes qualification. Qualification precedes low-rate production. Low-rate production precedes full-rate production. Each phase has its own paper trail, its own review board, its own external integrator. The institutional weight of CMMC, ITAR, AS9100, DCMA audit lineage, and the security-cleared contractor base creates a moat that is real, but the moat also enforces the slowness. The product of the legacy prime model is not the missile or the airframe. It is the program.
This description maps directly onto legacy infrastructure development. The owner-developer hires an owner's engineer. The owner's engineer hires a permitting consultancy, an environmental consultancy, a geotechnical firm, a transmission planning consultancy, a substation EPC firm, a civil EPC firm, a structural EPC firm. Each of these reports up. Each waits on the prior phase. The same firms reappear on every project because the qualification cycle for new entrants is brutal. The product of legacy infrastructure development is not the campus. It is the program.
What Anduril, Saronic, Shield AI, And Hadrian Compressed
Anduril's Arsenal-1 plant outside Columbus, Ohio is roughly five million square feet on roughly five hundred acres, drew an Ohio incentive package reported at around nine hundred and ten million dollars, broke ground in 2024, and targets initial production in the 2026 to 2027 window. The site is designed for tens of thousands of autonomous systems per year. The legacy comp for a defense production facility of that scale and throughput is a decade-long capital program. Arsenal-1 is closer to a thirty-month buildout.
The unit-cost story is sharper. Roadrunner, Anduril's reusable interceptor, has been publicly discussed in the high six figures per round against a legacy interceptor stack measured in low to mid seven figures and in some configurations higher. Saronic Technologies' autonomous surface vessel program produces hull cadence on a quarterly rhythm that the legacy shipbuilding base cannot match outside of a wartime mobilization. Shield AI ships V-BAT airframes against a unit economic profile no legacy ISR platform approaches. Skydio produces enterprise-grade autonomous airframes at consumer-electronic unit costs. Apex builds satellite buses on standardized platforms with order-of-magnitude lead time compression against legacy bus integrators.
Hadrian closed a Series C of approximately two hundred and sixty million dollars in January 2025 and stood up Factory 3, a roughly two hundred and ninety thousand square foot facility in Mesa, Arizona, with a stated expansion into shipbuilding and naval defense components. Hadrian's core claim is a factory-as-a-service model in which a software-defined precision machining floor replaces the qualified-supplier-network procurement model that the legacy primes depend on. The point is not that Hadrian will replace Lockheed. The point is that a defense prime that previously had to coordinate hundreds of qualified machine shops can now coordinate a single software interface and receive parts at compressed lead times and standardized tolerances.
The pattern across all of these companies is identical. Software-defined production. Vertical integration of historically fragmented supplier bases. Standardized platforms that absorb mission variance in software rather than in hardware respins. Autonomous and semi-autonomous operations replacing labor-bound bottlenecks. Capital efficiency that puts unit cost below the legacy alternative by a factor that varies between three and twenty depending on the category.
The Legacy Infrastructure Development Stack
Four to six engineering firms drive the bulk of US hyperscale and energy infrastructure work. Bechtel, founded in 1898, runs nuclear, petrochemical, defense, and large civil programs. Burns & McDonnell, founded in 1898, runs transmission, substation, and large utility EPC programs. Black & Veatch, founded in 1915, runs water, power, and telecom infrastructure. Fluor, founded in 1912, runs petrochemical and large industrial work. Sargent & Lundy, founded in 1891, runs nuclear and power generation engineering. Quanta Services, the largest US electrical infrastructure construction firm by revenue, operates as the union construction arm that builds what the engineering firms design.
Reported backlog at these firms has run between twelve and thirty-six months of forward capacity through the 2023 to 2025 hyperscale boom, with substation transformer and high-voltage circuit breaker lead times pushing past three years in some configurations. The engineering capacity, not the construction capacity, is the binding constraint on US data center buildout above one hundred megawatts. The constraint is a labor constraint expressed through institutions that hire, train, and certify engineers on multi-decade timescales.
The legacy development sequence is well known. Site control. Owner's engineer scoping. Environmental review. Interconnection study queue. Utility coordination. Land use entitlement. Detailed engineering. Procurement. Construction. Commissioning. Each phase is gated by the prior phase. Each phase is staffed by a different consultancy. Each phase is paper-driven, drawing-driven, and audit-driven. The phases do not parallelize because the legacy institutional structure does not let them parallelize. An engineering firm hired to deliver a substation design will not begin detailed work before the interconnection study is finalized because the interconnection study can move the substation footprint and force a redesign. A civil EPC firm will not begin trenching before geotechnical is finalized because the geotechnical results can move the foundation plan. The serial dependency is not technical. It is contractual and institutional.
What AI-Native Infrastructure Development Looks Like At The Edges
Crusoe operates a distributed power and modular data center model that places compute next to stranded gas or curtailed renewables and stands up megawatt-scale capacity in months rather than years. Stargate's Abilene campus targets an eighteen-month ground-to-power schedule. GridFree and Lancium operate flexibility and load-shifting models that turn data center load into a grid asset rather than a grid burden, which compresses interconnection timelines because the utility analysis changes shape. ExxonMobil's announced behind-the-meter natural gas power push is the clearest signal that the supermajor capital base has decided AI infrastructure is the next deployable use of its gas reserves, and the development model the supermajors will bring is not the legacy EPC model.
Hadrian's factory-as-a-service is the manufactured-precision analogue of what AI-native developers want from their substation, transformer, and switchgear supply chains. A small but growing set of equipment-reservation platforms now lets developers pre-position long lead time electrical equipment against site-agnostic standardized designs, which decouples procurement from final site selection and compresses the critical path. Site intelligence platforms compress the discovery and qualification phase of site selection from a six to twelve month consultancy engagement to a software query. Permit automation tooling is early but real, and the parts of the permit stack that are genuinely automatable (notices, schedules, standard form filings, agency calendaring) are already being automated by developer-side software rather than consultancy-side process.
The operative pattern is parallel execution against software-defined standards rather than sequential execution against bespoke consultancy work product.
Why The Compression Is Structural
Three mechanisms drive the compression. The first is parallelization. When site intelligence, equipment procurement, permit drafting, and interconnection modeling all run on shared software state rather than on serial paper handoffs, the critical path collapses. The work that previously sat in a queue waiting for the prior consultancy to deliver a drawing now runs concurrently because the dependencies are explicit in software and the non-dependencies are released to run.
The second is standardization. Anduril does not design a new airframe for every mission. Hadrian does not design a new factory floor for every customer. The AI-native infrastructure developers are converging on standardized campus blocks, standardized substation designs, standardized modular power blocks, and standardized cooling configurations that absorb site variance in software rather than in custom engineering. A standardized one hundred and fifty megawatt block can be permitted, procured, and constructed in parallel across many sites without re-engineering each one.
The third is autonomous operations. The labor cost and labor scheduling burden of legacy construction is substantial. Autonomous earthmoving, autonomous concrete pours on programmed schedules, drone-based site progress monitoring, and AI-managed construction sequencing remove the dominant source of schedule slippage on legacy projects, which is trade coordination on a multi-thousand-worker site.
None of these three mechanisms is hypothetical. Each is deployed at scale somewhere in the industrial base. The question is not whether the mechanisms work. The question is how fast they consolidate into a coherent development stack that competes head to head with Bechtel and Burns & McDonnell on hyperscale campus work.
What Becomes True By 2028
Three concrete claims define the 2028 endpoint. The first is dominance of the announcement flow. AI-native developers will account for the majority of new-build announcements above one hundred megawatts, and the share will rise with project size. The legacy development stack will continue to dominate transmission, large generation, and federal work, but the campus segment will shift.
The second is legacy consolidation. The engineering consultancies will not disappear. They will either acquire AI-native developers, partner with them through preferred-developer arrangements, or carve out segments where the cost-plus consultancy model still wins (nuclear, federal, complex brownfield, regulated transmission). The cost-plus consulting model will collapse specifically on the greenfield hyperscale campus side because the AI-native developers will deliver the same product on a fixed-price basis at half the schedule.
The third is the new normal of the development cycle. Five to seven year campus timelines will be remembered the way the F-35 program timeline is now remembered: as the legacy benchmark that a new generation of developers compressed by an order of magnitude. Eighteen to thirty months from site control to first power will be the AI-native standard. The legacy stack will rationalize this as the loss of a market segment that was always going to commoditize. The AI-native stack will recognize it as the same transition Anduril ran against Lockheed, run again on a different industrial base.
The Lockheed-to-Anduril transition is not a metaphor. It is the reference implementation. The infrastructure version is in flight.