Why Data Center Designs Work
Engineering Certainty, Not Hypothesis
Abstract
This publication develops the central claim that the modern AI-factory campus is the product of engineering certainty rather than engineering hypothesis. The contemporary industry’s tendency to frame each new generation of campus construction as a frontier-engineering project is rhetorically powerful and commercially useful, but it is fundamentally misleading: the engineering surface of the modern AI campus is bounded, the body of established practice is mature, and the disciplined application of settled engineering — rather than the perpetual reinvention of solved problems — is what distinguishes well-executed deployments from troubled ones.
The framework presented organizes the campus problem into six engineering domains: power architecture from substation through silicon (Agee, 2025a; Agee, 2025b; Agee, 2025c), cooling architecture from chiller plant through cold plate (Agee, 2025k; Agee, 2026j; Agee, 2026k), structural and spatial design from envelope through floor system (Agee, 2025n; Agee, 2026n), telemetry and control from field device through enterprise historian (Agee, 2025r; Agee, 2025s), lifecycle and operating-model discipline from initial commissioning through long-term adaptation (Agee, 2025p; Agee, 2025q), and governance and standards posture from change-control through regulatory compliance (Uptime Institute, 2024; NFPA, 2023a; IEEE, 2022a; IEC, 2021a; ASHRAE, 2021a; NIST, 2020).
The analysis draws on the published frameworks of the Uptime Institute (Uptime Institute, 2024), ASHRAE TC 9.9 (ASHRAE, 2021a), the Open Compute Project (OCP, 2023), the National Fire Protection Association (NFPA, 2023a; NFPA, 2023b), the Institute of Electrical and Electronics Engineers (IEEE, 2022a; IEEE, 2018), the International Electrotechnical Commission (IEC, 2021a; IEC, 2020), the National Institute of Standards and Technology (NIST, 2020), the International Organization for Standardization (ISO, 2018), the Federal Energy Regulatory Commission (FERC, 2023), and the U.S. Department of Energy (DOE, 2023). It is informed by the author’s prior published portfolio across power, cooling, structural, operational, governance, and capital domains, cited throughout in APA-style in-text references.
The publication’s principal findings are that approximately 60 to 90 percent of the engineering surface in each of the six domains is settled and should be executed by reference architecture rather than re-derived; that the small open-engineering surface is concentrated in four recurring decisions (the workload-bound cooling choice, the capacity scaling cadence, the bespoke-versus-extendable mode selection, and the operating-model maturity commitment); and that the campus’s engineering, capital, and operating commitments should be captured in a single written architecture envelope authored at campus-program scope rather than at the rack or hall level. The recommendations are operationalized through six body chapters that develop each engineering domain in turn, supplemented by appendix glossary, acronym, and reference material.
The geographic and temporal scope is the contemporary North American and European hyperscale and colocation industry from 2024 through the expected technology refresh cycle of 2028, with explicit attention to the rack-density envelope from 5 kilowatts at the edge through 250 kilowatts at frontier training scale. The analytical posture is evenhanded across operator tiers, equipment categories, and capital postures, and the publication’s recommendations are intended to be useful at the hyperscale boardroom, the federal regulatory review committee, the infrastructure investment firm, the utility planning authority, and the executive engineering steering committee simultaneously.
Executive Summary
The modern AI-factory data center is not a frontier-engineering hypothesis. It is the disciplined application of an established engineering vocabulary against an evolving but bounded set of workload, density, and operating-model constraints. The contemporary industry’s tendency to frame each successive campus build as a frontier-crossing project — driven by chip generation, workload class, or capital pace — obscures the underlying engineering reality that the principal architectural decisions have converged on a small number of settled reference patterns, and that the engineering attention each program genuinely requires is concentrated on four recurring open decisions rather than on the entire campus design surface (Agee, 2025a; Uptime Institute, 2024).
Finding one. Power architecture from substation through silicon is engineering-settled in three reference configurations: the 415 V AC chain for racks at or below 60 kilowatts, the 800 V high-voltage direct-current chain for racks in the 60 to 250 kilowatt density band, and the solid-state-transformer chain for hyperscale campuses where direct medium-voltage-to-rack DC is justifiable. Each is documented in the published reference-architecture literature and is supported by a mature equipment ecosystem at multiple supplier tiers (Agee, 2025b; Agee, 2025c; Agee, 2025e; IEEE, 2022a; IEC, 2021a; OCP, 2023). The engineering organization that designs a campus power chain in 2026 is applying a published vocabulary rather than inventing one.
Finding two. Cooling architecture from chiller plant through cold plate is engineering-settled in three configurations matched to chip thermal design power: single-phase direct-to-chip cooling for the 1000 to 1500 watt range, two-phase direct-to-chip cooling for the 1500 to 2000 watt range, and dielectric immersion cooling for the chip-package thermal envelope above 2000 watts and for custom-form-factor compute assemblies. The configurations are supported by the published ASHRAE TC 9.9 thermal guidelines and by the cooling-design literature accumulated across multiple decades of building HVAC and industrial process cooling experience (Agee, 2025k; Agee, 2026j; Agee, 2026k; ASHRAE, 2021a; ASHRAE, 2022).
Finding three. Structural, spatial, telemetry, lifecycle, and governance disciplines have similarly converged on a small set of settled patterns. The three-zone spatial framework separates IT halls, MEP galleries, and utility yards into distinct architectural zones (Agee, 2025n). The slab-on-grade floor system dominates contemporary liquid-cooled builds. The five-level commissioning workflow gates the construction phase. The operating-model maturity envelope characterizes the discipline progression from initial through optimized operations. The architecture change-control decision tree distinguishes within-envelope changes from envelope-breaking changes that require board review (Agee, 2025p; Agee, 2025q; Uptime Institute, 2024; ISO, 2018).
Recommendation one. Allocate engineering attention to the four decisions that are genuinely open at every campus program: the workload-bound cooling choice (Agee, 2025k; Agee, 2026j), the capacity scaling cadence (Agee, 2025d; Agee, 2025t), the bespoke-versus-extendable mode selection (Agee, 2026a; Agee, 2026e), and the operating-model maturity commitment (Agee, 2025p; Agee, 2025q). Each of these four decisions recurs at every modern campus and demands deliberate engineering judgment rather than reference application. Engineering organizations that allocate their scarce attention to these four decisions and execute the remaining design surface by reference architecture deliver consistent results across multiple campuses; those that re-derive the entire design surface at every program absorb avoidable engineering risk and schedule cost.
Recommendation two. Maintain a written architecture envelope at the campus-program scope rather than at the rack or hall scope. The envelope is the single document that the capital sponsor signs against, the engineering organization designs against, and the operating team operates against. It declares the campus’s workload class commitment, capacity and density bounds, power and cooling reference configurations, operating-model commitments, and governance and standards posture. Changes to the envelope require formal change control, and envelope-breaking changes require board or capital-committee review (NFPA, 2023a; IEEE, 2022a; Uptime Institute, 2024).
Recommendation three. Treat the operating-model maturity commitment as a first-class architectural decision rather than as a downstream operational concern. The operating team that converges on optimized maturity within the first two years of the operate phase delivers a campus that delivers its committed availability; the operating team that does not absorbs the operational cost of premature failure escalation, missed lifecycle adaptation, and inadequate telemetry posture (Agee, 2025p; Agee, 2025q; Agee, 2025r; Uptime Institute, 2023).
Scenario and forecast. The publication’s central forecast across the 2026 to 2028 technology refresh window is that the rack-density envelope will continue to shift upward, that liquid cooling will displace air cooling for the entire frontier and production-training workload class, that 800 V HVDC will become the default architecture for new builds above 60 kilowatt rack density, and that solid-state transformer deployment will emerge from the early-adopter hyperscale tier into the broader hyperscale and large-colocation tiers as the equipment ecosystem matures (Agee, 2025c; Agee, 2026b). The forecast supports the publication’s recommendation that engineering organizations treat the settled architecture as settled and concentrate engineering attention on the open decisions.
The paper is intended for executives, engineering organizations, capital sponsors, operating teams, hyperscale customers, utility planning authorities, government regulators, infrastructure investors, and the manufacturing and integration communities that build the modern AI-factory campus. Each audience is offered a body of analysis structured to its specific decision posture, with engineering content sized for engineering review, capital content sized for capital review, and operational content sized for operational review. The body chapters develop the six engineering domains in sequence; the appendices consolidate glossary, acronym, and reference material; and the closing note from the author offers a reflective conclusion on the publication’s strategic importance.
Full white paper below
