THE GRID-TO-CHIP REFERENCE ARCHITECTURE
A Unified Reference for Hyperscale AI Infrastructure from Bulk Transmission to Silicon-Level Power Delivery
Abstract
This paper presents a unified reference architecture for hyperscale artificial intelligence infrastructure spanning the complete grid-to-chip power chain from the bulk-transmission interface through silicon-level power delivery, including the cooling chain that operates concurrently with the power chain at two megawatts and greater rack densities. The reference architecture is organized as a single integrated object rather than as a sequence of independently managed subsystems. The integration is necessary because the operating envelope at hyperscale densities couples electrical and thermal subsystems dynamically, exceeds the historical bounds of single-discipline design practice, and requires governance maturity that no single dimension can produce alone.
The paper integrates the prior five publications in the FCG executive technical series. Voltage selection as capital governance is treated as the architectural foundation that determines downstream supply-chain, schedule, and operating consequences. Code compatibility is treated as the standards alignment posture that converts architectural commitment into permitted installations. Operator-model maturity is treated as the institutional capability that determines whether deployed architectures operate at design intent. Interconnect sequencing is treated as the upstream pacing function that determines whether programs proceed on schedule. Supply-chain readiness is treated as the procurement constraint that determines whether equipment is available at required scale and timing. The five dimensions interact across the chain, and the paper develops the interactions explicitly rather than treating each dimension independently.
Three reference architectures are presented. Architecture A is the 415-volt alternating-current baseline appropriate for densities up to approximately one megawatt per rack. Architecture B is the hybrid alternating-current and direct-current configuration appropriate for densities in the one to two megawatt per rack range and for operators in active migration toward higher voltages. Architecture C is the 800-volt direct-current end-state appropriate for densities at and above two megawatts per rack. A phased migration path from Architecture A through B to C is presented for operators with high uncertainty about density growth horizon. Five reference deployment archetypes are presented for 50-megawatt edge campuses, 100-megawatt colocation, 250-megawatt hyperscale, 500-megawatt-plus AI campuses, and behind-the-meter integrated deployments.
The paper is calibrated to programs operating at densities at or projected to exceed two megawatts per rack with campus capacities in the 50-megawatt to 500-megawatt range during the 2026 through 2028 deployment window. Reference materials from the Data Center World 2026 conference are cited where directly relevant and are catalogued in Appendix C. The paper distinguishes verified facts, considered analysis, structured inference, and explicit assumption throughout, and includes substantial diagrammatic and tabular content to support executive review and capital committee approval.
Executive Summary
Hyperscale artificial intelligence infrastructure deployment in the current cycle requires governance maturity across five interacting dimensions: architecture, standards, operations, interconnect, and supply chain. The five dimensions are addressed individually in the prior five papers of this FCG executive series. They are addressed together in this capstone paper because operators that apply only some of them produce gaps at the boundaries between applied and unapplied dimensions. Operators that apply all five concurrently produce coherent program governance that delivers programs on schedule, within capital, and at design intent. The integration is the point.
The unified grid-to-chip reference architecture presented in this paper organizes the power chain into nine layers from bulk transmission through silicon-level conversion and the cooling chain into six layers from heat rejection through cold plate to silicon thermal interface. The layers are not independent; at two megawatts and above per rack the time constants of the power and cooling chains overlap, the failure modes of each chain interact with the other, and the operating model must address the integrated chain rather than independent subsystems. The unified architecture makes the integration explicit.
Three reference architectures are presented. Architecture A is the 415-volt alternating-current baseline that is appropriate for sub-one-megawatt-per-rack densities and that is the mature default of the broader industry. Architecture B is the hybrid alternating-current and direct-current configuration appropriate for one to two megawatt per rack densities and for operators migrating toward higher voltages. Architecture C is the 800-volt direct-current end-state appropriate for two-megawatt-plus per rack densities and for hyperscale AI campuses at the leading edge of the operating envelope. A phased migration path from Architecture A through B to C is presented for operators with high uncertainty about density growth horizon. The architecture selection decision is supported by a decision tree, a quadrant analysis on capital exposure and density growth, and a comparative table across multiple architectural attributes.
Five reference deployment archetypes are presented. The 50-megawatt edge campus archetype favors Architecture A or hybrid B with limited DC pilot. The 100-megawatt colocation archetype favors Architecture B because customer mix varies in density. The 250-megawatt hyperscale archetype favors Architecture C with phased deployment to manage supply-chain exposure. The 500-megawatt-plus AI campus archetype favors Architecture C with solid-state transformer-based transformation and integrated energy storage. The behind-the-meter integration archetype is overlaid onto any of the four base archetypes for operators in regulatory environments that support behind-the-meter capacity and for operators that need to accelerate partial operating capacity ahead of utility energization.
Six recommendations integrate the prior five papers into a single executive posture. First, treat voltage architecture as upstream capital governance rather than as downstream electrical detail; this is the foundational recommendation from which the others follow. Second, complete standards alignment before architectural commitment so code compatibility is documented rather than assumed. Third, engage utilities at the architecture gate of the broader governance framework rather than at construction; pre-application briefing is the highest-leverage activity in interconnect engagement. Fourth, apply the operator-maturity model honestly rather than aspirationally; plan programs to assessed maturity and invest deliberately in maturity advancement. Fifth, manage supply chain as governance rather than procurement; apply the supplier readiness funnel at every procurement decision and structure operator-OEM relationships as multi-program partnerships where appropriate. Sixth, integrate the five dimensions into a single governance practice rather than maintaining each as an independent workstream; the integration is the strategic capability that distinguishes mature operators from average ones. Applied together, the six recommendations produce a campus program that operates at the operating-model maturity hyperscale customers expect and that the supply chain can support across deployment cadences (Uptime Institute, 2022; Woods & Hollnagel, 2006).
The paper is intended as a working capstone reference for executive sponsors, capital committee members, chief architects, standards leads, supply-chain leaders, operations directors, and senior engineering leaders with responsibility for AI infrastructure programs. The paper is also relevant to government, regulator, utility, and investor audiences because the unified reference architecture is at the intersection of multiple stakeholder interests in the hyperscale infrastructure transition.
Full white paper below
