Where Smoothing Belongs: White-Space, Gray-Space, and Black-Space BESS Topology
A Decision Framework for Pushing Energy Storage Outward as AI-Factory Build Size Scales
Abstract
This paper presents a four-tier framework for deciding where, physically, to place battery energy storage in AI-era data centers as the build size of the facility scales from sub-megawatt edge nodes to gigawat
t-class AI factories. The framework distinguishes among rack-sidecar placement inside the IT white space, sidecar placement at the power-distribution-unit tier in gray-space-adjacent alcoves, row-tier and hall-tier placement in dedicated gray-space rooms, and external placement in outdoor black-space yards beyond the building envelope. The placement decision is treated as a coupled engineering, code-compliance, cybersecurity, and capital-governance decision, rather than as an electrical-engineering optimization with safety and governance as afterthoughts. The paper draws on the predecessor analyses (Agee, 2026a) on the gray-space doctrine, (Agee, 2026b) on the transient electrical loads imposed by AI workloads, and (Agee, 2026c) on the reliability split between training and inference, and extends them with a synthesis that connects the time-domain of the transient being smoothed to the natural physical placement of the reservoir that smooths it.
The methodology integrates four threads. First, the electrical-engineering thread maps transient phenomena across nine decades of time-scale, from microsecond switching ripple to multi-hour ride-through, and identifies which technology and which placement tier naturally owns each band. Second, the code and standards thread evaluates the NEC, NFPA 855, NFPA 70E, NFPA 75, UL 9540, UL 9540A, IEEE 1547, NIST SP 800-53, and the Uptime Institute and TIA-942 frameworks against each candidate placement, identifying disqualifying conditions, manageable conditions, and naturally compatible conditions. Third, the cybersecurity thread applies zero-trust and least-privilege doctrine to the physical zone boundaries the placement creates or collapses, drawing on the architectural reasoning developed in the predecessor papers and codifying it as a procurement and audit obligation. Fourth, the capital and operational thread decomposes the cost stack and the operational complexity burden for each placement, drawing on Wood Mackenzie, BloombergNEF, EPRI, and hyperscaler-disclosed reference data alongside the author’s direct field experience.
The principal findings are that placement is a governance decision with electrical consequences rather than the reverse; that the time-domain of the transient determines the natural physical tier; that rack-sidecar placement above approximately 50 kW per cabinet creates compounding code, cyber, and life-safety exposures that cannot be retrofit out of the architecture; that above approximately 50 MW of contiguous IT load the conversation is no longer about which room inside the building but about which yard outside it; and that the four-tier framework, applied as a coordinated set rather than as a single-tier choice, produces lower total cost of ownership and higher operational maturity than any single-tier optimization.
The principal recommendations are that the basis of design for any new AI-factory build adopt the four-tier framework before site selection is locked; that placement decisions be approved by a defined governance circle (operations, facilities AHJ, cybersecurity, electrical engineering, insurer, and where applicable utility and regulator) rather than by electrical engineering alone; that builds above 50 MW default to a black-space external BESS yard with grid-forming power-conversion sub-systems unless an explicit and documented exception applies; that vendor qualification require UL 9540 and 9540A test data at the appropriate installation scale; and that revisits to the placement decision be scheduled at each two-times scale milestone in the program rather than treated as one-time decisions.
The geographic scope of the paper is the United States data center industry, with code references drawn from the 2023 National Electrical Code, the 2024 edition of NFPA 70E, NFPA 855 in its 2023 edition, and TIA-942-C in its 2024 edition. The temporal scope is the 2026 through 2030 deployment window during which AI-factory-class facilities are expected to consume an increasing share of new electrical demand and during which the regulatory and code environment surrounding behind-the-meter generation and battery storage is expected to mature substantially. The analytical posture is that of an independent advisory analysis prepared for owners, operators, investors, regulators, government agencies, and industry consortia. The framework is vendor-agnostic by construction; no manufacturer’s product portfolio is recommended over another, and the procurement guidance permits any qualified manufacturer to satisfy the placement-tier specifications it describes.
Executive Summary
The question this paper takes up is, on its surface, a narrow one. As battery energy storage becomes load-bearing in AI-era data center designs, where in the building should the storage actually sit? The question can be asked of any AI-factory program, from sub-megawatt edge pilots to gigawatt-class campuses, and the answers proposed in current industry literature range from rack-sidecar packages mounted alongside individual server cabinets, to power-distribution-unit-tier cabinets, to row-level pods, to hall-scale uninterruptible-power-supply-plus-storage banks, to outdoor yards beyond the building envelope. The choices look like an engineering optimization. They are not. They are a coupled engineering, safety, code, cybersecurity, and capital-governance decision whose first-order consequences are operational and whose long-tail consequences are financial.
The thesis of this paper is that as build size scales, the appropriate physical location for energy storage shifts outward. The framework introduced here treats the four candidate placements as a coordinated tier system rather than as competing options. Each tier owns a time-domain of the transient phenomena that AI workloads impose. Rack-level capacitance owns microsecond and short-millisecond ripple; the power-distribution-unit tier owns tens-of-milliseconds to single-second transients; the row and hall tiers own seconds to tens of seconds; and the external black-space yard owns the tens-of-seconds to hours range, including utility ride-through, demand-charge management, and behind-the-meter generation protection. A scale-appropriate design uses all four tiers together. An under-scaled design tries to make one tier do the work of two, and pays for the choice in operations, audit cycles, and lifecycle ownership.
Three findings frame the analysis. The first finding is that placement is a governance decision with electrical consequences rather than the reverse. The dominant variable is not the electrical-engineering optimization but the codified spatial separation between IT load and energy-storage equipment, expressed through NEC working clearances, NFPA 855 separation distances, NFPA 70E arc-flash boundaries, NFPA 75 fire-protection envelopes, NIST SP 800-53 physical-environment controls, and TIA-942 and BICSI 002 space classifications. The second finding is that the time-domain of the transient determines the natural placement. A nanosecond-class current step belongs to capacitance at the rack; a multi-second all-reduce cluster sync belongs to a hall-level reservoir; a multi-minute utility ride-through belongs in a black-space yard. The third finding is that above approximately fifty megawatts of contiguous IT load, the conversation is no longer about which room inside the building but about which yard outside it. Black-space yards are the dominant architecture above this threshold for engineering, code, and operational reasons that compound rather than offset.
Three recommendations follow. The first recommendation is that the basis of design for any new AI-factory build adopt the four-tier framework before site selection is locked. The placement decision sets the perimeter for code-of-record selection, cyber-zone classification, civil and structural design, and utility interconnect strategy, and treating it as a downstream electrical-engineering decision forces every upstream decision into reactive posture. The second recommendation is that the placement decision be approved by a defined governance circle, not by electrical engineering alone. The governance circle should include operations, the facilities authority-having-jurisdiction, cybersecurity, electrical engineering, the property and casualty insurer, and where applicable the serving utility and the regulator. The third recommendation is that builds above fifty megawatts of contiguous IT load default to a black-space external BESS yard with grid-forming power-conversion sub-systems, unless an explicit and documented exception applies. The exception path should require that the placement decision be revisited at each subsequent two-times scale milestone in the program rather than treated as a permanent commitment.
The companion findings of the paper, developed in the body chapters and appendices, include a placement-decision tree keyed to build size, workload class, and reliability tier; a capital and operating cost decomposition for each tier; a code-and-standards mapping across NFPA, UL, IEEE, NIST, TIA, and BICSI; reference architectures at five-megawatt, fifty-megawatt, five-hundred-megawatt, and one-and-a-half-gigawatt scale; and a consolidated recommendations table cross-referenced to standards. The framework is intended to be vendor-agnostic in substance and is presented in this FCG edition for advisory engagements with owners, operators, investors, regulators, government agencies, and industry consortia. A separate Delta edition exists for OEM and hyperscaler customer discussions; both editions share the same engineering substance, the same references, the same recommendations, and the same governance framework, and a reader who consults both will find the technical analysis identical.
The forecast posture of the paper is that the four-tier framework will move from being an emerging practice in 2026 to being the dominant practice by approximately 2028 in the United States and by approximately 2030 in the European Union, driven by code adoption cycles, insurance underwriting practice, and operator experience with the failure modes that single-tier under-scaled designs surface. Operators, engineering organizations, and OEMs that anchor their decisions to the framework now will avoid the retrofits that single-tier choices will force in two-to-five-year time frames. The paper closes with a closing note from the author placed after the references appendix, intentionally, so that the final voice in the publication is the author’s reflection rather than the bibliography.
Full white paper below
