The Structural Envelope of the Pre-Populated AI Module
Steel-vs-Aluminum, Multi-Module Sealing, and the Floor System Under Liquid Cooling
Abstract
The pre-populated modular AI data center module is increasingly procured as a unit of capital equipment, listed under UL 2755 and equivalent product-listing standards, and dropped into place at the deployment site with the expectation that the listing record substitutes for site-level engineering acceptance. The substitution proposition, addressed in the companion paper (Agee, 2026g), is incorrect at the regulatory and insurance levels and is also incorrect at the structural-engineering level addressed in the present paper. The structural envelope of the pre-populated module — the framing material, the multi-module seam, the floor system, the fire-rated assembly, the thermal-cycling behavior, the galvanic-corrosion service life, and the shipping integrity — is a coordinated engineering system that must satisfy six co-equal constraints simultaneously, and the contemporary procurement practice of evaluating envelope material on shipping cost per unit underweights the structural and fire penalties of the aluminum substitution argument.
The scope of the paper covers material selection (steel versus aluminum across modulus, density, coefficient of thermal expansion, fire performance, galvanic compatibility, and lifecycle cost), structural mechanics under liquid-cooled AI workloads (static loading, dynamic loading, deflection tolerance, levelness tolerance), multi-module sealing (the five continuity disciplines of thermal, vapor, fire, electromagnetic, and acoustic), floor systems (raised versus slab-on-grade across the rack-density range relevant to current accelerators), fire performance (envelope rating versus compartmentation strategy), thermal expansion and cycling behavior (seam dimensional change and bolted-joint preload retention), galvanic corrosion (dissimilar-metal coupling at module-to-module joints), shipping integrity (envelope-level bracing versus component-level shock thresholds for high-bandwidth memory and optical components), and the governance doctrine that binds these elements into an engineering acceptance program. The temporal scope is the 2024-through-2026 code cycle for the applicable standards. The geographic scope is the United States and Canada with reference to harmonized International Organization for Standardization and International Electrotechnical Commission frameworks.
The paper’s principal findings are three. First, aluminum’s elastic modulus is approximately one-third of steel’s and its coefficient of thermal expansion is approximately twice steel’s; equivalent deflection under load requires heavier aluminum sections that erase the mass advantage, and the thermal cycling at bolted seams produces seam-integrity loss that the procurement specification does not currently quantify. Second, multi-module sealing at the seam between adjacent modules carries thermal, vapor, fire, electromagnetic, and acoustic continuity obligations that are independently engineered to different standards; the integration of the five obligations is the responsibility of the operator rather than of any module manufacturer, and the absence of a coordinated seam-assembly engineering posture is the unappreciated single point of failure of the multi-module campus. Third, floor systems under liquid-cooled deployments must support static loads in the 300-to-500-pound-per-square-foot range and dynamic loads from rack-level shock and pump-driven harmonics; the contemporary raised-floor assumption from the legacy data center era is incompatible with the rack densities the present generation of accelerators demands.
The paper’s principal recommendations are three. First, adopt a six-axis material-selection doctrine at design freeze (mass, stiffness, coefficient of thermal expansion, fire performance, galvanic compatibility, lifecycle cost) and require the procurement specification to surface the deflection, thermal-cycling, and fire-rating consequences of the material choice as primary acceptance criteria. Second, engineer the multi-module seam as a single coordinated assembly carrying documented continuity strategies for the five disciplines and require the seam to carry a fire test, an air and vapor permeance test, an electromagnetic compatibility record, and a thermal-bridging analysis as conditions of engineering acceptance. Third, standardize the slab-on-grade floor system for liquid-cooled deployments at the 300-to-500-pound-per-square-foot static-loading range with documented drip-pan integration, leak-detection topology, and overhead service access; treat the raised-floor option as a legacy variant suitable only for specific brownfield retrofits where the slab cannot be poured.
Executive Summary
The thesis of this paper is that the structural envelope of the pre-populated modular AI data center is a single coordinated engineering system rather than a procurement line item, and that the contemporary practice of evaluating the envelope on shipping cost per unit underweights five engineering disciplines that the operator must satisfy across a thirty-year service life. The five disciplines are material selection, structural mechanics under liquid cooling, multi-module sealing, floor system engineering, and fire-rated assembly construction; the paper develops each in turn, identifies the failure modes by category, and prescribes an engineering acceptance program that converts the envelope from a purchased object into a governed engineering deliverable.
Finding 1. Material selection on shipping cost alone is wrong. Aluminum’s elastic modulus is approximately one-third of steel’s; equivalent stiffness therefore requires either deeper aluminum sections (which erase the nominal mass advantage at typical module geometries) or accepting three times the deflection (which violates the rack-levelness tolerance for liquid-cooled racks). Aluminum’s coefficient of thermal expansion is approximately twice steel’s; the seam between adjacent 12-meter modules cycling through ocean transport (-10°C), factory ambient (20°C), and operational heat rejection (45°C) experiences approximately 9.7 mm of dimensional swing in aluminum versus approximately 5.0 mm in steel, both of which exit the typical ±3 mm seam-movement budget at one or both extremes. Aluminum’s melting point of 660°C is well below steel’s 1370°C; the aluminum envelope is rating-limited to approximately one hour of standard fire exposure regardless of protective coating, while the steel envelope reaches two and three hours with conventional protection. Aluminum’s galvanic potential is approximately 0.20 volts more negative than carbon steel and 0.65 volts more negative than stainless steel; the dissimilar-metal joints typical of multi-module construction therefore produce galvanic-corrosion risk that the procurement specification typically does not address. The recommended remedy is the six-axis material-selection doctrine described in Chapter 2 and Chapter 10: evaluate the material on mass, stiffness, coefficient of thermal expansion, fire performance, galvanic compatibility, and lifecycle cost at design freeze, and surface the deflection, thermal-cycling, and fire-rating consequences of the material choice as primary acceptance criteria rather than as derivative consequences of a shipping-cost decision.
Finding 2. The multi-module seam is the unappreciated single point of failure of the multi-module campus. The seam carries five independent continuity disciplines: thermal continuity (governed by ASTM C518 and ASHRAE Standard 90.1), vapor continuity (governed by ASTM E96 and ASTM E2178), air-leakage continuity (governed by ASTM E283), fire continuity (governed by UL 263 and ASTM E119), and electromagnetic compatibility continuity (governed by MIL-STD-461 and IEEE 299). Each discipline is independently engineered by a different practitioner against a different test method; integration of the five at the seam is no party’s specified responsibility absent a deliberate engineering acceptance program. The contemporary seam assembly consists of six layers acting in series — an interior elastomeric sealant, an intumescent firestop strip, a mineral-wool thermal and acoustic core, a conductive electromagnetic-compatibility gasket, a vapor barrier tape, and an exterior elastomeric sealant — and the failure of any one layer breaks the corresponding continuity discipline regardless of how well the remaining layers perform. The recommendation is to treat the seam as a discrete engineered assembly with a single accountable party, a documented test record, and a service-life maintenance plan, rather than as a residual of the multi-module construction sequence.
Finding 3. The floor system question is settled in favor of slab-on-grade for liquid-cooled deployments at the rack densities the present generation of accelerators demands. The legacy raised-floor architecture is incompatible with the 300-to-500-pound-per-square-foot static loading produced by modern racks, exhibits pedestal-failure modes under the dynamic loading produced by pump-driven and fan-driven harmonics, and complicates the leak-detection topology that liquid cooling requires. The slab-on-grade alternative accepts the higher static load, eliminates the pedestal failure mode, consolidates leak detection at the slab perimeter and beneath drip pans, and supports the overhead cooling and cable distribution that liquid-cooled rack architectures favor. The floor system specification developed in Chapter 5 calls for an 8-to-12-inch reinforced concrete slab on a compacted granular subbase, with surface flatness of FF 50 minimum and FF 75 for rack rows, surface levelness of FL 35 minimum and FL 50 for rack rows, continuous perimeter leak detection, and drip-pan integration beneath each rack row.
Recommendation 1. Adopt the six-axis material-selection doctrine at the design-freeze stage of every modular AI data center deployment. Evaluate the envelope material on mass, stiffness, coefficient of thermal expansion, fire performance, galvanic compatibility, and lifecycle cost; weight the axes per the deployment context; document the decision with a single-page selection record; and surface the deflection, thermal-cycling, and fire-rating consequences as acceptance criteria for the manufactured module rather than as residual consequences of the shipping-cost decision.
Recommendation 2. Engineer the multi-module seam as a single coordinated assembly. Specify the six-layer composite seam (or an equivalent five-discipline assembly) at design freeze; require the manufacturer or the integrator to deliver the seam with a fire test record per UL 263 or ASTM E119, an air-permeance record per ASTM E283, a vapor-permeance record per ASTM E96, an electromagnetic-compatibility record per MIL-STD-461 or IEEE 299, and a thermal-bridging analysis demonstrating compliance with the energy code in the deployment jurisdiction; assign single-party accountability for the assembly across the operating life.
Recommendation 3. Standardize the slab-on-grade floor system for liquid-cooled AI deployments. Specify the slab thickness, reinforcement, concrete strength, surface flatness, surface levelness, joint spacing, and vapor retarder per the values developed in Chapter 5; integrate the leak-detection topology and the drip-pan layout into the slab construction documents; treat the raised-floor option as a legacy variant suitable only where the slab cannot be poured or extended.
Forecast. The substitution of aluminum for steel in module envelopes will continue to grow as the transport-mass argument becomes more visible in procurement conversations; the structural and fire-performance penalties of the substitution will become operationally visible in the public record within the next thirty-six to sixty months in the form of seam-integrity loss, rack-levelness drift, fire-rated assembly performance disputes, and the first thermal-cycling-driven fatigue events on bolted seams that have been in service for five or more years. The framework presented in this paper is the prevention.
Full white paper below
