AI data-center engineering

The pressure

Why AI compute breaks conventional facility engineering

The load is growing faster than the buildings. Data centers consumed about 1.5% of the world’s electricity in 2024, and the IEA projects that figure to roughly double by 2030, to about 945 TWh.^[1] In the United States, data centers drew 4.4% of national electricity in 2023; Lawrence Berkeley National Laboratory projects between 6.7% and 12% by 2028.^[2]At the same time, the industry’s average efficiency has stopped improving — Uptime Institute’s 2025 survey finds industry-average PUE roughly flat for about six years, even as typical rack densities climb into the 10–30 kW band.^[3]

Conventional engineering answers this with bigger projects: more shell, more chillers, more substation. A factory-built approach answers it with a repeatable unit — solve each domain once, then manufacture the solution. The trade-offs between the two approaches are examined in modular AI data center vs traditional data center; the product architecture that results is described in the platform overview.

This index summarizes each domain, explains how the constraints change when the facility is a factory-built unit rather than a construction project, and links to the deep dive for every one.

Domains 01–04

Cooling, power, density, and the boundary they all share

ENG-01 · Cooling: move heat with liquid, not air

At AI densities, air stops being a workable transport medium for heat. ASHRAE TC 9.9 — the committee whose thermal guidelines define data-center environmental classes — has documented the shift toward liquid cooling in mainstream facilities as rack power rises.^[4][5] The practical response is direct-to-chip liquid cooling: cold plates on the hottest silicon and a coolant loop in place of a room full of moving air, with the Open Compute Project now maintaining multi-vendor requirements for cold plates and coolant distribution units.^[6]The PODOS Pod’s cooling is designed as a closed direct-to-chip loop, which also concentrates heat into a recoverable stream — NREL’s ESIF facility, for example, heats its offices with waste heat from liquid-cooled supercomputers and reports an annualized PUE near 1.04.^[10] Both halves have their own explainer: direct-to-chip liquid cooling and data-center heat recovery.

ENG-02 · Power: from medium voltage to the rack

A megawatt of IT load is an industrial electrical project. The chain runs from a medium-voltage utility feed through transformation, switchgear, distribution, and power conversion down to the rack — and every stage adds losses, footprint, and failure modes. Reliability engineering for that chain is a discipline of its own; the IEEE 3006 series covers reliability analysis for critical-facility power systems.^[8] Field-built electrical rooms are engineered one project at a time. A factory-built unit integrates the same chain into a manufactured product, so the design is validated once and then repeated; the PODOS Pod is designed to accept a medium-voltage utility feed and carry distribution inside the unit. The full chain is walked stage by stage in data-center power architecture.

ENG-03 · Compute density: the number every other domain inherits

Density sets the requirements for cooling, power, envelope, and safety at once. Uptime Institute’s 2025 survey shows typical racks moving into the 10–30 kW band,^[3]and accelerator vendors have already moved past it: NVIDIA’s GB200 NVL72 packages 72 GPUs and 36 CPUs into a single liquid-cooled rack that behaves as one NVLink domain.^[7] Each PODOS Pod is designed for 128 GPUs inside its unit envelope — a design figure, not a measured deployment. The rack-level design problem is taken up in high-density GPU infrastructure. Signs a design has crossed the air-cooling threshold:

• Rack loads pushing past the range ASHRAE documents as air’s practical territory^[4][5]
• Accelerator inlet temperatures riding the top of the allowable class limits^[4]
• Hot-aisle containment already deployed — and still insufficient
• Fan energy becoming a visible share of facility overhead

ENG-04 · Thermal envelope: the boundary condition

The envelope decides how much of the outside climate the cooling plant has to fight. ASHRAE’s environmental classes — A1–A4 for air-cooled equipment, an H1 class for high-density gear, and liquid-cooling classes named by facility water temperature — define the machine-side climate that must hold regardless of weather.^[4] A conventional building maintains that climate with mass and mechanical plant, assembled on-site. A manufactured enclosure treats the envelope as a product surface: insulation and barriers engineered once, on a production line. The PODOS enclosure is designed as a fully insulated envelope, so cooling capacity is spent on silicon rather than on the weather — the subject of thermal enclosure design.

Domains 05–07

The fabric, the telemetry, and the code

ENG-05 · Networking: the east–west fabric

AI clusters live or die on east–west bandwidth — traffic between accelerators inside the cluster, not north–south traffic to the internet. Rack-scale designs make the point structurally: the GB200 NVL72 presents 72 GPUs as one NVLink domain because the interconnect is, in effect, the computer.^[7] At unit scale the open engineering questions are topology, cabling economics, and how a unit designed as a standardized 1-MW block joins a larger fabric without re-architecting it — the questions taken up in AI data center network architecture: fiber and leaf-spine.

ENG-06 · Monitoring: telemetry as a design input

Half of the operators surveyed by Uptime Institute in 2025 reported an impactful outage within the previous three years.^[3] Monitoring is the difference between a derate you catch and an outage you explain. In a factory-built unit, telemetry can be designed in — sensor points and alarms specified on the production line rather than commissioned ad hoc at each site. The open questions are which signals matter per subsystem, and what a fleet of standardized units makes possible that one-off facilities cannot: like-for-like comparison across every unit in service. Both are worked through in data-center monitoring and controls.

ENG-07 · Safety: fire, energy storage, and code

Data-center safety is governed by code. NFPA 75 covers fire protection for information-technology equipment spaces, and its 2024 edition moves stationary lithium-ion battery requirements out to NFPA 855 — so any unit that carries on-site energy storage inherits both standards.^[9] Factory manufacture changes the compliance surface: detection, suppression, and egress can be engineered into a repeatable product instead of re-derived per project. It does not remove the local permitting authority, whose review remains site-specific. The standards landscape is mapped in modular data center fire safety and physical security.