Construction Oversight & Validation: Preventing Design‑to‑Build Drift in AI Facilities
Uptime Institute’s AI Infrastructure Advisory – Part 3 of 5
You have a solid design (Part 1) and you have selected reliable vendors and technologies (Part 2). Now comes the moment of truth: construction. According to Uptime Institute, this is where many AI projects face their greatest risk – design‑to‑build drift.
Fast schedules, evolving AI hardware, and the sheer complexity of high‑density systems mean that what gets built often deviates from what was designed. If caught early, corrections are straightforward. If discovered during commissioning or operations, remediation costs can multiply – and schedules can slip by months.
This post, based on Uptime Institute’s Paper 3, covers the technical considerations for constructing AI data centers, including structural demands, liquid cooling integration, and a phased oversight strategy.
The Construction Challenge: Speed vs. Fidelity
AI projects are under immense pressure to go live. Business leaders want clusters online in months, not years. But cutting corners on construction quality can permanently compromise resiliency, scalability, and safety.
Uptime’s warning:
AI infrastructure is more complex than conventional data centers. The risk of design‑to‑build drift increases with complexity. Common sources of drift include:
Substitutions of materials or equipment (value engineering)
Installation errors in power or cooling distribution
Changes to floor plans without updating resiliency topologies
Inadequate coordination between multiple contractors (electrical, mechanical, civil)
The solution according to Uptime:
Independent, milestone‑based construction oversight that verifies each phase against the design intent before the next phase begins.
Physical Demands: Bigger, Heavier, Taller
Uptime Institute’s Paper 3 highlights several structural requirements that are often underestimated.
Rack Weight and Floor Loading
Conventional rack: 680 - 1,000 kg
AI high‑density rack (DLC, 130 kW): >2,000 kg (two to three times heavier)
Immersion tanks (if used): Even heavier, often requiring dedicated structural reinforcement
Technical implication:
Standard raised floors are usually inadequate. Uptime recommends solid concrete slabs or structurally braced raised floors with verified point‑load ratings. Floor loading verification must occur before equipment delivery.
Ceiling Height and Overhead Space
Conventional: ~42U racks (6 ft), standard ceiling height
AI: Racks can reach 48U–52U (7–9 ft) due to taller GPU servers and integrated cooling manifolds. Additional overhead space is needed for:
Thicker power cables (higher current)
Liquid cooling pipes and manifolds
Network cabling (high‑speed copper/fiber)
Uptime’s recommendation:
Design ceiling height for future densities. A facility built today may need to support 1 MW per rack within its lifetime – plan for at least 1–2 feet of extra overhead clearance.
Multi‑Story Facilities
Large training clusters require extremely low latency between GPUs. This often drives designs toward multi‑story buildings – systems on an upper floor can be closer to cooling and power on the floor below than in an adjacent building.
Construction implications:
Stronger floor slabs on every level (not just ground floor)
Pre‑fabricated openings for cable trays and piping between floors
Vertical risers sized for future capacity
Corridors, Lifts, and Loading Docks
AI equipment is not only heavier but also bulkier. CDUs, UPS modules, and switchgear may not fit through standard doorways or lifts.
Checklist for construction plans:
Corridor widths adequate for moving 2,000+ kg racks on pallet jacks or small cranes
Lift capacity (if multi‑story) – many standard freight lifts are rated for 1,500–2,000 kg; AI gear may exceed that
Loading dock height and turning radius for delivery trucks carrying oversized cratesCooling Subsystems: What You Must Specify
Liquid Cooling Infrastructure: Not Just Pipes
Direct liquid cooling (DLC) introduces a new set of construction considerations.
CDUs and Gray Space
Coolant distribution units (CDUs) are large, heavy heat exchangers. In Uptime’s reference design, each 5 MW data hall requires multiple CDUs (e.g., three 1.25 MW units for N+1). These often reside in gray space – the technical area outside the white space.
Construction nuance:
Gray space may need to be larger than in conventional facilities. Some designs allocate >50% of white space area to gray space for CDUs, pumps, and power gear. This must be reflected in floor plans and structural slabs.
Piping and Leak Containment
DLC systems circulate water/glycol mixtures. Even a small leak can damage IT equipment. Construction must include:
Double‑containment piping or drip trays under all coolant lines
Leak detection cables at every joint and manifold
Sloped floors or drainage channels in case of large leaks
Physical separation between liquid piping and electrical cables
Uptime’s advice:
Operators generally prefer to locate liquid circulation below the rack (underfloor) for leak containment. This affects the choice between slab and raised floor. A solid slab with trenches may be preferable to a traditional raised floor.
Hybrid Cooling: Two Systems in One
High‑density AI facilities have effectively two independent cooling systems:
Liquid system (CDUs, cold plates, manifolds) for GPUs
Air system (fan walls, CRAHs) for networking and storage
Construction complexity:
These systems share some components (chillers, pumps) but have separate distribution paths. Coordination between mechanical and electrical contractors is essential. Uptime recommends pre‑functional testing (Level 3) of each subsystem before integration.
Phased Construction: The Uptime Approach
Because AI technology evolves rapidly, Uptime Institute recommends phased construction rather than building the entire facility at once.
Construction Sequencing for Flexibility
Uptime advises against installing large plant items (chillers, UPS systems, switchgear) too early. Instead:
Complete civil, structural, and architectural work for the entire site (foundations, slabs, shell).
Build out the first phase completely – including all power and cooling.
Commission the first phase (Levels 4 and 5) before starting the second phase.
For subsequent phases, adjust designs based on lessons learned and new hardware roadmaps.
Why this matters:
If you lock in a 5 MW chiller plant for the whole facility and later need 10 MW per hall, you face expensive retrofits. Phased construction keeps options open.Vendor Evaluation: A Structured Approach
Uptime Institute emphasises a vendor‑neutral, disciplined process to avoid lock‑in and ensure long‑term flexibility.
Retrofit Considerations: Converting Existing Space
Uptime Institute notes that retrofitting a conventional data center for AI training is rarely practical or economical. Why?
Power capacity – AI training may require doubling the building’s existing power feed.
Gray space – Conventional facilities have limited space for CDUs, pumps, and extra switchgear.
Floor loading – Existing raised floors often cannot support 2,000+ kg racks.
Cooling – Air‑cooled halls lack the plumbing and CDU space for DLC.
When retrofitting might work:
Small AI inference clusters (not training)
Facilities originally built with high ceilings and heavy floor loading
Partial conversion of a few racks, with traditional workloads moved elsewhere
Uptime’s recommendation:
If you must retrofit, conduct a detailed feasibility study covering floor strength, equipment access (lifts, corridors), and the ability to add gray space. Then consider a pilot with a single row or hall.
Construction Oversight: Milestones and Inspections
To prevent design‑to‑build drift, Uptime Institute recommends independent oversight at specific milestones.
Critical Inspection Points
Civil/structural complete (before equipment delivery) - Verify: Floor loading, ceiling height, slab flatness, trench locations, leak containment provisions
Critical infrastructure distribution pathways - Verify: Busway/ cable tray routing, pipe routing (separation from electrical), manifold placements
Equipment placement - Verify: Weight distribution, anchoring, clearances for maintenance, alignment with CDU and rack layouts
Functional component testing (Level 3) - Verify: Individual operation of pumps, CDUs, UPS, generators, chillers manifold placements
Post‑installation, pre‑commissioning - Verify: No damage during handling, all connections torqued, fluid loops filled and pressure‑tested
The Cost of Late Detection
Uptime illustrates the economic case for early oversight:
Caught during construction: Change order cost = $10k–$50k, schedule impact = days.
Caught during commissioning: Re‑work cost = $100k–$500k, schedule impact = weeks.
Caught during operations: Outage + re‑work + lost revenue = $1M+, schedule impact = months.
Independent third‑party oversight provides an unbiased set of eyes at each milestone, catching deviations before they become embedded.
Summary: Technical Construction Must‑Haves for AI Facilities
Based on Uptime Institute’s Paper 3, your construction plan must include:
Structural verification – Floor loading >2,000 kg per rack, solid slab or braced raised floor, higher ceilings, wide corridors, heavy‑duty lifts.
Liquid cooling infrastructure – Double‑containment piping, leak detection, sloped floors or trenches, separate gray space for CDUs.
Hybrid cooling coordination – Independent verification of liquid and air subsystems before integration.
Phased construction – Build in 15 MW or 12 MW phases, each with independent power and cooling. Delay large plant commitments until needed.
Retrofit caution – Full AI training retrofits are rarely feasible; consider new build for large clusters.
Milestone inspections – Independent oversight at structural completion, distribution pathway completion, equipment placement, and functional testing.
Early deviation correction – Catch issues during construction, not commissioning or operations