Buyer's Guide_
Data Center Rack-and-Stack Services for GPU Builds: What's Included
A field engineer's guide to scoping and executing rack-and-stack services for GPU AI clusters, covering what's included, what's not, and how to avoid costly scope gaps in NVL72-class deployments.
Key facts
- Rack-and-stack for GPU builds includes mechanical installation, power wiring, structured cabling, and initial power-on testing—but excludes fiber termination, liquid loop filling, and network commissioning unless explicitly scoped.
- A standard GPU rack (e.g., NVL72) weighs over 1,500 kg fully loaded; floor loading must be verified per ASHRAE TC 9.9 before rack placement.
- MPO trunk cables for scale-out networks (InfiniBand/Ethernet) are factory-terminated; field work is patching, routing, cleaning, inspection, and testing—never field-crimping.
- Copper NVLink spine inside NVL72 racks is pre-installed by the OEM; rack-and-stack crews verify seating and torque per OEM spec but do not terminate or replace it.
- Structured cabling for GPU clusters follows TIA-942-B or ISO/IEC 24764 for pathway fill ratios and bend radius; exceeding 40% fill in a 4-inch ladder tray requires a second tray.
- Power wiring for GPU racks typically requires 3-phase 480 VAC at 60 A per rack; torque values for busbar connections are specified by the PDU manufacturer and must be verified with a calibrated torque wrench.
- Commissioning includes booting the baseboard management controller (BMC), verifying PSU status, and running a basic GPU discovery via nvidia-smi—but full cluster validation (e.g., NCCL tests) is a separate service.
Scope Boundaries: What Rack-and-Stack Includes and Excludes
Rack-and-stack for GPU AI clusters is a mechanical and electrical installation service. It covers unloading the rack from the truck, positioning it on the data center floor, installing slide rails and shelves, seating GPU servers (e.g., HGX baseboards or DGX nodes), connecting power cables from the PDU to each server PSU, and routing structured cabling (fiber MPO trunks, DACs, or AOCs) between switches and servers per the cabling matrix. It also includes initial power-on testing: verifying that each server's BMC is reachable, PSUs show green, and nvidia-smi detects all GPUs.
What it does not include: fiber termination (MPO connectors are factory-polished), liquid cooling loop filling or leak testing (separate service), network switch configuration, InfiniBand subnet manager setup, or cluster-level validation like NCCL all-reduce benchmarks. These are typically scoped as separate commissioning or integration services. A common mistake is assuming rack-and-stack covers cable testing; it does not—OTDR or MPO continuity testing is an add-on. Scope clearly in the SOW: 'Install, route, and label all cables per the provided cable matrix; perform visual inspection and continuity check with a calibrated MPO tester.'
Mechanical Installation: Rack Positioning, Leveling, and Floor Loading
Before a rack touches the floor, verify the data center's floor loading capacity. A fully loaded NVL72 rack with 72 GPUs, switches, and PDUs can exceed 1,500 kg. Per ASHRAE TC 9.9, raised-floor tiles must be rated for the point load—typically 1,200 kg per tile for heavy equipment. If the rack sits on a concrete slab, verify the slab thickness and reinforcement. Use a laser level to ensure the rack is plumb and level within the tolerance specified by the rack manufacturer (typically 1–2 mm per meter); an unlevel rack stresses server rails and can cause binding during insertion.
For GPU racks with liquid cooling, the rack must be positioned within the distance specified by the cooling system design (often 1 meter or less) of the facility coolant supply and return lines. Use a pallet jack or rack dolly rated for the weight; never drag a rack across the floor—it damages tiles and can shear bolts. Secure the rack to the floor using seismic brackets if required by local code (e.g., IBC 2018 in seismic zones). Document the installation with photos of the level bubble and torque values on floor anchors.
Power Wiring: PDU Connections, Torque, and Phase Balancing
GPU racks draw high current—typically 60 A at 480 VAC 3-phase per rack for an NVL72. Each server PSU is connected to a PDU outlet via a C19-to-C19 or C21-to-C21 power cord, depending on the PSU type. The PDU itself is fed by a hardwired busway or whip from the facility's power distribution. Critical: torque the busbar connections to the PDU manufacturer's specified value (commonly between 20 Nm and 40 Nm for M8 bolts; verify the exact value in the installation manual). Under-torquing causes resistive heating; over-torquing strips threads. Use a calibrated torque wrench and mark each bolt with a torque seal.
Phase balancing is essential. A 3-phase PDU serving 60 A per phase must have the load spread evenly across phases to avoid neutral current exceeding 50% of phase current per NEC 310.15(B)(5). For GPU racks, each server's PSUs are dual-feed; connect half to phase A and half to phase B, or use a phase rotation schedule. Label each power cord with the server slot and PSU number. After wiring, measure voltage at each PSU inlet with a multimeter—should be within ±5% of nominal. Document all readings.
Structured Cabling: Routing, Bend Radius, and Labeling for MPO Trunks
Scale-out network cabling in GPU clusters uses MPO-12 or MPO-24 trunk cables (OS2 single-mode or OM4 multimode) between leaf switches and GPU servers. These cables are factory-terminated and polished; field work is routing, patching, and testing. Per TIA-942-B, the minimum bend radius for a 12-fiber MPO trunk is 10 times the cable diameter under load and 15 times during installation. For typical 12-fiber cables (around 3 mm diameter), that translates to 30 mm installed, 45 mm during pull, but always defer to the cable manufacturer's spec. Exceeding this causes micro-bends that increase attenuation and can lead to link errors.
Route cables in overhead ladder trays or underfloor baskets. Maximum fill ratio per ISO/IEC 24764 is 40% for a single tray; if you exceed that, add a second tray. Use vertical cable managers on rack sides to manage slack. Label both ends of every cable with a machine-printed label that includes source port, destination port, and cable ID (e.g., 'SW1-P1 to RACK3-SVR4-P1'). After routing, test each MPO trunk with a calibrated MPO continuity tester or an OLTS for insertion loss. Acceptable loss per TIA-568.3-D for single-mode MPO is ≤0.75 dB at 1310 nm. Reject any cable that exceeds this.
Common Failure Modes: What Goes Wrong in the Field and How to Catch It
The most frequent failure in GPU rack-and-stack is loose power connections. A busbar bolt torqued below the manufacturer's spec can develop a hot spot over weeks, melting the insulator and causing a phase-to-ground fault. Catch it by using a torque wrench and marking each bolt. Second: bent MPO pins. When an MPO connector is inserted at an angle, the guide pins can shear off, causing permanent damage. Always inspect the connector endface with a 200x or 400x microscope before insertion—per IEC 61300-3-35, a single scratch >3 µm wide on the core is a fail. Third: cable tension. A trunk cable pulled tight across a sharp edge can have its jacket cut, exposing fibers. Inspect all cable pathways after installation; use a cable tension meter if pulling long runs.
Fourth: incorrect server seating. A GPU server not fully seated in its rails can have the backplane connector misaligned, causing intermittent NVLink errors. Verify that each server clicks into the backplane and that the retention latches are engaged. Fifth: labeling errors. A mislabeled cable can cause days of troubleshooting during network commissioning. Implement a two-person verification: one reads the label, the other checks the port. Finally, static discharge. GPU servers are ESD-sensitive; use grounded wrist straps and ESD-safe work surfaces. A single ESD event can damage a GPU's memory controller, leading to silent data corruption.
Commissioning: Power-On, BMC Check, and GPU Discovery
After all servers are seated and cabled, perform a controlled power-on. Start by energizing the PDUs and verifying that each PDU's input voltage and current are within spec. Then power on one server at a time, watching for PSU fan spin-up and status LED (green = OK, amber = fault). For each server, connect to the BMC via the dedicated management port (usually a 1 GbE RJ45). Verify that the BMC responds to ping and that the web interface or CLI shows all PSUs as 'present' and 'input OK'.
Next, boot the server into the BIOS or OS installer and run nvidia-smi. It should list all GPUs (e.g., 8 GPUs for an HGX baseboard). If a GPU is missing, check the PCIe slot seating and power cables. Document each server's BMC MAC address, GPU serial numbers, and firmware versions. This baseline is critical for later RMA claims. Do not proceed to cluster-level tests (NCCL, storage I/O) unless that is explicitly scoped—those are separate services. Hand off the completed rack to the network and storage teams with a signed checklist.
Scope Gaps: Liquid Cooling, Fiber Termination, and Network Commissioning
Rack-and-stack services for GPU builds often exclude liquid cooling loop installation. For NVL72 racks with direct-to-chip cooling, the facility coolant lines must be connected to the rack's manifold, the loop filled with dielectric fluid, leak-tested at the pressure specified by the manufacturer (typically 1.5× operating pressure), and bled of air. This requires specialized training and tools (pressure test kit, vacuum pump, fluid analysis). If your SOW says 'rack and stack' but the rack has liquid cooling, you need a separate scope for 'liquid cooling loop commissioning.'
Similarly, fiber termination is not part of rack-and-stack. MPO trunks are factory-terminated; field work is patching and testing. However, if the data center uses field-terminated connectors (e.g., for custom-length runs), that requires a fusion splicer and a certified technician—a different service. Network commissioning—configuring leaf/spine switches, setting up InfiniBand subnet managers, and running NCCL tests—is a separate engagement. A typical hyperscaler in Texas might have a dedicated network integration team that takes over after rack-and-stack. Clarify these boundaries in the SOW to avoid change orders.
Standards referenced: ASHRAE TC 9.9 (thermal guidelines for data centers) · TIA-942-B (telecommunications infrastructure standard for data centers) · ISO/IEC 24764 (generic cabling for data centers) · TIA-568.3-D (optical fiber cabling components standard) · IEC 61300-3-35 (fiber optic connector endface inspection) · NEC 310.15(B)(5) (neutral conductor sizing for nonlinear loads) · IBC 2018 (International Building Code for seismic restraints)
Frequently asked_
Does rack-and-stack include testing the fiber cables after installation?
Not by default. Standard rack-and-stack includes visual inspection of MPO connectors and a continuity check with a calibrated MPO tester, but not insertion loss testing with an OLTS or OTDR. If you need loss measurements to TIA-568.3-D limits (≤0.75 dB for single-mode MPO), you must add that as a separate line item. Many hyperscalers require this for warranty acceptance.
What is the typical timeline for rack-and-stack of a single NVL72 rack?
A crew of two experienced technicians can complete mechanical installation, power wiring, and structured cabling for one NVL72 rack in 8–12 hours, assuming the rack is pre-assembled and cabling is pre-terminated. Liquid cooling loop commissioning adds 4–6 hours. Network commissioning (switch config, NCCL tests) is separate and can take 1–2 days per rack.
How do I ensure the rack-and-stack vendor doesn't damage the MPO connectors?
Specify in the SOW that all MPO connectors must be inspected with a 200x or 400x microscope per IEC 61300-3-35 before insertion. Require that technicians use dust caps on all connectors when not in use, and that cables are routed with a minimum bend radius of 10 times the cable diameter (per TIA-942-B). Leviathan Systems includes these steps as standard practice.
What documentation should I expect from a rack-and-stack service?
You should receive a signed checklist for each rack: rack leveling measurements, torque values for all power connections, voltage readings at each PSU inlet, BMC MAC addresses, GPU serial numbers, and a cable matrix showing every cable's source and destination. Photos of the installation (rack front, rear, cable pathways) are also standard. This documentation is critical for RMA and troubleshooting.
Can rack-and-stack include liquid cooling loop filling?
Only if explicitly scoped. Most rack-and-stack vendors do not handle liquid cooling because it requires specialized training, pressure test equipment, and fluid handling procedures. If your rack has direct-to-chip cooling, you need a separate liquid cooling commissioning service. Leviathan Systems offers this as an add-on for NVL72 builds.