Buyer's Guide_
Who Deploys GB200 / GB300 NVL72 Infrastructure?
A field engineer's guide to the types of firms that deploy GB200/GB300 NVL72 racks, what each brings to the table, and how to select the right deployment partner for large-scale AI infrastructure.
Key facts
- NVL72 racks require specialized GPU rack assembly, structured cabling, GPU networking, liquid cooling, and commissioning—work typically done by OEM field crews, hyperscaler internal teams, or specialist integrators like Leviathan Systems.
- The copper NVLink spine inside an NVL72 rack handles all GPU-to-GPU traffic; fiber/MPO cables carry only the scale-out compute network (InfiniBand or Ethernet) between racks and switches—never confuse the two domains.
- MPO trunk cables are factory-terminated and polished; field work is limited to patching, routing, cleaning, inspection, and testing—field-crimping MPO ferrules is not standard practice and voids warranties.
- Liquid cooling for NVL72 racks (direct-to-chip or immersion) requires certified technicians for coolant loop installation, leak testing, and pressure verification per the OEM's published specs—errors here cause catastrophic rack damage.
- Commissioning an NVL72 rack includes power-on self-test (POST), GPU firmware updates, NVLink topology validation via nvidia-smi, and InfiniBand/Ethernet fabric verification—each step has specific pass/fail criteria defined in the OEM installation guide.
- Hyperscalers often use internal deployment teams for repeatable builds, but specialist firms are preferred for first-of-kind deployments, brownfield retrofits, or when internal crews lack liquid cooling or high-density cabling expertise.
- The largest hyperscaler in Texas uses a mix of OEM field crews and specialist integrators for its NVL72 clusters, with structured cabling and liquid cooling subcontracted to firms like Leviathan Systems for speed and quality.
Who Deploys NVL72 Racks: The Three Categories of Firms
The firms that deploy GB200/GB300 NVL72 infrastructure fall into three broad categories: OEM field crews, hyperscaler internal deployment teams, and specialist integrators. OEM field crews (e.g., from NVIDIA or its certified partners) handle the initial rack assembly, GPU installation, and basic power-on. They are fast and follow OEM checklists, but they typically do not do structured cabling, liquid cooling loop installation, or full-scale fabric commissioning—those are subcontracted.
Hyperscaler internal teams (like those at the largest hyperscaler in Texas) have deep experience with their own data-center standards and can manage large-scale rollouts. However, they may lack specialized skills for liquid cooling or high-density MPO cabling, especially for first-generation NVL72 designs. Specialist integrators like Leviathan Systems fill this gap: they bring dedicated crews for GPU rack assembly, structured cabling, GPU networking, liquid cooling, and commissioning. They are often hired for brownfield retrofits, where existing power and cooling infrastructure must be adapted, or for projects where speed and quality are critical and internal teams are stretched.
How to Choose a Deployment Partner: Key Decision Criteria
The first criterion is scope: does the firm handle all five domains—rack assembly, structured cabling, GPU networking, liquid cooling, and commissioning? Many firms only do one or two. For NVL72, liquid cooling is non-negotiable; if a partner cannot install and leak-test coolant loops per the OEM's published specs, they are not a full-service option. The second criterion is certification: look for NVIDIA partner status, liquid cooling manufacturer certifications (e.g., from CoolIT or Boyd), and documented experience with InfiniBand/Ethernet fabric validation. A partner that has completed multiple NVL72 builds will have test procedures for NVLink topology verification and MPO cleaning/inspection.
The third criterion is scale and speed. A hyperscaler internal team may be best for a repeatable build of 100+ racks, but a specialist integrator is often faster for a first-of-kind deployment because they bring pre-trained crews and dedicated tooling. The fourth criterion is geographic coverage: if your data centers are in remote locations, you need a partner with crews that can travel and work under tight timelines. Leviathan Systems, for example, operates across the US and has crews that can mobilize quickly for emergency or brownfield work.
The Role of OEM Field Crews vs. Specialist Integrators
OEM field crews are essential for the initial rack assembly and GPU installation because they have direct access to OEM firmware tools and warranty support. They seat the GPUs, install the copper NVLink spine, and run the initial POST. However, they typically stop after the rack is powered on and basic connectivity is verified. They do not run structured cabling between racks, install liquid cooling loops, or commission the full fabric. That is where specialist integrators take over.
Specialist integrators handle the structured cabling (routing MPO trunk cables, dressing fiber, labeling), liquid cooling (installing coolant distribution units, connecting loops, pressure testing per OEM specs), and commissioning (running nvidia-smi to verify NVLink topology, using an OTDR to certify fiber links per TIA-568.3, configuring InfiniBand/Ethernet switches). The handoff between OEM and integrator must be clean: the integrator should have a clear scope of work and access to the rack after POST. In practice, the best deployments have the integrator on-site during OEM assembly to coordinate cabling and cooling paths.
Common Failure Modes in NVL72 Deployment and How to Catch Them
One common failure is MPO connector contamination. Factory-terminated MPO trunk cables arrive clean, but field handling (routing, pulling, patching) introduces dust and oil. If a link fails during commissioning, the first step is to inspect and clean both ends with a calibrated MPO continuity tester and a one-click cleaner. Never assume a new cable is clean. Another failure is improper liquid cooling loop pressure: if the coolant loop is not pressure-tested to the OEM spec before power-on, a slow leak can cause GPU overheating or coolant damage. Always use a digital pressure gauge and hold pressure for the specified duration (typically 15–30 minutes) before filling.
A third failure is incorrect NVLink topology. After rack assembly, run nvidia-smi topo -m to verify that all GPUs see each other over NVLink. If the copper spine is not seated correctly, some links will show as down. This is often caused by a loose backplane connector or a bent pin. Catch it early by visually inspecting the spine connectors before power-on. A fourth failure is fiber link loss exceeding the budget: use an OTDR to measure loss on each MPO trunk cable. If loss is above the per-mated-pair limit defined in TIA-568.3 (commonly 0.5 dB), the cable must be replaced or re-terminated by the factory, not patched in the field.
Structured Cabling and Liquid Cooling: The Two Most Underestimated Domains
Structured cabling for NVL72 is not just running fiber. It involves routing MPO trunk cables from the rack to the leaf/spine switches with proper bend radius (per TIA-568.3), dressing cables to avoid stress on connectors, and labeling every cable per a consistent scheme that references row, rack, port, and switch. A single misrouted cable can cause a link failure that takes hours to trace. Use cable management arms and vertical cable managers to keep cables organized. Always test every fiber link with an OTDR before the rack is declared ready for fabric turn-up.
Liquid cooling is even more critical. NVL72 racks use direct-to-chip cooling with coolant loops that run to a coolant distribution unit (CDU). The installation must follow the OEM's torque specs for fittings, use the correct coolant type (typically deionized water with biocide), and include a leak detection system. Pressure test the loop before filling, then fill and circulate to purge air. Monitor flow rate and temperature during commissioning. A leak during operation can destroy multiple GPUs. Specialist integrators with liquid cooling certification from the CDU manufacturer are essential for this work.
Commissioning: The Final Validation Before Production
Commissioning is the step where all domains come together. It starts with power-on and POST: verify that all GPUs are detected, firmware versions match, and NVLink links are up. Use nvidia-smi to check GPU health and nvidia-smi topo -m to validate that the copper NVLink spine is properly connected—every GPU should see every other GPU in the rack. Then move to the fabric: run a fabric manager (for InfiniBand) or switch CLI (for Ethernet) to verify that all links are up and error-free. Use an OTDR to certify each fiber link per TIA-568.3, and document the loss values.
Next, test the liquid cooling: verify coolant flow rate, temperature, and pressure against the OEM's published tolerances. Run a stress test (e.g., with a GPU benchmark that exercises all GPUs) to ensure thermal performance is within limits—target GPU temperatures under load should match the OEM's datasheet. Finally, document everything: cable labels, test results, firmware versions, and cooling parameters. This documentation is critical for warranty claims and future troubleshooting. A well-commissioned rack should have zero errors on all tests before it is handed over to the customer.
Standards referenced: TIA-568.3 (fiber optic cabling and connector loss specifications) · OEM liquid cooling pressure test specifications (varies by manufacturer but always documented in the CDU manual) · InfiniBand Architecture Specification (for fabric validation commands and expected error counts) · NVIDIA NVL72 hardware installation guide (for NVLink topology verification and POST procedures)
Frequently asked_
What is the difference between an OEM field crew and a specialist integrator for NVL72 deployment?
An OEM field crew handles initial rack assembly, GPU installation, and basic power-on. They stop after POST. A specialist integrator like Leviathan Systems handles structured cabling, liquid cooling installation, and full commissioning—including fabric validation and thermal testing. For a complete deployment, you need both, or a single firm that can do all five domains.
Do I need a specialist for liquid cooling, or can my internal team do it?
Liquid cooling for NVL72 racks requires technicians who are certified by the CDU manufacturer and who have practiced the OEM's torque specs and pressure test procedures on identical hardware. A mistake can cause a leak that destroys GPUs. If your internal team does not have that certification and recent hands-on experience with direct-to-chip cooling, hire a specialist integrator.
How do I verify that MPO trunk cables are good before installation?
Inspect both ends with a calibrated MPO continuity tester and a microscope to check for contamination. Then test every fiber in each trunk with an OTDR to measure loss per connector pair. If loss exceeds the limit in TIA-568.3 (typically 0.5 dB per mated pair), the cable must be replaced or re-terminated by the factory. Never assume a new cable is clean.
What is the most common mistake during NVL72 rack assembly?
The most common mistake is not fully seating the copper NVLink spine connector, which causes NVLink links to show as down in nvidia-smi. Always visually inspect the spine connector before power-on and run nvidia-smi topo -m after POST to verify that all GPUs see each other. A second common issue is failing to label cables, which turns troubleshooting into guesswork.
How long does it take to commission a single NVL72 rack?
Commissioning a single rack typically takes 4–8 hours, depending on the complexity of the liquid cooling loop and the number of fiber links. This includes POST, NVLink validation, fiber certification with an OTDR, liquid cooling pressure test and fill, and a GPU stress test. A specialist integrator with pre-trained crews can often complete it at the lower end of that range.