Buyer's Guide_
How to Choose a Data Center Liquid Cooling Company
A field-proven buyer's guide for AI data-center operators selecting a liquid-cooling integration partner, covering the specific technical, operational, and reliability criteria that separate competent crews from costly failures.
Key facts
- Direct-to-chip liquid cooling for NVL72 racks typically removes 80-90% of GPU heat, requiring coolant supply temperatures between 25-35°C per OEM spec.
- A single NVL72 rack can dissipate over 100 kW; a 10-rack cluster needs a cooling distribution unit (CDU) with redundancy and a secondary fluid loop rated for that total thermal load.
- Coolant purity must meet the OEM's specified conductivity (typically <1 µS/cm) and particle count (e.g., <100 particles/mL >5 µm); field testing with a calibrated particle counter is mandatory before commissioning.
- Leak-detection systems must be installed at every quick-connect, manifold, and GPU cold-plate; a single undetected drip can short a high-end GPU module within minutes, potentially causing tens of thousands of dollars in damage.
- The relevant standard for liquid-cooled IT equipment is ASHRAE TC 9.9 (2021) for allowable coolant temperatures and humidity; for safety, IEC 62477-1 applies to power interfaces.
- Field-terminated coolant hoses must be cut square and deburred; a burr or angled cut causes o-ring failure and leaks that can appear within weeks.
- Pressure testing the entire secondary loop to 1.5x the maximum operating pressure (per ASME B31.9) and holding for 30 minutes with zero drop is the industry-accepted leak-check protocol.
Verify Coolant Chemistry and Purity Capabilities
The single most common cause of liquid-cooling field failures is contaminated or incorrect coolant. A qualified partner must demonstrate they can test and maintain coolant to the GPU OEM's published specifications—typically deionized water with a biocide and corrosion inhibitor, conductivity below 1 µS/cm, and particle counts under 100 particles per milliliter above 5 microns. They should bring a calibrated particle counter and conductivity meter to every site, not rely solely on the CDU's built-in sensors.
Ask for their written coolant-handling procedure: how they flush the loop, how they sample at multiple points, and how they document baseline chemistry. A crew that skips this step or uses tap water will cause galvanic corrosion in cold plates and manifolds, leading to leaks and GPU failures within 6-12 months. The best integration partners include a full coolant analysis and certification in every deployment, because the cost of a shortcut is far higher than the effort to do it right.
Assess Leak-Detection and Containment Strategy
No matter how well the hoses are crimped, leaks happen—from a loose quick-connect, a micro-crack in a manifold, or a seal that wasn't seated. The integration partner must have a documented, multi-layer leak-detection plan: spot sensors under every cold plate and quick-connect, rope sensors along hose trays, and floor sensors in the aisle. They should also specify a containment strategy—drip trays under each rack and a secondary containment trough for the CDU—because even a small amount of coolant on a live GPU row can cause catastrophic downtime.
Demand to see their leak-response protocol: what triggers automatic CDU shutdown, how they isolate the affected rack, and how they dry and re-commission. A partner who says 'we've never had a leak' is either inexperienced or lying. The best crews have a written procedure and have practiced it. They also pressure-test every secondary loop to 1.5x the maximum operating pressure per ASME B31.9, hold for 30 minutes, and log the results—no exceptions.
Evaluate Rack-Level Integration and Cable Management
Liquid cooling doesn't eliminate cable management—it adds another layer. The GPU rack still has power cables, fiber MPO trunk cables for the scale-out network (InfiniBand or Ethernet), and now coolant hoses. A competent partner plans the hose routing to avoid crossing fiber paths, because a coolant leak onto an MPO connector causes immediate optical loss and requires a full cleaning and re-termination. They also ensure hoses have a minimum bend radius (per the hose manufacturer's spec) and are supported every 18-24 inches to prevent sagging and stress on quick-connects.
Ask to see their rack-level CAD or BIM model showing hose and cable paths before installation. If they can't produce one, they're guessing. The best crews also color-code hoses (supply vs. return) and label every connection point. This matters during maintenance: a technician should be able to trace a hose from CDU to cold plate without a schematic.
Check Commissioning and Testing Procedures
Commissioning is where most liquid-cooling problems are caught—or missed. The partner should have a written commissioning checklist that includes: pressure test the secondary loop, flush and fill with certified coolant, run the CDU at full flow for a full day while monitoring pressure and temperature, then check every quick-connect for leaks with a thermal camera or dry paper. They should also verify the GPU cold-plate temperatures under load using the BMC or nvidia-smi, ensuring they stay within the OEM's specified range (typically 25-35°C inlet).
Do not accept a verbal 'it's all good.' Demand written test results for every rack: pressure test log, coolant chemistry report, thermal images of all connections, and a final sign-off sheet. A partner who can't produce these documents is not a partner—they're a liability. The commissioning phase is also the time to train your on-site facilities team on coolant handling, leak response, and CDU alarms. If the partner doesn't offer this training, find one who does.
Common Failure Modes in the Field and How to Catch Them
The most frequent failure is a loose quick-connect that wasn't fully seated during installation. It doesn't leak immediately—it drips slowly for days until the coolant finds a power cable or GPU backplane. Catch it by pressure-testing every connection at 1.5x operating pressure for 30 minutes, then running a thermal soak over a full day with the CDU at full flow and checking all connections with a thermal camera. A cold spot on a connection indicates a leak path.
Second, coolant contamination from improper flushing. If the loop wasn't flushed with deionized water before filling, residual debris from hose cutting or manifold assembly will clog cold-plate microchannels, causing GPU throttling and eventual failure. Prevent this by requiring a particle count test after flushing and before final fill.
Third, hose routing that violates bend radius, causing kinks and flow restriction. This shows up as high pressure drop across the rack and higher GPU temperatures. Catch it during the pre-installation walkthrough by verifying hose paths against the manufacturer's bend radius spec. A partner who ignores these details will cost you in downtime and GPU replacements.
Demand Service-Level Agreements for Response and Spares
Liquid cooling is not set-and-forget. Pumps fail, valves stick, sensors drift. Your integration partner should offer a service-level agreement (SLA) that specifies response time for critical failures, spare parts availability, and scheduled maintenance intervals such as quarterly coolant sampling and filter replacement. Without an SLA, you're gambling that nothing breaks—and in a high-power rack, something will.
Ask for their spare parts inventory list and where it's stored. A partner who keeps spares at a distant central warehouse is not responsive. The best crews maintain a local stock of the most common failure items: quick-connects, hose segments, o-rings, and a spare CDU pump. They also have a remote monitoring system that alerts them to pressure or temperature anomalies before you notice. Some leading integration partners provide a 24/7 NOC with direct access to field engineers, because a delay in detecting a leak can mean a significant loss.
Standards referenced: ASHRAE TC 9.9 (2021) – Thermal Guidelines for Data Processing Environments · IEC 62477-1 – Safety requirements for power electronic converter systems and equipment · ASME B31.9 – Building Services Piping (pressure testing protocol)
Frequently asked_
What coolant should I use for NVL72 direct-to-chip cooling?
The GPU OEM specifies the coolant chemistry in their installation manual. Typically it's deionized water with a biocide and corrosion inhibitor, with conductivity below 1 µS/cm and particle count under 100 particles/mL >5 µm. Never use tap water or automotive coolant—they cause galvanic corrosion and void warranties. Your integration partner should provide a coolant certification report after filling.
How do I know if a liquid-cooling partner is competent?
Ask for their written commissioning checklist, coolant handling procedure, and leak-response protocol. A competent partner will show you pressure test logs, coolant chemistry reports, and thermal images from previous deployments. They should also have a local spare parts inventory and a 24/7 monitoring NOC. Leviathan Systems, for example, provides a full coolant certification and 24/7 NOC with direct field engineer access. If they can't produce documentation, they're not ready for your site.
Can I mix liquid cooling from different vendors in the same rack?
No—mixing cold plates, manifolds, or quick-connects from different vendors risks incompatible materials, different pressure ratings, and different coolant chemistry requirements. Stick with a single vendor for the entire rack's liquid cooling loop. Your integration partner should source all components from the same OEM to ensure compatibility and warranty coverage.
What is the most common mistake in liquid-cooling installation?
Not fully seating quick-connects. They can appear locked but still have a tiny gap that leaks over time. Always pressure-test every connection to 1.5x operating pressure and hold for 30 minutes, then run a thermal soak over a full day with thermal camera inspection. Also common: cutting hoses at an angle or leaving burrs, which causes o-ring failure. Use a square cutter and deburr every cut.
How often should I replace coolant in a liquid-cooled GPU rack?
Most OEMs recommend coolant sampling every 3-6 months and full replacement every 2-3 years, depending on water quality and system usage. Your integration partner should provide a schedule for quarterly conductivity and particle count checks. If conductivity rises above 1 µS/cm or particle count spikes, replace the coolant immediately to prevent corrosion and clogging.