Vertiv NVIDIA DSX Ready CDU Qualification Sets a New Cooling Benchmark

Vertiv has qualified a 2.3 MW coolant distribution unit under NVIDIA DSX Ready, giving AI data center builders a new benchmark for liquid-cooling procurement. The company says its CoolChip model is the first CDU qualified specifically through this program.

The milestone matters because buyers no longer select accelerators, power systems, and cooling equipment as separate infrastructure layers. Higher rack densities force those systems into one engineering decision. NVIDIA is now extending its influence from computing hardware into the facilities that support it.

However, qualification is not equivalent to approval for a particular building. NVIDIA uses a self-qualification suite for CDU products, while each operator remains responsible for facility design and operating validation. That distinction creates the central tension behind the Vertiv NVIDIA DSX Ready CDU announcement.

What Vertiv’s 2.3 MW Qualification Actually Changes

The qualification gives buyers a common screening framework, but it does not deliver a finished cooling design.

Vertiv announced the qualification on September 21, 2026. The product is a 2.3 MW model from the company’s CoolChip coolant distribution unit family.

A coolant distribution unit, or CDU, controls the flow and temperature of liquid that removes heat from computing equipment. It also separates the facility water loop from the cleaner secondary loop serving servers.

This separation protects sensitive cold plates and server components from contaminants in facility water. It also lets operators control coolant temperature, pressure, filtration, and flow closer to the computing load.

According to the Vertiv qualification, the 2.3 MW model supports direct-to-chip cooling and rear-door heat exchangers. Direct-to-chip systems circulate liquid through cold plates attached to processors. Rear-door systems remove heat as air exits a rack.

The unit uses a liquid-to-liquid design. Instead of rejecting heat directly into room air, it transfers heat between the server-side coolant loop and a facility-side water loop.

Vertiv designed the model for installation at the end of a row or along a data hall perimeter. Those positions let one CDU serve several high-density racks without occupying space inside every rack.

Its 2.3 MW rating describes available thermal capacity under specified conditions. It does not represent a newly installed 2.3 MW computing deployment. It also does not identify the number of racks that every installation can support.

That number depends on rack heat loads, water temperatures, required flow, redundancy, and operating margins. A design serving lower-density racks will distribute capacity differently from one supporting next-generation accelerator systems.

The NVIDIA DSX Ready program gives builders a way to identify products that meet applicable requirements within NVIDIA’s broader AI factory architecture. For CDU suppliers, qualification uses NVIDIA’s CDU self-qualification suite.

The word “applicable” is important. A qualified CDU meets relevant functional requirements for its product category. It does not certify every pipe, pump, chiller, sensor, control system, and rack connection around it.

Vertiv also says the product is intended for edge, colocation, and hyperscale deployments. Those markets have very different operating models.

A hyperscaler can standardize complete buildings around one computing architecture. A colocation provider must often accommodate several customers with different rack designs and service agreements. Edge sites can face strict space and staffing limits.

The qualification therefore changes the beginning of procurement more than the end. It helps engineering teams create a credible shortlist. Detailed site analysis still determines whether the selected CDU fits the actual deployment.

This distinction prevents the news from becoming a simple certification story. Vertiv has gained a recognized place inside NVIDIA’s infrastructure framework. Buyers still carry the engineering risk once a product enters their facility.

Why NVIDIA DSX Ready Cooling Matters Now

AI computing has turned cooling compatibility into a system-level constraint rather than a secondary facilities question.

NVIDIA describes an AI factory as infrastructure designed to produce AI outputs at scale. The concept joins compute, networking, software, power, cooling, and building systems into one operating model.

The DSX reference design organizes those elements into repeatable scalable units. Each unit combines compute areas with supporting networking, power, cooling, and facility infrastructure.

This approach reflects a practical limit on accelerator deployments. More GPUs do not create useful capacity when a facility cannot deliver enough power or remove enough heat.

Traditional enterprise data centers often treated cooling as a room-level service. Computer room air handlers supplied cold air, while server fans moved that air across components.

High-density AI racks change that arrangement. Air can still cool networking equipment and other components, but the hottest processors increasingly require liquid delivered close to the silicon.

NVIDIA’s published DSX guidance specifies a liquid-to-liquid CDU form factor. It also calls for technology cooling system flow of at least 1.5 liters per minute per kilowatt and an N+1 CDU redundancy approach.

N+1 means an installation includes one more unit than the minimum needed for its planned load. That spare capacity allows maintenance or a single equipment failure without immediately losing the entire cooling service.

The guidance also shows cabinet thermal design power increasing across computing generations. NVIDIA lists cabinet values from 198 kW for one MGX generation to 330 kW for Vera Rubin NVL72 systems.

These are reference-design figures rather than universal operating loads. Still, they illustrate why megawatt-class cooling equipment is entering facility plans.

At 330 kW per cabinet, a few racks can consume a meaningful share of a 2.3 MW CDU’s nominal capacity. Engineers must also reserve headroom for transient conditions, maintenance, equipment degradation, and future changes.

NVIDIA is responding by standardizing more of the surrounding facility. Its framework gives partners shared assumptions for interfaces, redundancy, layout, and operational planning.

For buyers, that standardization can reduce early design uncertainty. A qualified component has already been evaluated against relevant functional requirements instead of arriving as an entirely independent specification.

For suppliers, the framework creates pressure to align product development with NVIDIA’s infrastructure roadmap. A cooling vendor can offer excellent equipment, yet still face friction if buyers cannot map that equipment into their selected AI architecture.

Vertiv Chief Product and Technology Officer Scott Armul summarized that dependency in the company’s announcement. He said compute, power, and cooling decisions can no longer be made independently.

The point is less promotional than it first appears. A cooling constraint can reduce available computing capacity even when every GPU is installed and powered.

Operators might cap processor performance, delay rack activation, or reserve more physical space for thermal equipment. They could also face unplanned construction if facility piping cannot support the necessary flow.

This is why the Vertiv NVIDIA DSX Ready CDU qualification matters beyond one equipment model. It places cooling products within the same compatibility conversation as AI servers and network systems.

It also gives NVIDIA more influence over the facilities supply chain. NVIDIA is not manufacturing the CDU, but its reference requirements can shape which products reach an AI factory shortlist.

That influence is valuable when GPU architecture moves faster than conventional data center construction. A server platform can change within a product cycle, while large facilities require years of planning and permitting.

Reference designs help bridge those schedules. They let developers plan repeatable infrastructure blocks around defined assumptions, then adapt individual elements for local conditions.

The model is attractive, but it also concentrates attention on NVIDIA’s definition of an AI factory. Operators supporting several accelerator vendors must decide whether one reference architecture can remain sufficiently flexible.

Vertiv NVIDIA DSX Ready CDU Faces a Crowded Supplier Field

Vertiv gained an early qualification, but competing capacities and broader cooling portfolios prevent it from owning the category.

The most direct pressure comes from other DSX-aligned cooling suppliers. LG Electronics, Schneider Electric’s Motivair business, and LiquidStack have also appeared in reporting about qualified CDU providers.

LG presents an especially clear capacity comparison. Its 2.6 MW CDU is listed as NVIDIA DSX Ready, placing its nominal rating above Vertiv’s 2.3 MW model.

The LG CDU portfolio includes 600 kW, 1 MW, and 2.6 MW equipment. LG describes the 2.6 MW system as a liquid-cooling product for mission-critical AI data centers.

Capacity alone does not identify the better product. Engineers also compare flow and pressure performance, operating temperatures, footprint, filter design, controls, maintainability, redundancy, and local service.

A larger nominal rating can reduce the number of units in one design. However, several smaller units can provide more granular redundancy and simplify maintenance in another design.

LG’s presence shows how quickly the competitive threshold is moving. Vertiv’s qualification is notable, but the industry is already comparing different megawatt-class options inside the same framework.

Schneider Electric adds another competitive route through Motivair. Its MCDU-70 provides 2.5 MW of cooling capacity and is designed for deployments that can scale beyond 10 MW.

Schneider describes a six-unit arrangement as one way to provide four active units and two redundant units for a 10 MW requirement. That example shows how CDU procurement expands into complete system architecture.

The company has also connected its cooling work with electrical design, operational software, and digital twins. Its DSX collaboration includes facility simulation and power-system tools alongside liquid cooling.

Vertiv competes with a similarly broad infrastructure proposition. It sells power, thermal management, modular systems, controls, and related services.

That breadth matters because operators increasingly want accountability across connected facility systems. A CDU fault might originate in a control setting, facility-water condition, server-side restriction, or upstream heat-rejection problem.

Vendors that understand the whole chain can help diagnose those interactions. They can also coordinate equipment schedules and service plans across more of a project.

Vertiv’s product page describes the CoolChip CDU as using redundant variable-speed pumps. It also lists a 50-micron filter, leak detection, and precise temperature control among the product family’s features.

Those details address practical operating requirements. Pumps must respond to changing computing loads, filters must protect server loops, and controls must avoid sudden coolant-temperature changes.

Still, product-page features remain vendor claims until tested within a customer’s configuration. A qualified component can behave differently when connected to unfamiliar cold plates, piping layouts, controls, and water systems.

LiquidStack represents another category of competitor. Its background in liquid and immersion cooling brings experience with thermal designs that extend beyond conventional room cooling.

The competitive contest therefore involves more than maximum megawatts. Suppliers are competing over integration depth, deployment speed, service coverage, controls, and compatibility with future server generations.

NVIDIA’s program can make those comparisons more structured. It cannot make them simple.

A standardized qualification may also reduce one traditional vendor advantage. Buyers can compare several products that satisfy a common baseline instead of relying entirely on a long-standing supplier relationship.

At the same time, qualification can favor larger vendors that already possess test facilities, engineering staff, global support, and established relationships with computing manufacturers.

Vertiv sits in that group. Its challenge is to turn qualification into selected designs and operating deployments before the designation becomes routine across competing portfolios.

The first qualified model receives attention. The long-term winner will be the supplier whose equipment performs predictably across changing loads and real maintenance events.

The Qualification Does Not Remove Site-Level Cooling Risk

DSX Ready narrows product uncertainty, while leaving the hardest integration decisions with the data center designer and operator.

Vertiv explicitly says qualification does not replace the engineering required for a particular site or configuration. Customers must determine how the CDU fits their facility design, operating requirements, and system configuration.

That disclaimer defines the limit of the announcement. It also identifies the work that remains after procurement teams recognize the badge.

The self-qualification structure deserves careful attention. NVIDIA supplies a CDU testing suite, but the program relies on providers to evaluate their specific offerings against applicable requirements.

This process gives suppliers a consistent technical target. It should not be read as an independent guarantee covering every field condition.

Buyers need access to test methods, boundary conditions, and complete performance curves. A single 2.3 MW label cannot describe performance at every supply temperature, return temperature, coolant mixture, or flow rate.

Thermal capacity also depends on the temperature difference across a system. Changes in that difference affect the flow required to transport the same amount of heat.

Pumps must provide that flow against resistance from pipes, valves, manifolds, cold plates, and heat exchangers. If pressure losses exceed design assumptions, the nominal CDU rating becomes less useful.

Water quality creates another risk. Facility loops can contain particles, corrosion products, biological material, or incompatible chemicals.

The CDU’s heat exchanger separates that water from the technology cooling system. However, each loop still needs suitable treatment, filtration, monitoring, and maintenance.

Materials compatibility matters on the server side. Metals, seals, hoses, coolants, and inhibitors must work together over years of operation.

A qualification cannot anticipate every combination deployed by every customer. Operators need approved material lists and clear ownership when components come from several suppliers.

Controls introduce another integration layer. The CDU must respond to server demand, pumps, valves, facility-water conditions, alarms, and backup equipment.

A poor control sequence can cause unstable temperatures or unnecessary pump energy. It can also produce conflicting commands when facility and server systems use different operational assumptions.

Redundancy must be tested as behavior, not counted only as equipment. An N+1 design works when the remaining units can assume the load during a failure or maintenance event.

That transfer must occur without violating server coolant limits. Operators need to test valve positions, control timing, electrical backup, communications, and alarm escalation.

Maintenance access can also undermine a strong diagram. Technicians need space to replace pumps, clean filters, isolate heat exchangers, and reach sensors without shutting down unrelated equipment.

Vertiv’s end-of-row or perimeter placement can improve access compared with an in-rack system. It also requires careful piping routes and leak-management planning.

The facility’s heat-rejection system remains a separate constraint. A CDU transfers heat, but it does not make that heat disappear.

Chillers, cooling towers, dry coolers, pumps, and local climate conditions determine where the thermal load goes next. Water availability and permitted operating temperatures can change the preferred design.

Efficiency claims require the same context. Liquid cooling can reduce fan work and enable warmer coolant loops, but total facility efficiency depends on the entire heat-removal chain.

A poorly matched chiller plant can erase gains achieved near the rack. Conversely, favorable climate conditions can support free cooling for more operating hours.

Future accelerator generations add planning uncertainty. A CDU installed today might serve several server refreshes, yet rack power and coolant requirements will continue changing.

NVIDIA’s reference documentation already illustrates that movement across product generations. Buyers need to know whether their piping, pumps, controls, and heat-rejection systems can support the next load profile.

The 2.3 MW rating offers useful scale, but scale can concentrate consequences. A failure affecting one large centralized CDU could influence several racks at once.

Redundancy, isolation, and rack-level controls reduce that exposure. Those protections consume space, capital, and engineering time that a headline capacity figure does not show.

Operators also need reliable technical documentation throughout the facility lifecycle. Design decisions, commissioning results, alarm procedures, and maintenance records often sit across disconnected systems.

A searchable knowledge base can help engineering teams preserve those records. It does not replace controls or commissioning, but it can reduce operational knowledge gaps.

The skeptical reading is therefore straightforward. Vertiv has met a meaningful product-level requirement, according to the company and NVIDIA’s program structure.

It has not established that every deployment will deliver 2.3 MW of usable cooling under customer conditions. It has not removed dependence on facility water, redundancy, controls, or commissioning.

That is not a flaw unique to Vertiv. It is the normal boundary between component qualification and system performance.

The badge has value when buyers understand that boundary. It becomes risky when procurement teams treat it as a substitute for engineering evidence.

Three Signals Will Show Whether the Qualification Matters

Real deployments, transparent operating data, and competitive responses will determine whether DSX Ready becomes a procurement standard.

The first signal is named customer adoption. Vertiv needs to move the qualified model from a product announcement into disclosed AI factory, hyperscale, or colocation deployments.

A customer reference would provide more than marketing validation. It could reveal the number and type of racks served, the redundancy model, commissioning process, and facility-water design.

Operating evidence would strengthen the case further. Buyers should watch for performance data across changing loads rather than one maximum-capacity result.

Useful disclosures would include temperatures, flow, pump behavior, failover performance, maintenance requirements, and the surrounding heat-rejection configuration.

A public deployment would strengthen Vertiv’s position if it shows stable service under realistic conditions. Continued silence would not prove a technical problem, but it would leave the qualification largely at the screening stage.

The second signal is the evolution of NVIDIA’s requirements and marketplace. The DSX Ready framework is most valuable when its criteria remain clear, technically demanding, and updated across accelerator generations.

NVIDIA’s documentation says the validated CDU partner list is dynamic. Buyers should expect products and status labels to change as new equipment completes qualification.

More qualified products would make DSX Ready a broader purchasing baseline. That development could help customers while reducing the marketing advantage held by early participants.

The marketplace also needs to distinguish different readiness levels and test scopes clearly. A buyer should be able to tell which model, configuration, and requirement set a qualification covers.

Changes tied to Vera Rubin systems will be especially important. Higher cabinet thermal loads can alter flow, redundancy, and facility-water assumptions.

If Vertiv maintains qualification as those requirements evolve, the announcement becomes evidence of an adaptable product platform. If the product requires substantial redesign, its current advantage becomes more temporary.

The third signal is competitor execution. LG already lists a 2.6 MW DSX Ready CDU, while Motivair offers a 2.5 MW system within a broader scalable architecture.

LG says it plans to pursue qualification for a 4 MW CDU by the end of 2026. Delivery against that goal would push nominal capacity comparisons higher.

Vertiv does not need to win a capacity race to remain competitive. It does need to show why 2.3 MW offers the right balance of footprint, redundancy, serviceability, controls, and deployment flexibility.

Watch how suppliers package products into complete reference designs. A strong CDU paired with incompatible controls or inadequate heat rejection cannot support an AI factory alone.

Also watch whether operators standardize around one supplier or assemble multi-vendor systems. Single-vendor designs can simplify accountability, while multi-vendor designs can improve flexibility and negotiating leverage.

The most revealing competitive evidence will come from repeat orders. A first installation can reflect experimentation or schedule pressure. A repeat deployment suggests the equipment survived commissioning and operational review.

For developers and enterprise buyers, the immediate action is not to select a CDU from its headline rating. It is to convert the Vertiv NVIDIA DSX Ready CDU qualification into a structured engineering review.

Ask which test conditions match the planned site. Compare performance curves, failure behavior, service access, controls, water requirements, and upgrade paths.

Then examine the complete thermal chain, from processor cold plates to the final heat-rejection system. If any link remains undersized, the qualified CDU cannot rescue the design.

NVIDIA and its infrastructure partners are building a common language for AI factory planning. The next question is whether disclosed deployments prove that language can produce repeatable operating results.