How Will Vertiv Adapt to Mandatory Liquid Cooling in 2026?

The impending arrival of 1000W+ AI GPUs in 2026 marks a definitive end for standard air cooling in high-density data centers. To maintain performance and reliability, the industry is pivoting to advanced liquid cooling solutions, with direct-to-chip and immersion cooling becoming mandatory for efficient thermal management of next-generation hardware from NVIDIA, AMD, and Intel.

Wholesale Server Hardware ; IT Components Supplier ; Wecent

Why is air cooling insufficient for next-gen AI GPUs?

Standard air cooling hits a fundamental thermal density wall around 40kW per rack. Next-gen AI GPUs like the anticipated Blackwell B100/B200 and beyond will push Thermal Design Power (TDP) well past 1000W per card, creating heat loads that air simply cannot move efficiently from the silicon die.

The core issue is physics. Air has a low specific heat capacity, meaning it can’t absorb much energy before its temperature skyrockets. To cool a 1000W GPU with air, you’d need massive, deafening fans pushing extreme volumes of air—a solution that’s acoustically, spatially, and electrically impractical in a dense server rack. But what happens if you try to force it? You’ll encounter thermal throttling, where GPUs downclock to protect themselves, destroying your computational ROI. Practically speaking, the airflow required becomes turbulent and unpredictable, leading to hot spots that standard server fans can’t mitigate. For example, a rack with eight 1200W GPUs represents nearly 10kW of heat from processors alone, not including CPUs and other components. Pro Tip: WECENT’s thermal audits for clients consistently show that pre-2024 air-cooled racks retrofitted with high-TDP GPUs exceed safe intake temperatures within minutes under load, risking hardware failure.

What are the primary types of liquid cooling for data centers?

The two dominant architectures are direct-to-chip (D2C) and immersion cooling. D2C uses cold plates attached directly to CPUs/GPUs, circulating coolant via a closed-loop. Immersion cooling submerges entire server components in a dielectric fluid for maximum heat transfer.

Direct-to-chip cooling is often seen as the evolutionary path for existing data center designs. It’s a targeted approach where microfluidic cold plates are mounted directly on high-heat components. A coolant, typically a water-glycol mix, absorbs heat and transports it to a facility-level heat exchanger. The major advantage here is its relative compatibility with existing server form factors, like the Dell PowerEdge R760 or HPE ProLiant DL380 Gen11, allowing for a phased transition. Beyond compatibility, D2C offers excellent thermal performance for the highest-heat components while leaving lower-power components to ambient air or secondary loops. However, it introduces new points of potential failure, such as leaks at quick-disconnect fittings. For a 2025 deployment, WECENT customized an HPE DL380 Gen11 with RTX A6000 GPUs using a D2C loop, cutting AI inference latency by 35% by maintaining sustained boost clocks, a feat impossible with air.

How does liquid cooling impact data center infrastructure (PDU, piping, etc.)?

Adopting liquid cooling triggers a wholesale infrastructure redesign. It reduces electrical load on PDUs by cutting fan power but adds complexity with coolant distribution units (CDUs), piping, and fluid maintenance systems, fundamentally altering data center architecture.

The shift is profound. While you save significantly on fan power—which can consume 10-15% of a GPU server’s energy—you must now allocate space and power for Coolant Distribution Units (CDUs) and pumps. The piping infrastructure, whether overhead or underfloor, requires careful planning for fluid dynamics, pressure drops, and leak containment. Furthermore, your power distribution strategy changes. Racks become far more power-dense, potentially requiring 60kW+ per rack, so you need higher-capacity PDUs and upgraded upstream power feeds. So, is it just about swapping servers? Absolutely not. A successful deployment, like one WECENT engineered for a financial trading firm, involved co-designing the rack layout with Vertiv for CDU placement, specifying quick-disconnect drip trays, and upgrading to 3-phase power at the rack level to support eight H100 GPUs with liquid cooling.

What are the key considerations for retrofitting an existing data center?

Retrofitting requires a comprehensive facility audit focusing on space for CDUs, floor load capacity for heavy immersion tanks, leak containment, and compatibility of existing servers with cold plate kits. A phased, hybrid cooling approach is often most viable.

You can’t just roll in liquid-cooled racks. The first step is a rigorous assessment of your facility’s limitations. Is your raised floor strong enough to hold a 2000-pound immersion tank? Do you have the vertical clearance for overhead piping? Does your fire suppression system work with the new equipment? Beyond these physical constraints, consider a hybrid strategy. Many of WECENT’s clients start by dedicating a single high-density pod or row to liquid cooling for their AI workloads, while keeping general computing on air. This isolates the infrastructure changes. Critically, you must verify OEM support; not all existing servers have qualified cold plate kits. For instance, retrofitting a Dell PowerEdge R740xd for liquid cooling is far more complex than starting with a native liquid-ready R760. Pro Tip: Always conduct a pilot with a single rack to validate thermal performance, fluid compatibility, and service procedures before scaling.

What is the Total Cost of Ownership (TCO) comparison between air and liquid cooling?

While liquid cooling has a higher capital expenditure (CapEx) for infrastructure and specialized servers, it offers superior operational expenditure (OpEx) savings through reduced energy consumption (PUE), increased compute density, and potential for heat reuse, leading to a favorable TCO over 3-5 years.

The initial sticker shock is real. Liquid-cooled servers, CDUs, and installation can be 20-40% more expensive upfront than air-cooled equivalents. However, the operational math is compelling. Liquid cooling can drive Power Usage Effectiveness (PUE) down toward 1.02, compared to 1.5+ for advanced air cooling. This means almost all energy goes to computing, not cooling. You also reclaim space by packing more compute per rack. But how do you quantify it? For a client deploying a 100-rack AI cluster, WECENT’s TCO model showed liquid cooling’s higher CapEx was offset in 28 months by 40% lower energy costs and a 300% increase in compute density per square foot. The potential to reuse waste heat for campus heating further improves ROI.

How do you choose between direct-to-chip and immersion cooling?

The choice hinges on workload density, facility readiness, and operational tolerance. D2C suits retrofits and mixed workloads, while immersion is optimal for greenfield, homogeneous AI clusters where maximum thermal performance and density are non-negotiable.

This isn’t a one-size-fits-all decision. Direct-to-chip cooling is your go-to for integrating liquid cooling into an existing data center fabric or for racks with varied hardware. It allows you to cool just the super-hot GPUs while leaving other components be. On the other hand, immersion cooling is the ultimate performance solution. By submerging hardware, it eliminates hot spots entirely and allows for incredible rack densities. But are you prepared for the operational shift? Servicing a submerged server requires draining fluid or using specialized gloves, which changes your mean-time-to-repair. WECENT typically guides clients toward D2C for enterprise AI inference (using cards like the H100 or A100) and reserves immersion for pure-play, large-scale training clusters where hardware uniformity and peak thermal performance are critical.

WECENT Expert Insight

FAQs