AI Liquid Cooling: Dry Loops vs Evaporative Towers

AdvancedUNO
11 Aug, 2026
AI Liquid Cooling: Dry Loops vs Evaporative Towers

6 min read

The Liquid Cooling Deployment Playbook

  • The Friction: Doubling rack densities and 500% grid connection backlogs mean operators cannot afford to waste power or water on inefficient cooling.
  • The Architecture: A stark operational choice between water-free, space-hungry dry closed-loops and hyper-efficient, water-guzzling evaporative towers.
  • The Immediate Action: Map your local meteorological peak dry-bulb temperatures against your chip-inlet thresholds before specifying your heat rejection loop.

Why 100kW Racks Are Melting the Traditional Datacenter Air Handler

As rack densities double annually to support frontier models, operators face a stark thermodynamic limit: air can no longer carry the heat away. When an AI cluster pushes past 40kW per rack, attempting to cool it with traditional air handlers is like trying to cool a jet engine with a hand-held paper fan. The air simply cannot move fast enough through the chassis without turning the server room into a screaming wind tunnel.

This physical barrier is driving an unprecedented industrial scramble. nVent Electric recently announced a massive 160,000 square-foot manufacturing expansion in Blaine, Minnesota, marking their third liquid cooling capacity expansion in three years to meet this surging demand. With a projected market size of $23.24 billion by 2035 according to DataM Intelligence, liquid cooling is no longer an exotic option for overclockers; it is the baseline requirement for modern silicon.

The urgency is compounded by utility constraints. Steven Carlini of Schneider Electric recently highlighted that grid connection queues in the US and UK face a staggering 500% backlog. Because cooling can consume up to 40% of a traditional data center's total energy budget, shifting to liquid cooling is the most direct path to clawing back power capacity that can be redirected straight into the compute stack.

The Thermodynamics of High-Density Heat Rejection

To understand the engineering behind this transition, we have to look at how heat moves from a microscopic transistor to the outside world. Modern liquid cooling relies on a two-loop system. The secondary loop runs clean, low-conductivity fluid directly across copper cold plates bolted to the processors. This fluid absorbs the heat and carries it to a Coolant Distribution Unit (CDU).

Think of a CDU as a corporate translator: it sits between the dirty, high-pressure industrial facility water loop and the pristine, low-pressure fluid running directly across your million-dollar GPUs, keeping them completely isolated. The heat is transferred across a plate heat exchanger inside the CDU from the secondary loop to the primary facility loop, which carries the heat outside the building.

How NVIDIA's 45-Degree Celsius Spec Changes the Primary Loop

The real magic happens at the roof. Historically, datacenters had to keep facility water chilled to around 15°C to 20°C to keep servers cool. Keeping water that cold requires massive, energy-hungry chillers. However, platforms like NVIDIA Blackwell are designed to operate with inlet fluid temperatures as warm as 45°C (113°F).

Because the fluid returning from the racks is even hotter, the temperature difference between the coolant and the outside air is massive. This delta allows operators to dump heat directly into the ambient air using dry coolers (essentially giant car radiators with fans) without needing any refrigeration. This shift is what makes near-zero water consumption technically viable.

How to Sequence Your Liquid Cooling Migration: The 4-Step Playbook

Transitioning a facility to liquid cooling requires a tightly orchestrated sequence of engineering decisions. Skipping a step or miscalculating a thermal threshold can result in choked performance or catastrophic fluid leaks.

  1. Step 1 — Map the meteorological envelope: Audit your local historical p99 dry-bulb and wet-bulb temperatures. If your peak dry-bulb temperatures stay safely below your target chip-inlet temperature minus your system's approach temperature, you are a prime candidate for water-free dry cooling.
  2. Step 2 — Size and select the Coolant Distribution Units: Match the CDU flow rate, measured in liters per minute, to the collective Thermal Design Power (TDP) of your server racks. Ensure the CDU has redundant pumps and can maintain a stable pressure differential across the secondary loop.
  3. Step 3 — Deploy secondary-loop manifold plumbing: Install stainless steel or high-grade polymer manifolds along the rack frame. Every connection to a server blade must use dripless, quick-disconnect couplings from specialized manufacturers like CPC or Parker to prevent micro-leaks during hot-swaps.
  4. Step 4 — Balance and commission the primary loop: Fill the system, purge all air pockets, and run hydrostatic pressure tests at 1.5 times the operating pressure. Calibrate your telemetry to poll flow meters, pressure transducers, and leak-detection ropes at sub-second intervals to catch anomalies before they trigger automatic shutdowns.

Weighing the Trade-offs: Dry Closed-Loops vs. Evaporative Towers

Operators must choose how to reject heat from the primary loop to the environment. There is no single winner here; the choice is a direct trade-off between water conservation and physical footprint.

  • Dry Closed-Loops (Zero-Water Cooling): This approach uses sealed radiators to dump heat directly into the air. It consumes virtually zero water, completely bypassing municipal water permit delays and environmental scrutiny. The catch is that dry coolers require a massive physical footprint on the roof or ground, and their heat rejection efficiency drops sharply when ambient summer temperatures spike.
  • Evaporative Wet Towers (Latent Heat Cooling): This method sprays water over a media fill, using the cooling effect of evaporation to reject heat. It is incredibly efficient and keeps the facility loop much cooler, which can improve GPU performance. The downside is massive water consumption, often chewing through 2.6 million gallons of water per megawatt annually, making it a target for regulators and local communities.

The Hidden Failure Modes of Rapid Liquid Deployments

When engineering teams rush liquid cooling deployments to meet aggressive AI compute schedules, they frequently stumble into three specific operational traps.

  • The Metallurgy Blindspot: Mixing incompatible metals in the fluid loop is a recipe for disaster. If you couple copper cold plates with aluminum manifolds or fittings without perfect chemical isolation, you create a battery. Galvanic corrosion will rapidly eat away the aluminum, clogging the microscopic channels in your cold plates with metallic debris within months.
  • The Flow-Velocity Erosion Trap: To squeeze more cooling capacity out of existing plumbing, operators are tempted to crank pump speeds. However, pushing fluid through copper tubes faster than the manufacturer's recommended velocity limits strips away the protective copper oxide layer, leading to rapid mechanical erosion and eventual pipe wall failure.
  • The Chemical Maintenance Gap: Unlike air systems, liquid loops require constant chemical monitoring. Neglecting biocide and corrosion inhibitor levels leads to rapid biological fouling. A thin layer of slime inside a cold plate acts as an insulator, causing chip temperatures to spike and triggering thermal throttling under heavy training workloads.

Frequently Asked Questions

What happens to our primary loop heat rejection when ambient dry-bulb temperatures exceed 45°C during a summer peak?

If ambient dry-bulb temperatures rise above your target coolant temperature, a pure dry-cooler loop loses its ability to transfer heat. To prevent your GPUs from thermal throttling, you must either deploy a trim-chiller to mechanically assist the cooling loop or use an adiabatic spray system that wets the dry-cooler coils to lower the entering air temperature via localized evaporation.

How do we handle secondary loop fluid volume expansion during rapid workload transitions from idle to peak TDP?

AI workloads cause rapid temperature swings, which in turn cause the coolant to expand and contract. Your Coolant Distribution Unit must be equipped with an expansion vessel (diaphragm tank) sized to absorb these volume changes. If the expansion tank is undersized, the resulting pressure spikes will stress your quick-disconnect fittings and trigger high-pressure alarms that can trip the entire rack.

The Architect's Field Verdict: If you are building in a region with strict environmental regulations or facing multi-year delays on municipal water hookups, design your system around a warm-water dry-loop architecture. However, this path only works if you verify that your chosen GPU platform can run reliably at 45°C inlet temperatures during your local summer peaks. Map your local wet-bulb and dry-bulb profiles before you commit to a hardware vendor.

How many megawatts of your current AI pipeline are actually pre-plumbed for facility water, and are you willing to bet your H1 2027 deployment dates on a municipal water connection?

Related from this blog

Sources

Next Post Previous Post
No Comment
Add Comment
comment url