Delta shows a 150 kW liquid-to-air CDU for running Vera Rubin NVL72 racks in air-cooled halls
A rack-sized radiator that draws 18 kW to shed 150 kW shows what it now costs to move heat away from an AI rack.
Delta Electronics has shown a liquid-to-air coolant distribution unit (CDU) capable of removing 150 kW of heat from a single rack, paired with ASRock Rack's NVIDIA Vera Rubin NVL72 system, ServeTheHome reported on 8 August. The GoCool-150 was on display at ASRock Rack's Computex 2026 booth and is aimed at data centres that want to install direct-liquid-cooled racks without a facility-wide liquid cooling loop.
The unit circulates 225 litres of coolant per minute and is designed to return coolant to the rack at 45°C, which ServeTheHome notes is in line with NVIDIA's design targets for the Vera Rubin generation. Heat from the rack is carried into two heat exchangers that run almost the full height of the CDU, each with five copper pipes, and is then pushed into the room by an array of 32 Delta PFD2048HT fans. These are 200 mm units running on 48 V DC, rated at up to 336 W each, with a maximum speed of 8,250 RPM and a noise figure of up to 81 dBA. Together the fans move 17,658 CFM of air. The fans and the five pumps are hot-swappable, so a failed unit can be replaced while the rack keeps running.
The physical scale is notable. The CDU stands 2.3 metres tall and weighs 1,200 kg before coolant is added. ServeTheHome also reports that the CDU itself requires 18 kW of electrical power to operate, an overhead of more than a tenth of the heat it removes.
The design proposition is straightforward: many existing halls were built for air-cooled servers, with air handlers and hot-aisle containment, and lack chilled-water plumbing at the rack. A liquid-to-air CDU keeps the liquid loop local to one rack and hands the heat off to the existing air infrastructure, at the cost of floor space, weight and fan power.
Why it matters
Rack power for AI systems has climbed to a level where air alone cannot carry the heat away from the chips, but the buildings housing those racks were largely designed for air. Products like this one are a bridge: they let operators deploy current-generation accelerators in older halls while facility-scale liquid cooling is retrofitted or built new. The trade-off is visible in the numbers, with more than a tonne of hardware and tens of kilowatts spent just to move heat. Expect cooling overhead, not just GPU count, to become a headline figure in AI deployments.
Rack power for AI systems has climbed to a level where air alone cannot carry the heat away from the chips, but the buildings housing those racks were largely designed for air.
Accelerators and cold plates
Chips in the NVL72 rack transfer heat into coolant flowing through cold plates
Rack coolant loop
225 litres per minute carry the heat out of the rack to the CDU
CDU heat exchangers
Two full-height exchangers with copper pipes pass heat from liquid to air; coolant returns at 45°C
32-fan array
17,658 CFM of air carries the heat into the hot aisle for the building's existing air handlers
What you can learn from this
Direct liquid cooling uses water because it carries far more heat than air. Water has roughly four times the specific heat capacity of air by mass and is hundreds of times denser, so a modest flow of coolant can absorb heat that would otherwise require enormous volumes of air. Cold plates sit on the chips, coolant flows through them, and the warmed coolant leaves the rack. The 225 L/min figure here is what it takes to move 150 kW while keeping the temperature rise across the loop small.
A CDU separates the rack loop from whatever sits outside it. A coolant distribution unit contains pumps, a heat exchanger and controls; it keeps the clean, treated coolant that touches the servers isolated from the facility side. In a liquid-to-liquid CDU the facility side is chilled water; in a liquid-to-air CDU like this one, the facility side is the room's air. The exchanger transfers heat across a metal boundary without mixing fluids, which protects expensive cold plates from contamination.
Supply temperature targets are set by the chips, not the room. A 45°C coolant supply sounds warm, but accelerators are designed to run with junction temperatures well above that, so a warm loop still provides a useful gradient. Higher supply temperatures reduce or eliminate the need for chillers, because rejecting heat to ambient air is easier when the coolant is hotter than the room. That is why vendors publish a target temperature and why cooling hardware is built to meet it.
Moving heat costs energy, and that cost shows up as overhead. The 18 kW this CDU draws is spent on pumps and fans, and it adds to the facility's power without doing any computing. Data centre efficiency is often summarised as PUE, the ratio of total facility power to IT power; every kilowatt of cooling pushes PUE upward. Liquid-to-air is less efficient than liquid-to-liquid because fans must move a large air volume, which is the price of not having chilled water at the rack.
Hot-swappable pumps and fans are about availability, not convenience. A rack drawing 150 kW will overheat within minutes if cooling stops, so a single failed pump cannot be allowed to take the system down. Redundant units that can be replaced without shutting off the loop let maintenance happen while the servers keep working. This N+1 style of redundancy is the same principle used for power supplies and network links.
We teach this
How to use this in practice
Calculate the airflow needed to remove a given heat load. Use the common rule-of-thumb formula CFM ≈ watts × 3.16 ÷ temperature rise in °F, and work out the air volume needed to remove 150 kW with a 20°F rise, then compare it to the 17,658 CFM figure above. Done means a short worked calculation in a notebook showing your assumptions and how close you land to the published number. Then rerun it with a 30°F rise to see how sensitive the airflow requirement is to the allowed temperature difference.
Draw the heat path from silicon to outside air. Sketch chip, cold plate, rack manifold, CDU heat exchanger, fan array, hot aisle, air handler and the building's heat rejection. Mark which stages carry heat in liquid and which in air, and where the fluid boundary sits. Done means a one-page diagram where each arrow is labelled with the medium and, where you know it, the temperature.
Read your own machine's thermal data. On Linux run
sensorsafter installing lm-sensors, or on Windows open Task Manager's performance tab or a tool such as HWiNFO, and record CPU or GPU temperature at idle and under a sustained load like a compile or a benchmark. Done means a small table of idle and loaded temperatures plus fan speed, which gives you a felt sense of the gradient that cooling hardware must maintain. If you are on a laptop, also note how quickly fan speed responds once load starts.Estimate the PUE impact of cooling overhead. Take a 150 kW IT rack and add 18 kW of CDU power, then add a plausible allowance for the facility's air handlers and lighting, and compute total divided by IT. Do it once for liquid-to-air and once assuming a liquid-to-liquid CDU with lower fan power. Done means two PUE figures side by side and a sentence explaining why they differ, with the assumptions left visible in the sheet so someone else can challenge them.
Sources
- Delta's GoCool-150 Goes Big To Enable 150kW Liquid-To-Air Cooling for ASRock Rack's NVIDIA VR NVL72 — ServeTheHome
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report