Signal & Supply
← Archive
November 4, 2026 LIQUID COOLING

A Modern AI Rack Draws About 130 Kilowatts. Fans Stop Working Around 30 — and the Part That Replaced Them Is in Short Supply.

A GPU rack from a few years ago pulled roughly 15 kilowatts, and a fan could keep it cool — the same trick that's cooled computers since the 1960s. Nvidia's current flagship rack pulls around 130 kilowatts in the same footprint. Air physically can't carry that much heat away. The box that can is harder to build than it sounds, and nobody can make enough of them.

Key takeaway An Nvidia GB200 NVL72 rack draws roughly 120–132 kilowatts, by vendor spec sheets — well past the point where ordinary airflow can remove heat fast enough, commonly cited around 30 kilowatts per rack. Above that line, liquid cooling isn't an optimization; Nvidia specifies it as a requirement. The part that makes it possible is the Coolant Distribution Unit (CDU): a heat-exchanger box that pumps a sealed coolant loop through cold plates on every chip without ever mixing fluids with the building's water system. Demand for CDUs is reportedly outpacing what cooling vendors can build, and even the market leader, Vertiv, holds under a third of global CDU revenue — there's no single company positioned to just scale up and close the gap the way one fab might for chips.

Push a rack past 30 kilowatts and watch which one survives

Drag the slider to set a rack's power draw. Left is a fan-cooled rack with a fixed air-cooling ceiling; right is the same rack with a Coolant Distribution Unit (CDU) feeding liquid-cooled cold plates. Ceilings (≈30 kW for air, ≈230 kW for a modern in-rack CDU) are simplified, rounded figures from vendor specs and industry benchmarks — real limits vary by rack design, airflow, and climate. "Homes" uses a rough 1.2 kW average continuous U.S. household draw.

Within cooling capacity Over capacity — overheating Cooling ceiling
20 kW
Rack power draw
20 kW
Air-cooled rack
—
Liquid-cooled rack
—
At 130 kW — a real GB200 NVL72 rack — the air-cooled version is already deep into overheat territory while the liquid-cooled version is barely past half its CDU's rated capacity. That gap is why every GPU generation since has made the CDU, not the fan, the part standing between a rack and running at full power.

The plain version

A rack of GPUs a few years ago drew about 15 kilowatts — roughly fifteen hair dryers running flat out. A fan blowing across the servers could carry that heat away, the same basic trick that's cooled computers since the 1960s.

Nvidia's current flagship rack, the GB200 NVL72, draws around 130 kilowatts in that same footprint — enough continuous power to run roughly a hundred typical homes, out of one cabinet barely wider than a refrigerator. At that density, air just can't carry the heat away fast enough. Blow harder and you mostly get a wind tunnel of hot air circling back on itself. Industry benchmarks put the practical ceiling for air cooling around 30 kilowatts per rack. Past that, you need to put liquid directly onto the chip.

That's the job of a part almost nobody outside data centers has heard of: the Coolant Distribution Unit, or CDU. Picture a radiator about the size of a vending machine sitting at the end of the row. It pumps a sealed loop of coolant through cold plates clamped onto every GPU, picks up the heat, and hands it off — through a heat exchanger, never mixing fluids — to the building's main water system. No CDU, and the rack can't run at full power without overheating and shutting itself down.

Here's the squeeze: AI data centers are being built faster than the CDU industry can make these things. One major cooling vendor said recently it's racing to expand two factories just to cover a fraction of the AI capacity coming online this year. Vertiv, the single biggest CDU maker, holds under a third of the global market — meaning even the leader can't single-handedly close the gap. The GPU everyone talks about being sold out is only half the story. So, more quietly, is the box that keeps it from cooking itself.

The expert version

Thermal design power scales with transistor density and clock, and the current GPU generation has outrun air's achievable heat-transfer rate. A single Nvidia GB200 NVL72 rack — 36 Grace CPUs paired with 72 Blackwell GPUs over NVLink — carries a published rack-level draw in the roughly 120–132 kW range, versus on the order of 15 kW for air-cooled racks only a couple of GPU generations back. Air cooling removes heat via convection into a moving airstream; its ceiling is set by achievable airflow volume and the delta-T a server can tolerate before thermal throttling, which caps most air-cooled racks around 30 kW in practice. Past that, direct liquid cooling isn't an optimization — Nvidia specifies it as a requirement for GB200-class deployments.

Mechanically: cold plates with internal microchannels sit in direct thermal contact with the GPU and CPU dies (via their heat spreaders), circulating a treated-water or dielectric coolant loop that's fully sealed and isolated from the facility's chilled-water loop. The Coolant Distribution Unit is the interface between those two loops — a plate heat exchanger plus redundant pumps, filtration, and leak/flow sensing, with in-rack units sized to 200–250 kW of rejected heat, enough for a full GB200 cabinet with margin. It's what lets a facility keep its existing chilled-water plant while running a separate, higher-purity loop next to silicon worth tens of thousands of dollars per chip that can't tolerate particulate contamination or corrosion.

Manufacturing a CDU at AI-cluster volume means precision-brazed heat exchangers, redundant pump assemblies, and sensor packages engineered to run unattended for years — a narrower, more specialized supplier base than server chassis. One market tracker puts Vertiv, the category leader, at under 30% of 2025 global CDU revenue, with Schneider Electric, Boyd Thermal, CoolIT, and nVent splitting most of the rest; none has the single-vendor dominance of an Nvidia or TSMC in their own layers of the stack. Trade press reports demand for liquid cooling rising faster than vendors anticipated, pushing expansion rounds — one cooling-system maker, LiquidStack, said it's scaling two plants to cover roughly 3 GW of near-term capacity — that are still running behind the pace of GPU rack deployment.

Why it matters for tech + supply chain: Everyone tracks the GPU shortage. The thermal hardware that lets a GPU rack actually run at full power is a thinner, less-watched supply chain underneath it.

Why it matters for tech + supply chain: CDU capacity, not GPU allocation, is increasingly the gating factor on how fast new AI racks can actually be energized at rated power.

Get it in your inbox every morning