The conversation about data center cooling efficiency happens at the wrong altitude. Every operator knows their PUE number. Every investor asks about it. Every sustainability report leads with it. PUE is a facility-level metric. Total facility power divided by IT equipment power. It tells you how much overhead the building adds to every watt of compute.
It does not tell you where the waste is.
Roughly 30% of total data center electricity consumption goes to cooling infrastructure. Chillers, pumps, fans, cooling towers, CDUs, compressors. In older air-cooled facilities running above 1.6 PUE, that number can push toward 40%. In modern liquid-cooled facilities at 1.2 PUE, it drops toward 15 to 20%. The industry has spent a decade driving that number down through architectural changes. Moving from air to liquid. Moving from chilled water to warm water. Moving from mechanical chillers to free cooling. Moving from evaporative towers to dry coolers.
These are the right moves. They work. But they are all macro-level interventions. Swap out the cooling plant. Redesign the air handling. Replace the chillers. Each one is a capital project measured in millions.
Almost nobody is looking at the components inside the cooling loop.
Inside a direct-to-chip liquid cooling loop, the primary energy consumers are the pumps. Pumps push coolant through manifolds, across cold plates, and back to the CDU. The energy a pump consumes is a function of flow rate and pressure. More flow, more watts. Higher pressure, more watts.
Here is the problem. Most pumps in data center cooling loops are overdriven. They push more flow than the system needs because the system was designed for worst-case thermal load, and worst-case thermal load almost never occurs simultaneously across every chip in the rack. The pump runs at a speed that guarantees adequate cooling for the hottest chip in the loop, which means it is delivering excess flow to every other chip.
That excess flow is wasted energy. Pure overhead. The pump does not know which chips are hot and which are idle. It pushes the same total volume through the loop regardless of workload distribution. A rack with eight GPUs, two running at 100% and six at 20%, receives the same pump energy as a rack with all eight at 80%.
The waste compounds with scale. A row of racks sharing a CDU means a shared pump system delivering flow to dozens or hundreds of cold plates. The pump speed is set by the worst-performing branch. Every other branch gets more flow than it needs. Multiply that across a facility with hundreds of racks and the energy consumed by pump over-drive becomes a measurable fraction of the total cooling bill.
When coolant flow is unevenly distributed across cold plates in a shared loop, the operator has two options. Accept the imbalance and risk thermal throttling on the starved chips. Or overdrive the pump to push enough total flow that even the most disadvantaged branch receives adequate cooling.
Every operator picks option two. Nobody throttles GPUs that cost $30,000 each to save pump energy.
But that overdriven pump is pushing excess flow to every cold plate that was already receiving adequate cooling. Those chips do not cool faster because they get more flow. The excess flow just passes through the cold plate, absorbs less heat per unit volume, and returns to the CDU carrying less thermal load than it could. The CDU then rejects this partially warmed coolant, burning energy on heat rejection for a thermal delta that could have been larger if the flow had been distributed efficiently in the first place.
The inefficiency hits twice. Once at the pump (excess flow). Once at the heat rejection (suboptimal thermal delta across the coolant).
Precision flow control at the cold plate level eliminates most of this waste. If each cold plate receives exactly the flow it needs based on the thermal load of the chip beneath it, the pump does not need to be overdriven. Total system flow drops. Pump speed drops. Energy consumption drops.
The components that enable this are small. Pressure-compensated flow regulators the size of a pipe fitting. Proportional solenoid valves at each cold plate inlet. Temperature sensors feeding real-time data to a control system that adjusts flow distribution dynamically. None of these are exotic technologies. Hydraulic systems in aerospace and industrial automation have used pressure-compensated flow control for decades. The adaptation to data center cooling is an engineering project, not a research project.
The savings are direct. A pump running at 80% speed instead of 100% consumes roughly 50% less energy (pump power scales with the cube of speed). In a facility where pump energy represents 5 to 10% of total cooling energy, and cooling represents 20 to 30% of total facility energy, a 50% reduction in pump energy translates to a 1 to 3% reduction in total facility power consumption. At a 100 MW facility, that is 1 to 3 MW saved. Permanently. From a component that costs a fraction of a chiller replacement.
Three reasons.
First, the cooling industry's sales cycle is built around systems, not components. CDU vendors sell CDUs. Chiller vendors sell chillers. Manifold vendors sell manifolds. Nobody's commission is tied to selling a $15 flow regulator at each cold plate inlet. The economic incentive to optimize at the component level does not exist in the current vendor ecosystem.
Second, PUE does not decompose well. The metric tells you the total overhead but not where it comes from. A facility at 1.25 PUE could be running highly efficient chillers and wildly inefficient pumps and you would never know from the PUE number alone. There is no standard metric for pump efficiency in a cooling loop, for flow distribution effectiveness, or for the energy cost of flow imbalance. What does not get measured does not get managed.
Third, the industry is still in deployment mode. The primary challenge for most operators in 2026 is getting liquid to the chip at all. Optimizing how that liquid is distributed chip by chip is a second-order problem. It will not stay second-order for long. When every facility has liquid cooling and the differentiation shifts from "do you have it" to "how efficiently do you run it," the component-level optimizations will be where the margin lives.
The cheapest megawatt in a data center is the one you stop wasting. Right now, a meaningful fraction of that waste is happening inside the cooling loop at the component level, in overdriven pumps and static flow restrictions that nobody has bothered to optimize. The components to fix it exist. The engineering to integrate them exists. The incentive structure to deploy them is the part that has not caught up.