The Parasitic-Power Ledger: Pumps, Distribution Units, And The Watts Liquid Cooling Still Costs
When a facility moves from air to liquid, it is tempting to imagine that cooling has become someone else's problem, tidily piped away. In fact liquid cooling relocates power rather than removing it, and it introduces a set of electrical loads that did not exist in the air-cooled world. To read a liquid-cooled design honestly, you have to keep a parasitic-power ledger, an accounting of every watt the cooling system itself consumes, because those watts land squarely on the power plant's bill and on the site's efficiency.
The centerpiece of that ledger is the coolant distribution unit, the device that stands between the facility's water and the fluid flowing over the chips. It holds a heat exchanger that passes heat from the technical loop to the facility loop, and it holds pumps that push coolant through the cold plates. Those pumps are the new parasitic load. Moving liquid through narrow cold-plate channels against real flow resistance takes continuous electrical power, and at scale the pumping load across a large hall is far from trivial. It is smaller than the fan power an equivalent air design would have burned, which is precisely why liquid wins on efficiency, but it is not zero, and it must be fed by the same electrical system that feeds the compute. The pump is a compute-adjacent load the power plant now carries.
The ledger has more lines. The facility water loop that receives heat from the distribution units still has to reject that heat to the environment, and depending on the climate and the loop temperature that rejection may need cooling towers, dry coolers, or chillers, each with its own pumps and fans and, in the chiller's case, a compressor that is one of the largest single electrical loads in the building. Liquid cooling at the rack does not abolish the rejection plant; it changes its size and its duty. Warm-water designs shrink the compressor's role dramatically by letting much of the heat go to ambient air directly, but the pumps that circulate the warm water run all the time. Every one of these devices is a watt drawn and, by the conservation identity, a watt of heat that the plant must in turn handle.
Why does this ledger matter commercially? Because the headline efficiency claims that sell liquid cooling only hold if the parasitic loads are counted honestly. A design that quotes a beautiful overhead ratio while quietly omitting its pumping and rejection power is quoting a fiction. The right comparison is total facility power against compute power, with every pump, every fan, every compressor, and every conversion loss in the numerator. Liquid cooling genuinely does move that ratio toward efficiency, often substantially, but the gain comes from replacing large fan loads with smaller pump loads and from running warm enough to idle the compressor, not from making the cooling free. An operator who cannot point to each line of the parasitic ledger cannot defend the efficiency number.
There is a subtler point about where the parasitic heat goes. In an air design the fan heat is dumped into the same air stream that cools the chips, so it directly burdens the compute cooling. In a liquid design the pump heat is added to the coolant, which raises the temperature of the fluid arriving at the next component and slightly reduces the loop's remaining capacity. The parasitic load is not only a power cost; it consumes a share of the very cooling capacity it exists to provide. Sizing a loop means budgeting for the heat the loop's own pumps add to it, a small self-referential tax that a careless design forgets.
The distribution units also concentrate risk in a way the ledger should flag. A single unit may serve a whole row of dense racks, so its pumps and its heat exchanger sit on the critical path for a large block of compute. The power feeding those pumps is therefore not ordinary building load; it is critical load, deserving the same backup and redundancy as the servers, because if the pumps stop, the racks they serve begin to overheat within seconds. Many efficiency-first designs learn this the hard way, having powered the pumps from ordinary building supply only to discover that a power blip took down cooling and compute together. On the ledger, pump power must be marked as critical, not incidental.
For the investor, the parasitic-power ledger is the tool that separates a real efficiency story from a marketing one. Ask to see total facility power against compute power with every cooling load itemized, ask whether the pumps are on protected power, and ask how much of the year the compressors are expected to run. A team that answers crisply has treated the cooling system as an electrical system, which is exactly what liquid cooling turns it into. A team that waves the questions away is selling you the piped-away illusion, and the illusion strands capital the moment the ledger is finally, inevitably, tallied.