Direct-to-chip cooling captures 70-80% of rack heat. A technical guide to cold plate loops, ASHRAE water classes, and the residual air-cooled load that.
Direct-to-chip liquid cooling stopped being optional somewhere around the 100 kW rack. NVIDIA's GB200 NVL72 draws roughly 120 kW in a single cabinet and ships as a liquid-cooled architecture; air cooling is not offered as an alternative. The industry conversation has moved from whether to plumb the rack to how to plumb it.
That conversation has a blind spot. A cold plate cools what it is bolted to. Everything else in the chassis still depends on moving air, and as the liquid loop absorbs the processors, the remaining air-cooled load gets harder to serve, not easier. This is a guide to both halves.
A cold plate is a machined or skived metal block, usually copper, clamped against the lid of a die package with a thermal interface material between them. Coolant channels are routed through the block. Heat conducts from junction to lid, from lid through the plate wall, and then transfers by convection into the fluid. The whole design question is how little temperature rise you can accept at each of those steps for a given flow rate and pressure drop.
It is worth placing the approach against the two alternatives. Immersion cooling submerges the entire board in a dielectric fluid, so every component is cooled and nothing needs an air path, at the cost of a different serviceability model and a long materials compatibility qualification. Rear-door heat exchangers do the opposite: they leave the server untouched and hang a liquid coil on the cabinet exhaust, converting rack air heat into liquid heat at the rack boundary. Direct-to-chip is the middle path. It targets the highest heat-flux components and leaves the rest of the rack, and the rest of the service procedure, broadly as it was. That is the main reason it won.
The technique is also written direct liquid cooling, or DLC. Two variants are in use. Single-phase loops circulate water or a water-glycol mixture and carry heat away as a sensible temperature rise; this dominates deployments because the fluid, the fittings and the pumps are all mature. Two-phase loops use a refrigerant that boils inside the plate, exploiting latent heat to hold the plate close to isothermal at higher heat flux, in exchange for pressure management and refrigerant handling. Neither is universally correct.
Accelerator package power has moved from the low hundreds of watts to the high hundreds and past a kilowatt. The number that binds, though, is not the package total. It is heat flux, watts per square centimetre of die and lid area, because that is what sets the temperature drop across every layer in the stack. A 1 kW package spread over a large multi-die substrate and a 500 W package on a small die can present very different problems, and the smaller one is often the harder one.
Air is a poor working fluid at high flux for one reason: its volumetric heat capacity is low. To carry more heat at a fixed temperature rise you need proportionally more mass flow, and fan power rises roughly with the cube of flow while the heat removed rises linearly. Acoustic output follows the same unhappy curve. So the cost of the last few kilowatts per rack is not linear, it is punitive.
In practice most operators find air remains workable to roughly 30 to 40 kW per rack, given aisle containment, disciplined blanking and elevated inlet temperatures. Above that band the fan power share, the noise and the achievable die temperatures all deteriorate together. The GB200 NVL72 at approximately 120 kW per rack sits far outside it, which is why liquid cooling there is an architectural requirement rather than a configuration option. The wider picture of how these bands map to cooling architectures is covered in our guide to data center thermal management.
Follow the heat in order.
The specification that ties the loop to the facility is supply temperature. ASHRAE's W-class categories give operators a shared vocabulary for it, from tightly chilled supply through to warm-water classes. The direction of travel is upward: the higher the supply temperature the technology loop can tolerate, the more hours per year the site can reject heat without mechanical cooling, and the better the case for heat reuse. Warm supply water costs you plate temperature margin, so the class you choose is a negotiation between the silicon and the facility.
On interoperability, the Open Compute Project's ACS Liquid Cooling Cold Plate Requirements document is becoming the baseline reference for plate geometry, fitting types, pressure ratings and leak test procedure. Asking whether a plate is conformant is now a reasonable procurement question rather than a purist one.
Finally, the failure modes engineers actually plan for:
Cold plates typically capture 70 to 80 percent of rack thermal load. That figure is not a coincidence or a product limitation; it is roughly the CPU and GPU share of chassis power. The balance is real, it is distributed, and it does not go away:
Then there is the part that surprises people. When liquid removes the dominant heat source, it also removes the reason for the bulk airflow that used to sweep past everything else on its way to the processor heatsink. The residual components were never designed with their own airflow budget; they were cooled incidentally, downstream of something large and hot. Take the large hot thing out of the air path and the incidental cooling goes with it.
Server designers respond in the obvious way, by keeping chassis fans anyway. The consequence is that a rack described as liquid-cooled very often still contains dozens of rotating parts, sized for a load that is now a minority of the total.
Chassis fan power does not vanish with the transition, and in a hybrid design it can remain a meaningful share of IT load, because fans sized for the residual load are still fans running continuously. Acoustic output is the second cost, and in dense halls it has become a workplace exposure question rather than an aesthetic one.
Reliability is the third and the least discussed. Once the processors are on liquid, the fan is frequently the only remaining electromechanical component in the box, which makes it the dominant wear-out failure mode by default. Fourth, any air path is also a particulate path, so filtration, ingress protection and the maintenance those imply stay on the list.
It is worth being precise about what rear-door heat exchangers fix here. They intercept heat at the rack boundary and relieve the room. They do nothing about airflow inside the chassis, because the heat still has to leave the components by air before the coil can see it.
To be clear about scope: ionic wind does not replace a cold plate, and it is not a candidate for cooling a kilowatt-class die. Anyone claiming otherwise is describing a different physics than the one at work.
Its relevance is the residual load specifically. Those components share a profile: localised, low to moderate heat flux, space constrained, acoustically sensitive, and today served by rotating parts.
The mechanism is straightforward. A dielectric barrier discharge actuator places electrodes either side of a thin dielectric and drives them with a high-voltage AC waveform. Air ionises at the exposed electrode edge, the ions accelerate in the local field, and momentum transfer to neutral molecules entrains bulk air along the surface. The result is directed airflow from a flat device with no blades and no bearings. The full physics is set out in our explainer on how ionic wind works, and the position of the approach among fanless alternatives in our guide to solid-state cooling.
The useful framing is hybrid rather than competitive: liquid for the processors, solid-state air for the components the plate cannot reach. The same argument applies outside the data hall wherever enclosures are sealed or noise limited, which is why it recurs in edge AI thermal management.
Questions worth asking a vendor, in roughly this order:
No. Cold plates typically capture 70-80% of rack heat. VRMs, memory, optics and NICs remain air-cooled, so most liquid-cooled servers retain chassis fans.
Direct-to-chip clamps a coolant-carrying plate to specific high-power packages and leaves the rest of the board in air. Immersion submerges the entire board in dielectric fluid, cooling everything but changing serviceability and materials compatibility.
Roughly 30-40 kW per rack with good containment. Above that, fan power, acoustics and achievable die temperatures make air impractical.
Single-phase water or water-glycol dominates deployments today because it is simpler and the supply chain is mature. Two-phase offers higher heat flux capability and more uniform plate temperature but adds refrigerant handling and pressure management.
No. Ionic wind is an air-mover technology. It is relevant to the residual air-cooled components in a liquid-cooled chassis, not to the processor die itself.
YPlasma builds solid-state DBD plasma actuators that move air with no blades, no bearings and no moving parts, for the components a cold plate never touches.