# Data Center Cooling Every watt a facility consumes becomes heat, and all of it must leave. Cooling is the chain of hand-offs that carries that heat from the surface of a chip to the open air, and each link in the chain has its own machinery, its own failure modes, and its own limits. The chain is worth learning whole, because the capability of a modern rack is set by it as firmly as by the silicon inside. > [!note] The one-sentence version > Silicon to plate, plate to coolant, coolant to facility water, facility water to the sky; five hand-offs, each with its own equipment, and the weakest one sets the density of the building. ## Why it is the governing constraint Electrical capacity decides how much a facility may draw; thermal capacity decides how much it can actually run. The two are rarely equal, and the thermal side is usually the tighter. Accelerator packages now dissipate on the order of a kilowatt each, racks pass one hundred kilowatts and current rack-scale systems reach the region of two hundred and twenty-five to two hundred and forty-five kilowatts, and heat flux at the die has reached levels once discussed in aerospace. Those numbers are why thermal engineers now sit with chip architects, and why a rack arrives from the factory with its plumbing installed. ## The chain, link by link 1. **Die to package to plate.** Heat crosses from silicon through the package lid and the thermal interface into a [[wiki/Cold Plate|cold plate]], or into a heatsink where air cooling still serves. This joint, microns thick, governs everything above it. 2. **Plate to coolant.** Liquid in microchannels carries the heat away; see [[wiki/Direct-to-Chip Liquid Cooling|Direct-to-Chip Liquid Cooling]]. Where boards are submerged instead, this and the previous step collapse into one; see [[wiki/Immersion Cooling|Immersion Cooling]]. 3. **Technology loop to facility loop.** A [[wiki/Coolant Distribution Unit|coolant distribution unit]] transfers heat between the clean, controlled loop serving the servers and the building's water, while isolating one from the other and holding flow, pressure, and temperature to specification. 4. **Air-cooled residue.** Memory, network interfaces, power conversion, and drives still exhaust to air, captured either by a [[wiki/Rear-Door Heat Exchanger|rear-door heat exchanger]] or by the room's own system; see [[wiki/Air Cooling|Air Cooling]]. 5. **Facility water to atmosphere.** The [[wiki/Chilled Water Plant|central plant]] cools the water, with compressors when outdoor conditions require and without them when they do not, and a [[wiki/Cooling Tower|tower or dry cooler]] gives the heat to the sky. Where a customer exists for warm water, part of the heat leaves as a product instead; see [[wiki/Heat Reuse|Heat Reuse]]. ## The temperatures that decide the design The single most consequential number in a cooling design is the temperature at which heat is captured. Capture it cold and the plant must run refrigeration most of the year. Capture it warm, and the same heat can often be rejected by outdoor air alone, which removes the compressors from everyday service, improves [[wiki/Power Usage Effectiveness|power usage effectiveness]], cuts water consumption, and makes the heat valuable to a neighbor. Liquid cooling helps precisely here: a cold plate works well with water far warmer than an air coil needs. The industry's move toward warm-water liquid capture is therefore an efficiency and water story at least as much as a density story. ## Continuity Thermal continuity is more urgent than electrical continuity. A liquid-cooled hall loses thermal margin within seconds of flow stopping, because the mass of coolant near the chips is small and the heat is intense. Pumps, coolant distribution units, and plant controls are therefore carried on uninterruptible power, thermal storage covers the generator start interval, and loops are valved so any single pump, exchanger, or unit can be isolated while the rest runs; the redundancy discipline of [[wiki/Redundant Power Feeds|A and B paths]] applies to pipes as rigorously as to conductors. Leak detection runs beneath the piping, and the quick disconnects that make trays serviceable are specified and sampled as a fleet because a facility holds thousands of them. ## Stack position - **Upstream:** electrical supply, water supply, and the weather. - **Downstream:** every component, and therefore the facility's achievable density and its sustained performance. - **Function and vendor:** heat removal is unavoidable; each link has several mature technologies and many suppliers, and the choices interact, which is why cooling is designed as a chain rather than assembled from parts. - **Degraded case:** devices throttle to protect themselves. The failure is quiet: the hardware keeps working and delivers a fraction of what it was bought to deliver. ## Lineage and trajectory Mainframes were water-cooled, the microcomputer era made air universal, and the accelerator era has brought liquid back for physical reasons rather than fashionable ones. Each step of the return has been a step toward integration: plates bolted on, then plates shipped with the tray, then racks delivered as plumbed systems, then cooling geometry designed into the package. The trajectory runs toward capture fractions approaching the whole rack, warmer supply temperatures, two-phase operation, heat delivered to customers rather than to the sky, and eventually cooling structures inside the silicon. ## Relationships - **composed of:** [[wiki/Cold Plate|Cold Plate]] · [[wiki/Direct-to-Chip Liquid Cooling|Direct-to-Chip Liquid Cooling]] · [[wiki/Coolant Distribution Unit|Coolant Distribution Unit]] · [[wiki/Rear-Door Heat Exchanger|Rear-Door Heat Exchanger]] · [[wiki/Air Cooling|Air Cooling]] · [[wiki/Chilled Water Plant|Chilled Water Plant]] · [[wiki/Cooling Tower|Cooling Tower]] - **alternative path:** [[wiki/Immersion Cooling|Immersion Cooling]] - **constrains:** [[wiki/Data Center Rack|Data Center Rack]] · [[wiki/Rack-Scale Computing|Rack-Scale Computing]] · [[wiki/Utilization|Utilization]] - **measured by:** [[wiki/Power Usage Effectiveness|Power Usage Effectiveness]] · [[wiki/Data Center Water Use|Data Center Water Use]] - **produces value through:** [[wiki/Heat Reuse|Heat Reuse]] - **part of:** [[wiki/Facilities and Thermal Management|Facilities and Thermal Management]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** the five-stage heat path and its hand-offs; accelerator packages in the kilowatt class; the seconds-scale thermal margin of a liquid-cooled hall after flow loss; uninterruptible support of pumps and controls; valved redundancy and leak detection; the relationship between capture temperature, compressor hours, water consumption, and heat value. - **Established (datestamped):** as of October 2026, current rack-scale systems operate in the region of two hundred and twenty-five to two hundred and forty-five kilowatts per rack with closed-loop liquid cooling. - **Strongly indicated:** warm-water capture as the dominant design direction; vendor and operator practice converge on it and long-run facility data is still accumulating. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; rack power envelopes, copper cold plates, closed-loop chilled water, and circulated flow at a current facility. - ASHRAE thermal guidelines and liquid-cooling classes for data processing environments, as current October 2026.