About this transcript: This is a full AI-generated transcript of How Liquid Cooling Moves Heat Out of AI Data Centers from TechEx-K, published November 22, 2026. The transcript contains 1,768 words with timestamps and was generated using Whisper AI.
"Inside an AI data center, thousands of processors are performing billions of calculations every second. But almost all the electricity consumed by those processors eventually becomes heat. If that heat cannot escape quickly enough, chip temperatures rise, performance drops, and servers may shut..."
[00:00:00] Speaker 1: Inside an AI data center, thousands of processors are performing billions of calculations every second. But almost all the electricity consumed by those processors eventually becomes heat. If that heat cannot escape quickly enough, chip temperatures rise, performance drops, and servers may shut themselves down to prevent permanent damage. This means the true limit of an AI system is not just how fast its processors can calculate, it is also how quickly the facility can remove the heat they produce. For decades, data centers relied on cold air and powerful fans. But modern AI hardware is generating too much heat inside to little space. Air cooling is beginning to reach its practical limits. The new solution sounds dangerous. Bring liquid to within millimeters of some of the most expensive electronic components ever built. So how can liquid cool thousands of AI processors without destroying them? To understand that, we need to follow the complete journey of heat, from the surface of a silicon chip, to the moment it finally leaves the building. An AI GPU contains billions of transistors switching on and off to process data. As electrical current moves through those transistors, memory circuits, and microscopic metal connections, energy is converted into heat. That heat is concentrated inside a very small area of silicon, creating an extremely high heat density. One processor already consumes a significant amount of power. Multiple processors are then installed inside one server, and dozens of servers can be packed into a single rack. The result is a concentrated heat load that can reach tens or even hundreds of kilowatts. But total heat is only part of the problem. Before it can escape, heat must pass through the silicon, the chip package, a thermal interface material, and a cooling device. Every layer creates thermal resistance. If any part of this path transfers heat to slowly, energy begins accumulating inside the chip, even while the rest of the cooling system appears to be operating normally. In a traditional air-cooled server, heat moves from the processor into a metal heatsink. Fans force cold air through hundreds of thin metal fins, allowing the moving air to absorb heat, before leaving the rear of the rack. The hot air is directed toward the cooling equipment of the building, chilled again, and returned to the front of the servers. Data centers separate these flows into cold aisles and hot aisles to prevent them from mixing. This system works well at moderate power densities, but air cannot carry nearly as much heat per unit of volume as a liquid. As processors become hotter, the data center must move more air. That requires larger heat sinks, faster fans, more electricity, and more space inside the server. Airflow can also be blocked by cables, memory modules, and other components. Some areas receive plenty of air, while critical hot spots remain difficult to cool. Eventually, increasing fan speed produces less and less improvement. Instead of moving enormous amounts of air through an entire room, engineers can move a much smaller volume of liquid directly to the main source of heat. In direct-to-chip liquid cooling, water does not touch the GPU itself. A sealed metal component called a cold plate is mounted on top of the processor. Between the chip and the cold plate is a very thin thermal interface material. It fills microscopic gaps between the two surfaces, allowing heat to move more efficiently into the metal. Inside the cold plate is a network of narrow channels. Coolant flows through these channels, but remains completely enclosed and separated from the electronics. Heat travels from the silicon through the chip package and into the cold plate. The moving coolant absorbs that energy, becomes warmer, and carries the heat away. The word cold can be misleading. A cold plate does not need to be close to freezing. In many systems, the coolant is deliberately kept warm enough to prevent condensation. If a pipe or cold plate falls below the dew point of the surrounding air, water vapor can condense on its surface. A few drops of water inside an operating server could be catastrophic. The coolant temperature must therefore remain low enough to cool the processor, but high enough to keep every exposed surface safely above the dew point. After absorbing heat from the chip, the warmer coolant leaves through a return hose. Each cold plate has a supply line and a return line. Hoses from multiple processors connect to a manifold inside or beside the rack. This manifold distributes coolant evenly to individual servers and collects the heated liquid returning from them. Quick disconnect fittings allow technicians to remove a server without draining the entire cooling network. Isolation valves can shut off a damaged branch, while sensors monitor temperature, flow rate, and pressure throughout the system. But the liquid flowing through the servers is usually not connected directly to the building's water system. Between the two networks is a device called a coolant distribution unit, or CDU. The CDU contains pumps, controls, sensors, filters, and a heat exchanger. On one side is the carefully controlled technology cooling loop that serves the processors. On the other side is the facility water loop connected to the building. The two liquids flow past opposite sides of thin metal surfaces inside the heat exchanger. Heat passes through the metal, but the liquids never mix. This separation allows the server loop to maintain its own pressure, temperature, cleanliness, and chemical quality while remaining protected from changes in the larger building system. The CDU is the transfer station that moves heat from a rack of AI hardware into the cooling infrastructure of the entire facility. Once the facility water receives that heat, it travels toward equipment designed to reject it outdoors. Depending on the climate and the design of the data center, the system may use dry coolers, cooling towers, chillers, or a combination of them. Heat moves from the facility water into the surrounding environment, and the cooler water returns to the CDU to begin the cycle again. The coolant does not create cold, and it does not make heat disappear. It simply transports concentrated heat from a tiny area inside a processor to a much larger outdoor system where that heat can be released efficiently. In some installations, the returning liquid is warm enough for its energy to be reused. Waste heat from computing can help warm nearby buildings, heat water, or support certain industrial processes. From a thermal perspective, an AI data center behaves almost like a factory that converts electricity into computation and heat, then continuously transports that heat away. Does liquid cooling eliminate every fan? Usually not. GPUs and CPUs may produce most of the heat, but memory, storage devices, power supplies, and other board components still become warm. Many direct-to-chip systems continue using a smaller amount of airflow to cool these lower-density components. The important difference is that liquid removes most of the concentrated heat directly from the processors. The fans no longer need to carry the entire GPU heat loadout of the rack, so they can operate with less air flow and consume less power. This creates a hybrid cooling system. Liquid handles the most intense hot spots, while air removes the remaining heat. Of course, putting liquid near electronics introduces a serious risk. Every cold plate, hose, seal, and connector must remain reliable under pressure for years. Drip-resistant quick connectors reduce coolant loss during maintenance. Moisture sensors can detect small leaks, while pressure and flow sensors identify abnormal conditions before the problem becomes visible. If one branch fails, valves can isolate that server from the rest of the rack. Many coolant distribution units also include redundant pumps, allowing circulation to continue if a primary pump stops. If coolant flow decreases and chip temperature begins rising, the processor can lower its operating frequency and reduce power. If temperatures continue climbing, the server can perform a protective shutdown before permanent damage occurs. Reliability therefore depends on more than preventing leaks. The system must detect, isolate, and respond to a failure before it spreads. Cold plates are not the only way to cool computers with liquid. In immersion cooling, entire server boards are submerged in a special dielectric fluid that does not conduct electricity. The fluid touches processors, memory, and other components directly absorbing heat across the whole system. In single-phase immersion cooling, the fluid becomes warmer but remains liquid. It is pumped through a heat exchanger and then returned to the tank. In a two-phase system, a specially selected fluid can boil at the surface of a hot component. The vapor rises, releases its heat at a condenser, and falls back into the tank as liquid. Immersion cooling can remove large amounts of heat and eliminate many server fans, but it also changes how equipment is designed, installed, and maintained. Direct-to-chip cooling fits more easily into familiar rack layouts. While immersion cooling requires tanks, compatible materials, and different service procedures. Neither approach is automatically best for every facility. The correct choice depends on power density, existing infrastructure, maintenance requirements, and operating cost. The next challenge is moving cooling even closer to the source of heat. A cold plate currently sits above the chip package, but future processors may contain more layers of silicon and vertically stacked memory, allowing heat to become trapped deeper inside the device. Researchers are developing microscopic liquid channels that can operate closer to the silicon or even between stacked layers. Other designs allow coolant to change phase inside a cold plate, using boiling to absorb large amounts of energy, without requiring a large temperature increase. But bringing liquid closer to the silicon also raises the requirements for cleanliness, sealing, and reliability. A tiny blockage or contaminant could prevent coolant from reaching a critical hot spot. Liquid cooling does not stop a processor from generating heat, and it does not provide free cooling. What it does is shorten the path that heat must travel. Instead of forcing all that energy through a large heat sink and into a room full of air, a cold plate captures it directly at the GPU. Coolant carries it to a manifold. The CDU transfers it into the facility water system, and the external cooling equipment finally releases it outside the building. The more computing power engineers place inside a small area, the more powerful this thermal transport network must become. In the age of artificial intelligence, the fastest processor may not always determine how powerful a data center can be. Sometimes, the real limit is simply how quickly we can keep it from overheating.