Two Roads to a Cool Chip: Direct-to-Chip and Immersion

Once you commit to liquid, you face a fork. There are two dominant architectures for getting heat out of a modern server, and they imply very different components, suppliers, and building designs. Understanding the fork matters because an investor or operator who backs the wrong ecosystem can strand real capital. Let me walk through both with the honesty each deserves.

The first road is direct-to-chip cooling, sometimes called cold-plate or direct liquid cooling. The idea is surgical. You leave the server mostly as it is, air-cooled, but you replace the heat sinks on the hottest components, the processors and accelerators, with sealed metal cold plates. Coolant flows through microscopic channels inside each plate, absorbs the heat, and carries it away through flexible tubing to a manifold at the back of the rack. Direct-to-chip captures roughly 70 to 80 percent of a server's heat into liquid; the remaining fraction, from memory, power supplies, and network cards, is still handled by a modest amount of air. This is the approach the large accelerator vendors have effectively standardized on for their reference designs, which is the single biggest reason it dominates deployments today. It slots into existing racks and existing operational habits with the least disruption.

The second road is immersion cooling. Here you abandon air entirely. You take the server, remove its fans, and submerge the whole board in a tank of electrically non-conductive fluid. In single-phase immersion, the fluid stays liquid and is pumped past a heat exchanger. In two-phase immersion, the fluid is engineered to boil at a low temperature right at the chip surface; the vapor rises, condenses on a cooled coil at the top of the tank, and rains back down, a beautifully passive cycle that moves enormous heat with very little pumping. Immersion captures essentially 100 percent of the heat into fluid and can handle the very highest densities. Its advocates, and I count myself sympathetic, point out that a tank has no fans to fail and no air to condition.

So why has direct-to-chip won the current round? Three practical reasons, and I try to state them without ideology. First, serviceability. A technician can open a rack and swap a direct-to-chip server much as before; pulling a dripping board out of a fluid tank and letting it drain is slower and messier, and it changes the whole maintenance workflow. Second, the fluids. Immersion depends on specialized dielectric fluids, and the best-performing two-phase fluids have been exactly the chemistry now caught in environmental regulation, which we will cover in depth later. That regulatory cloud stalled two-phase investment at a critical moment. Third, and most decisively, the chip vendors validate and warrant their parts for direct-to-chip reference designs. Operators are conservative about warranties on hardware this expensive, and they follow the validated path.

None of this means immersion is a dead end. For the densest future workloads, for edge sites where simplicity beats serviceability, and in regions where the fluid regulations are clearer, immersion remains compelling, and single-phase immersion in particular sidesteps much of the fluid controversy. What I want you to take away is that these are two distinct component ecosystems, not two settings of one dial. Direct-to-chip creates demand for cold plates, flexible hoses, quick-disconnects, rack manifolds, and coolant distribution units. Immersion creates demand for tanks, specialized dielectric fluids, condenser coils, and fluid-handling gear. The two share very few parts. A factory tooled for one does not simply pivot to the other.

For anyone allocating capital or planning a build, the fork has a clear near-term reading. The weight of deployment, the vendor validation, and the supply investment sit with direct-to-chip right now, so that is where the volume constraints and the component shortages will bite first and hardest. Immersion is the option value, the hedge, the thing to watch as fluid chemistry and density curves evolve. Hold both roads in mind as we go deeper, because the supply-chain story differs sharply depending on which one a given operator has bet on. The heat has to go somewhere; the question is only through which set of parts, made by which set of companies.