Immersion and the Operating Questions
On December 20, 2022, 3M announced that it would exit the manufacture of per- and polyfluoroalkyl substances, known as PFAS, by the end of 2025. The announcement, also filed with the Securities and Exchange Commission, meant the end of 3M's Novec and Fluorinert engineered fluids. Most readers had never heard of them. For one branch of data center cooling, though, those fluids were close to the whole story, and 3M's decision is a good way into the other main family of liquid cooling, which is immersion.
Immersion cooling does away with cold plates altogether. Servers are submerged in a bath of dielectric liquid, one that does not conduct electricity, so the whole board, every component, is in contact with the coolant. There are two versions. In single-phase immersion, the liquid, usually an engineered hydrocarbon or synthetic oil, stays liquid. Pumps circulate it through a heat exchanger that passes the heat to facility water. In two-phase immersion, the liquid boils at a low temperature on the hot surfaces of the chips, the vapor rises, condenses on a cooled coil above the bath, and drips back down. Boiling is an extremely effective way to remove heat, which is why two-phase systems attracted interest for the hottest parts.
The fluids that boil at the right temperature and are safe around electronics have mostly been fluorinated chemicals, which is where PFAS regulation and 3M's exit come in. Other chemical makers have offered alternatives, and research continues, but the episode illustrates a risk that is easy to miss: a cooling architecture can depend on a narrow and politically exposed supply chain. Single-phase immersion avoids the most contested chemistry, though it raises its own questions about fluid cost, fluid life and compatibility with plastics and cables.
Immersion has other practical costs. Tanks change the shape of the data hall, laying servers horizontally in baths rather than standing them vertically in racks. Servicing a board means lifting it out of liquid and letting it drain. Hardware warranties and component qualification have to account for long-term exposure to the fluid. None of these problems is insurmountable, and immersion is used in production in some settings, but they explain why the industry's main path has been direct-to-chip. The large accelerator vendors ship cold-plate designs, and the rack-scale systems discussed earlier arrive plumbed for them. When the hardware comes ready for one method, the building follows.
Whichever liquid method is chosen, operators face a common set of questions that did not arise with air. The first is leaks. Quick-disconnect fittings, hoses and manifolds are new failure points, and operators add leak detection cables, drip trays and valves that can isolate a rack. The second is water quality in the technology loop, which must be monitored and treated much as a boiler would be, because corrosion or particles can block the fine channels in a cold plate. The third is resilience. With air, a cooling failure gives a room some minutes of thermal inertia. With a liquid-cooled rack running at very high power, the margin before chips throttle or shut down can be much shorter, which puts emphasis on redundant pumps, backup power to the distribution units, and careful control logic.
The standards and specification work reflects these concerns. The Open Compute Project runs an Advanced Cooling Solutions effort that publishes requirements and guidance for cold plates, immersion and the equipment around them, and the technical committee of the American Society of Heating, Refrigerating and Air-Conditioning Engineers, or ASHRAE, has published guidance on liquid cooling for data processing environments. These documents are where terms get defined and where operators, hardware makers and suppliers agree on things like fitting types and fluid chemistry. Reading them is dull. It is also the best available evidence of what the industry has actually converged on, as opposed to what a single vendor hopes it will converge on.
Faraday's read is that immersion will remain a real but specialized option, with direct-to-chip as the default for accelerator racks for the foreseeable future, because the default is set by the hardware vendors and they have chosen cold plates. That view should be revisited if a major accelerator maker ships a product designed for immersion first. Until then, the question for most buyers is not which liquid method, but how quickly their buildings can accept liquid at all.
There is one more consequence to carry into the last module. Liquid cooling moves heat efficiently to the facility water system, but it does not by itself decide how the building rejects that heat to the outdoors. That choice, between evaporating water and not evaporating it, is where the water debate lives.