When you walk into a normal data center, the screens and servers are not the first thing that you see. Why is there so much noise? It’s because of a wall of industrial fans that are working nonstop to keep millions of dollars worth of computer hardware from melting. For many years, that noise was just how things worked. It seems like an issue that can’t be set aside any longer.
As AI workloads get more complex and dense, the heat loads inside modern data centers are getting too high for air cooling to handle. Clusters of high-performance computers, GPU arrays, and the kind of infrastructure that runs big language models all make a lot of heat that standard cooling systems were not made to handle. Engineers have known for years that this is where things are going to crash. What’s new is the sense of urgency.
For example, liquid immersion cooling, in which servers are submerged directly in a thermally conductive fluid, is no longer just a niche experiment. The industry is now taking it very seriously. The idea isn’t completely new. Different kinds of liquid cooling have been used for a long time in supercomputers. But using it at the scale of a data center in ways that make business sense is a whole different issue. It seems like technology has finally caught up with what people need.
The appeal is easy to understand. Liquids are much better at moving heat around than air. Immersion systems can handle heat loads that would be too much for even the most aggressive air-cooling systems because they bathe hardware in dielectric fluid instead of blasting it with cool air. The number of power density per rack has gone up a lot as AI infrastructure has grown. However, when liquid is doing the heating, that number is not as scary. Facilities that used to limit deployments to 10 to 15 kilowatts per rack now have to deal with ones that go well over 100 kilowatts. At that point, cooling with air is basically a lost cause.
What happens inside these liquid systems over time is something that isn’t talked about as much. It turns out that the purity of the coolant is a surprisingly important factor. Even very small particles, like metal debris, ionic contaminants, and fine particulate matter, can stop fluids from moving through microchannels that are as little as 100 to 300 micrometers wide, weaken the fluid’s dielectric properties, and speed up corrosion in ways that you can’t see until something breaks. Taking care of the coolant could become one of the most important operational challenges of next-generation data centers, similar to how cable management and airflow optimization used to be. The hardware gets all the attention. How long that hardware lasts depends on the fluid it’s sitting in.
There are also questions about the fluids themselves that come up with immersion cooling. In the beginning, fluorinated fluids were used a lot because they are stable and don’t conduct electricity, but environmental concerns have made people look more closely at them. Biodegradable fluids break down biologically within standard testing windows are being looked at by some operators as possible alternatives. However, the trade-offs between thermal performance and long-term stability are still being worked out. It’s still not clear if the industry will come together around a single type of fluid or split up into regional and application-specific preferences.
One thing that is clear is that AI infrastructure is changing, and not just at the chip level. Everything is being rethought: the buildings, the fluids, the filtration systems, and the rules for how to run the business. It takes longer and isn’t as exciting as announcing a new model or a faster processor. But as this happens, it’s hard not to notice that some of the most important decisions are being made at the physical level of AI computing.
