Mr. Guth, why should data centers switch from air cooling to liquid cooling?

Because air cooling is simply no longer enough. More powerful AI applications are leading to ever higher power densities in data centers. From 30 kW per rack, it is practically impossible to cool with air. This is due to the lower heat storage and conductivity of air. However, we are already talking to our customers about rack capacities of over 150 kW. The 1 MW cooling capacity provided by our new solution is quickly used up. One example is Nvidia GPUs: while 80% of Nvidia chips sold in the first quarter of 2024 were still designed for air cooling, 85% are expected to be ready for liquid cooling by the first quarter of 2025.

What are the options for liquid cooling?

The most important alternatives are direct liquid cooling, or DLC for short, of the processors, as we offer with the new CDU, and immersion cooling. For DLC, a water-polyglycol mixture is usually used, which has long been tried and tested in industry and automotive engineering. It flows from the distribution unit into the rack and through the heat sinks directly onto the heat-generating components. From there, it takes the heat to the heat exchanger, where it is released. Immersion cooling allows for simpler system designs, but there are still disadvantages and a need for development with regard to the substances in which the entire installation is immersed in this technology. Most liquids are either flammable, harmful to the environment or less compatible with the material. We are therefore currently focusing on DLC with water: the technology can be retrofitted, even if we see it primarily in new data centers, and is closer to what their operators are used to.

Users perceive liquid cooling of servers as unfamiliar, expensive and potentially dangerous. How do you overcome customers' reservations?

This is hardly necessary today. The fundamental reservations about liquids in data centers have very quickly given way to an awareness of the factual necessity. In our discussions with planners and customers, it's hardly a question of 'if', but much more a question of 'how'. However, there are still unanswered questions and hurdles here, and not just with the actual cooling solution, for which no design has yet become established as a standard. What needs to be considered in the piping for the primary circuit on the building side? What changes does the high power density bring to power distribution? How does DLC affect service during operation and ultimately the entire data center? Where possible, we contribute not only our many years of experience with cooling, but also with data centers as an overall system. For example, we know and, if required, can also arrange the appropriate facility trades.

When it comes to the solution itself, it helps that we can supply ready-to-use, secure and certified systems worldwide, in the 21-inch OCP rack or in the classic 19-inch format as part of our system platform. We are convinced that the OCP approach will gain relevance not only for hyperscalers, but also for colocation and other larger data centers up to the enterprise sector. Power, cooling and monitoring are increasingly being integrated directly into the standardized rack as elementary pillars of the IT infrastructure. The 'hot-swappable' modules of our solution can also be moved like server racks. During development, it was important to us that, despite the new technology, we enable handling that fits as well as possible into the familiar processes of a data center. We need to remove the technological, practical and structural hurdles as quickly as possible in order to bring DLC into data centers as an enabling technology for AI. To this end, we are working closely with several large data center developers and are installing a test setup under real conditions in use for a physical research institute at short notice.