Cool calculation

Electronics and liquids don’t mix—every child knows that. Lucas Voigt is an engineer working to disprove this supposed certainty. In a laboratory at Evonik’s Darmstadt location, he opens a server cabinet, pulls out one of the drawers containing a flashing computer module, and lifts it up. An oily liquid is dripping from circuit boards, processors, and memory modules.

What at first glance looks like extreme damage is actually intentional. The entire server was submerged in a cooling fluid. “The electronic components don’t mind,” says Voigt. He and his colleagues are conducting pilot-scale tests with biodegradable fluids for what is known as immersion cooling. It’s an answer to key questions currently facing the IT industry: How can high-performance computers be cooled when a cold airflow is no longer sufficient? And how might waste heat be utilized for other purposes?

The AI boom is fueling the heat problem here. The major American technology companies such as Amazon, Microsoft, Alphabet, and Meta are expected to invest US$620 billion in data centers in 2026. These companies need increasingly powerful chips to manage their workloads. The AI models run on graphics processing units (GPUs) that are capable of processing unimaginable amounts of data in a very short time. To do this, they need significantly more energy than the CPUs in conventional servers.

Power consumption, measured in kilowatts (kW) per rack (server cabinet) is a key metric. Standard data centers are designed for 5 to 10 kW per rack, while AI-optimized facilities can handle 40 to 80 kW.

“The current generation of Nvidia graphics cards already uses as much power as a one-family home, and the next one will use significantly more” says Professor Dieter Kranzlmüller from LMU Munich. (see Debate on pages 22–26). Kranzlmüller also serves as Chairman of the Board of Directors of the Leibniz Computing Center. Anyone building a new data center today must therefore plan for power densities that didn’t exist two years ago. Floors, grid connections, and power distribution systems must be designed to accommodate new dimensions. Chips ultimately convert nearly all of the electrical energy they absorb into heat. As a result, the heat generated inside a server increases along with the server’s performance, and it is crucial to ensure adequate cooling.

“The demand for new cooling systems is growing, because as power increases, established designs are reaching their limits,” says Mario Gomez. “Efficient cooling is becoming a bottleneck for data center operators.” Gomez, a chemist, heads Evonik’s program for the development of immersion cooling as part of the oil additives business. Together with Voigt and other colleagues, he is looking to develop an alternative to existing cooling technologies.

To date, two methods in particular have been used (see the diagram on pages 20–21). Air cooling is the established standard. In this process, air flows through the server racks, absorbs heat, and releases it into the surrounding environment via cooling systems.

This type of cooling is still suitable for conventional data centers. In the age of AI, however, as supercomputers generate ever-increasing amounts of heat, air alone is no longer sufficient. The problem is not so much the temperature in individual server racks as the extremely high heat density of individual AI chips. Huge fans would be needed to constantly dissipate heat from these hotspots using conventional methods, and electricity consumption would rise even more sharply.

In AI data centers, air cooling is therefore often supplemented today by another method: direct-to-chip cooling (DTC). Here, a fluid—usually a mixture of water and glycol—flows directly past the server components that generate the most heat. It thus captures waste heat directly at the source. The liquid then flows through a heat exchanger, where it is cooled again and returned to the chip.

Combined air and DTC cooling currently keeps most data centers running reliably. However, this approach also has its limitations, given the rapid advances in chip architecture and server performance. When training an AI model, numerous graphics processing units run at a consistently high load for long periods of time—generating a corresponding amount of heat.

A third cooling method—immersion cooling, also known as submersion cooling—could herald a paradigm shift. In this process, the entire server is submerged in dielectric fluids. These are liquids that do not conduct electricity—mineral oils or ester compounds, for example. A computer can be completely submerged in them: The electronic components come into direct contact with the liquid without causing a short circuit or other damage. The technology is not yet being used on a large scale in data centers, but that could soon change..

The experts Mario Gomez and Lucas Voigt from Evonik want to help make immersion cooling a widespread application by providing the chemistry needed for it. “There’s a lot going on in the fluids market right now,” says Gomez. “Of course, other suppliers are working on similar products. But Evonik is getting an early start and is pursuing an approach based on biodegradable substances. It’s a big opportunity for us.”

According to the market research firm “Markets and Markets,” the global market for data center cooling fluids is expected to grow to approximately US$2 billion by 2032 at an annual growth rate of over 30 percent, with about half of that growth expected to come from immersion cooling. Around half of that might be accounted for by the cooling fluids used. Evonik is targeting the emerging market with biodegradable ester-based fluids. In doing so, the experts can draw on the company’s many years of experience in the field of oil and lubricant additives. “We specialize in modifying the thermal and physical properties of liquids,” says Gomez. “This allows us to adapt them to the requirements of different applications.” In addition to data centers, he is looking at, among other things, industrial energy storage and electric mobility applications.

According to Gomez, immersion cooling for servers offers several advantages over established cooling technologies: “It enables particularly efficient heat exchange because the electronic hardware is directly surrounded by the cooling fluid.” Immersion cooling also allows for operation at significantly higher power densities without any risk of overheating. “The uniform cooling reduces thermal fluctuations,” Gomez explains. “That’s good for the service life of the server components.” In addition, more servers can fit into the same space, while the need for additional cooling systems decreases. This saves space and energy and lowers operating costs.

The fluids used in immersion cooling systems must be easy to handle, non-toxic, and environmentally friendly. Chemical stability is also essential to ensure their performance over the years without the need to replace them.

Meeting these requirements is Holger Becker’s job. The chemical lab technician at Evonik tests the properties of the ester fluids that will later circulate around the electrical components in the server. He has already tested viscosity, flash point, density, and thermal and electrical conductivity in more than 30 different candidates. “There’s no such thing as a jack-of-all-trades,” he says. “We’re looking for the best combination.” And—of all things—artificial intelligence is helping him do that. It proposes ester fluids that seem particularly promising due to their properties.

To test the cables, Becker places them in the ester fluids and then puts them in a heating block—a kind of oven—for several weeks. This shows whether the cable materials can withstand contact for a long period of time without sustaining damage.

The most promising fluids are sent to Lucas Voigt’s testing lab, which is where the server rack that Evonik uses to test the liquids under realistic conditions is located. To ensure better comparability, the server units run through a fixed program. “The tests are intended to show which fluids perform well in practice and achieve the desired effectiveness,” says Mario Gomez.

Evonik thus aims to position itself as a partner in the emerging value chain for immersion cooling. The goal is a portfolio that includes both individual ester components and custom blends. The key factor here is, first and foremost, collaboration with the manufacturers of the cooling systems. They supply tanks, cassettes, and pipes and must ensure that the fluids and hardware work together reliably over the long term. Looking ahead, server manufacturers themselves will also play an important role, as they are increasingly offering their own programs for validating fluids in direct cooperation with tank manufacturers.

This isn’t just about future data centers, as existing systems designed for air cooling could also be converted to immersion cooling, provided they are compatible. A startup from the Netherlands has set itself exactly that goal. iXora develops cassettes in which individual servers can be operated in a dielectric fluid. The key feature here is that instead of submerging the server cabinets in tanks, the cooling technology can be integrated into the widely used standard 19-inch racks. The cabinets themselves remain dry and are still accessible for maintenance work.

Immersion cooling offers another advantage that is likely to become increasingly important in the future: The fluids used transfer heat at a higher temperature than other coolants. This makes them particularly attractive for use in heating systems. Could servers thus become providers of both computing power and heat? “This is no longer a vision—it’s already a reality,” says Patrick Petzenhauser of the Lindner Group—a building systems manufacturer. He is leading a project in which Lindner aims to convert the waste heat from servers into heat for heating systems, thereby making it usable.

The heart of the project is a module called BitHeat, which can be connected to a building’s heating system. This device takes up less than half a square meter of space; inside, two fluid circuits meet: A fluid transports waste heat from a server to a heat exchanger. There, the thermal energy is transferred into the water circuit.

“For this water to be used effectively for heating, it must have a temperature of 55–60 degrees Celsius,” says Petzenhauser. “We can reach this target temperature when we use immersion cooling.” Together with partners such as Evonik, Petzenhauser is working at Lindner to optimize the interfaces between servers, cooling fluids, and building systems.

So far, the BitHeat module has not been used in central data centers, but rather in locations where servers operate locally—for example in technical rooms in hotels or office buildings. If there are a large number of servers, it is possible to operate several of the devices in one location, says Petzenhauser. Together with his team, he is exploring scalability in several directions. He sees one potential application in factories, for example, since quite a few industrial production processes require an energy supply in the form of heat. In the case of larger data centers, Petzenhauser believes local district-based concepts are feasible. Municipalities or neighborhoods would be supplied by a server farm via a district heating network; in the case of decentralized computers, the heat would be used directly within the building. “The heat is there in any event, so it would be a shame not to use it,” says Petzenhauser.

A value chain can emerge anywhere—from the chemical industry to server infrastructure or the underfloor heating system of an office. With its bio-based fluids from Darmstadt, Evonik aims to become a part of such developments. Will immersion cooling become widespread from here on out? “At the end of the day, the server workload determines the cooling strategy,” says Gomez. However, in an era of explosive growth in AI workloads, immersion cooling can evolve from a niche technology into a key strategy. Where servers achieve extreme power densities, this technology definitely becomes a viable option. The same is true in situations in which any waste heat generated needs to be utilized as efficiently as possible. In this way, immersion cooling could herald a “thermal revolution” in several respects in the age of artificial intelligence.