Liquid cooling for servers: How immersion cooling works and what its limitations are

Immersion cooling sounds like the future of data centers: servers are submerged in a tank of non-conductive liquid that draws heat directly from the components. The reality is more cautious. Immersion cooling is just one form of liquid cooling, and according to available reports, a more common method for high-performance AI systems involves cooling where the liquid flows through heat sinks directly on the chips.

Why Air Is No Longer Enough

Typical data centers still operate at low power densities. According to a 2026 Uptime Institute survey, the most commonly reported density is 11 kW per rack; in a 2025 survey, 82% of respondents had a rack with a maximum load of less than 30 kW, and only 9% had at least 50 kW. AI systems, however, are in a completely different league: NVIDIA’s reference architecture for the GB300 NVL72 system specifies up to 142 kW per rack.

According to analysts at the Uptime Institute, air cooling above approximately 50 kW per rack is becoming less practical due to space constraints within the rack and chassis. Therefore, direct liquid cooling is gaining traction at higher densities. Data on average density vary depending on the study (AFCOM reports much higher values for 2026 than Uptime), so they should be considered approximate.

How Immersion Cooling Works

In immersion cooling, components are submerged in a dielectric—that is, an electrically non-conductive—liquid. There are two basic approaches.

Single-phase cooling

The liquid remains in a liquid state throughout the entire process. The heated liquid is pumped through a heat exchanger, cooled, and returned to the tank. Such systems are simpler and use fluids free of fluorinated compounds, such as those based on hydrocarbons or esters.

Two-phase cooling

When the liquid comes into contact with the hot chip, it boils; the vapor rises to the condenser, condenses, and flows back down. Evaporation removes a great deal of heat, so this method is theoretically extremely efficient. In practice, however, it relies on special fluorinated liquids, and it is precisely these that are the problem today.

Immersion cooling and direct-to-chip cooling

In direct-to-chip (D2C) cooling, the liquid flows through cold plates mounted on processors and accelerators. The rest of the server—such as memory and power supplies—is often still air-cooled. The advantage is that it can be implemented in existing server racks and data centers. Several market analyses cite D2C as the dominant liquid cooling solution, but the specific market shares vary significantly across sources (ranging from approximately 43% to 55% in various reports), so I do not present them as fact.

Immersion cooling requires tanks, different floor load-bearing capacity, different maintenance, and often modified servers. For existing data centers, it is therefore more challenging to retrofit (this is my explanation based on the nature of the solution, not data from a source).

The Problem with PFAS Substances

Two-phase immersion cooling has long relied on Novec-type fluids from 3M, which are classified as PFAS, the so-called “forever chemicals.” 3M states that it completed the phase-out of PFAS production by the end of 2025. According to industry reports, this has significantly slowed the growth of two-phase systems. In its July 2026 report, the environmental organization Food & Water Watch also highlights the risk of these substances leaking into the environment. It is an advocacy organization, so its claims should be viewed with this in mind.

According to one overview, the European Chemicals Agency (ECHA) is expected to issue final opinions on a general restriction of PFAS by the end of 2026. For suppliers of two-phase systems based on fluorinated liquids, this means legal uncertainty. I could not find this information in the agency’s primary source, so please treat it with caution.

Advantages and Limitations

• Advantages (according to manufacturers and overviews): high heat transfer density, fewer or no fans, potentially lower energy consumption for cooling.
• Maintenance: When submerged, it is more difficult to replace components and handle the fluid.
• Compatibility: Servers and materials must withstand long-term contact with the liquid.
• Lack of standardization: According to a 2022 Uptime Institute survey, insufficient standardization and a limited selection of suppliers were among the main barriers to the adoption of direct liquid cooling. This data is older, so it does not necessarily reflect the current situation.

Realistic Outlook

According to available analyses, immersion cooling is most commonly deployed in facilities designed from the outset for this solution, and in a single-phase configuration. According to multiple sources, on-chip cooling remains the primary path forward for high power densities. The pace of the transition to liquid cooling also depends on how quickly power density grows in conventional data centers, which, according to the Uptime Institute, remains relatively low today.

Conclusion

Liquid cooling is not a single technology but rather a range of approaches with varying trade-offs. Immersion cooling is attractive for extreme densities, but two-phase systems face the PFAS issue, and retrofitting is challenging. When reading reports about “the cooling of the future,” it is therefore worth asking: Is it immersion cooling, or on-chip heat spreaders? Single-phase or two-phase? And is it a pilot project or routine operation?