Submerge a server in mineral oil. The oil absorbs heat from every component on the board. A pump circulates the heated oil out of the tank, through a heat exchanger, and back in. The oil never boils. Never changes state. Never leaves the loop as vapor. That is single-phase immersion cooling in its entirety. The engineering is almost aggressively simple. And right now, simplicity is the single most valuable attribute a cooling technology can have.
Single-phase immersion held 80.9% of the data center immersion cooling market in 2024. That number is climbing. The reason has less to do with thermal performance and more to do with a chemistry crisis that kneecapped the competition.
Servers go into an open tank. The tank is filled with a dielectric fluid, something electrically nonconductive that can absorb and transfer heat. Options range from mineral oil at $19 to $38 per gallon to synthetic hydrocarbons and natural esters like Cargill's NatureCool, a vegetable-oil-based fluid that biodegrades in about 10 days and carries 10% higher heat capacity than most synthetics. Engineered fluids from the likes of Shell and Castrol run $150 to $400 per gallon, depending on chemistry and volume.
Heat transfers from CPU, GPU, memory, VRMs, and every other component directly into the surrounding fluid. No fans. No air plenums. No hot aisle, cold aisle. The fluid circulates through a heat exchanger that rejects the thermal load to a facility water loop or dry cooler, then returns to the tank. Inlet fluid temps typically run 15 to 35 degrees Celsius. Some newer systems push that to 45C.
PUE lands between 1.02 and 1.10 in production deployments. For context, conventional air-cooled data centers average 1.58 to 1.65. The energy savings come from eliminating CRAC and CRAH units entirely. No compressor energy. No fan energy. Water consumption drops 45% compared to air-cooled equivalents. Some deployments hit zero water use when paired with dry coolers.
Submer in Barcelona raised $55.5 million in October 2024 and used the capital to launch two new business units for data center design and operations. They signed an MoU with the government of Madhya Pradesh for up to 1 GW of liquid-cooled AI data centers in July 2025. A deployment deal with Stellium Datacenters in the UK followed in September. Their fluid is a synthetic hydrocarbon co-engineered with Castrol. PFAS-free. Biodegradable. Submer's CEO, Daniel Pope, has been positioning the company as the turnkey operator for immersion-cooled AI factories, not just a tank vendor.
GRC (Green Revolution Cooling) in Austin pulled Samsung Ventures in June 2025, followed by a three-way MOU with LG Electronics and SK Enmove in October. They launched the ReliaSys IR500 CDU in November: 500 kW cooling capacity, purpose-built for AI density. Claims deployments in 22 countries. GRC's Jim Weynand has been the loudest voice in the room about immersion displacing direct-to-chip at scale.
Asperitas in Amsterdam closed a scale-up round in September 2025 led by STECON GROUP GLOBAL, with Shell Ventures participating. They partnered with Shell to develop a gas-to-liquid immersion fluid and signed a commercial agreement with UNICOM Engineering in February 2026 for single-contract immersion-cooled compute procurement. One purchase order. Server, tank, fluid, cooling loop. That is the pitch.
Single-phase immersion was growing before the PFAS crisis. After it, the growth became structural.
Two-phase immersion cooling, the technology that can push PUE below 1.03 and handle heat fluxes of 1,500 W/cm2, relied almost entirely on fluorinated fluids from 3M. Novec 7100. Novec 649. Fluorinert FC-72. All PFAS-based. All discontinued. 3M exited all PFAS manufacturing by end of 2025 under the weight of $12.5 billion in litigation. The EPA designated PFOA and PFOS as hazardous substances under Superfund. The EU is working through a restriction proposal covering over 10,000 PFAS substances, with ECHA final opinions expected by the end of 2026.
Single-phase fluids sit entirely outside this regulatory blast radius. Mineral oils, synthetic hydrocarbons, natural esters, gas-to-liquid fluids. None of them contain PFAS. None of them carry Superfund liability. None of them face supply chain concentration risk. Cargill, Shell, Castrol, and dozens of regional suppliers can all deliver. The procurement team never has to call one company in Zwijndrecht, Belgium and hope the inventory lasts.
Single-phase immersion has a thermal ceiling, and the GPU roadmap is heading straight toward it.
Conventional single-phase immersion handles sustained heat flux of roughly 250 W/cm2 at the chip surface. Enhanced surfaces and forced convection within the tank can push that higher, but the physics of convective heat transfer without a phase change set a hard boundary. Two-phase systems, by contrast, leverage the latent heat of vaporization and can reach 1,500 W/cm2. That gap matters when NVIDIA's Rubin architecture starts shipping at power draws that make Blackwell look modest.
Maintenance is the other real constraint. Pulling a server out of a tank of dielectric fluid, letting it drain, servicing a DIMM or GPU, and resubmerging it takes longer than swapping a cold plate. The fluid gets on everything. Material compatibility testing never ends. Seals, gaskets, thermal interface materials, cable insulation: each one needs validation against the specific fluid chemistry. And the tanks are heavy. A facility designed for standard rack loading may not support the structural requirements of fluid-filled immersion baths without floor reinforcement.
Then there is the OEM question. Dell, HPE, and Lenovo all offer factory-integrated direct-to-chip cold plate options. You can order a liquid-cooled server through the same procurement channel you use for everything else. Immersion-cooled servers require either purpose-built hardware or heavy modification. The ecosystem is growing. Iceotope cooled chips to 1,000W using single-phase precision liquid cooling in February 2024. But the server catalog for immersion remains a fraction of what direct-to-chip can accommodate.
Single-phase immersion cooling will not be the technology that handles the absolute thermal extremes of the next GPU generation. That crown probably belongs to direct-to-chip cold plates at scale, and possibly to two-phase systems if anyone solves the fluid problem. But single-phase immersion will be the technology that captures the widest range of use cases where air cooling has already failed and operators need a solution they can actually procure, install, and maintain without a chemistry degree or a Superfund lawyer.
Greenfield AI factories in India and the Middle East. Edge deployments in places with no raised floor and no water infrastructure. Sovereign compute in countries that cannot afford to depend on a single American or European fluid vendor. The market is projected at $11.1 billion by 2030. Single-phase will take the majority of it.
The technology that wins at scale is rarely the one with the best spec sheet. It is the one with the widest supply chain, the lowest operational risk, and the fewest ways to go wrong on a Tuesday night when the on-call technician has never seen dielectric fluid before. Single-phase immersion is that technology. For now, that is enough.