Immersion Cooling (Single-Phase & Two-Phase)

Judged on heat capture alone, immersion looks like the cooling that ought to have won. Heat capture is not the only score, because the rack’s service boundary moves into a fluid bath. Submerge the whole board in dielectric fluid and you remove every air gap, every fan and the external cold-plate interface, though the package and component-to-fluid resistance remain. The thermodynamics are excellent: high heat capture, halls quieted by removing server fans, and a concentrated warm fluid stream that suits heat reuse, if the fluid and exchanger can carry the duty. Pumps and rejection fans still use power and make noise. For a decade, cooling roadmaps drew immersion as the inevitable endpoint past the air-cooling cliff. As of 2026 it has not arrived, blocked not by thermodynamics but by chemistry, serviceability, supply chain, and insurability. The frontier racks of 2025-2026 are cooled by direct-to-chip liquid, not by tanks. → Chapter 5.4.

This chapter weighs a technology you may correctly choose not to deploy. Two architectures get lumped together. In single-phase immersion the fluid stays liquid and moves past the boards. In two-phase immersion it boils on the hot components and condenses on a coil. Their 2026 maturity differs sharply. Single-phase is a viable, supply-secure niche. Two-phase is still constrained by qualification: 3M exited PFAS manufacturing, and Chemours’ Opteon fluid has been qualified in current-generation 2CRSi servers. Two forks follow: immersion vs direct-to-chip at the architecture level, and single- vs two-phase within immersion. Each carries downstream costs in capex, serviceability MTTR, floor loading, PFAS liability, and the insurance and fire-suppression posture that decides whether a lender or insurer will let you build at all. → Chapter 2.6, Chapter 6.5.

Both architectures put servers into a dielectric fluid that does not conduct electricity, so live boards sit safely submerged. The servers are de-fanned, fitted with immersion-grade thermal-interface compound, and free of unsealed spinning disks: flash, or sealed helium drives where the drive, fluid and server are qualified together. The difference is what the fluid does with the heat.

Single-phase immersion keeps the coolant liquid throughout. A synthetic hydrocarbon or fluorinated oil carries heat by natural or forced convection to a liquid-to-liquid heat exchanger, which hands it to the facility water loop. The tank drawn below uses forced circulation, with a pump moving the fluid through an external exchanger; some designs use no pump at all. Thermal duty is set by the fluid, the component surface, the flow mechanism, the temperature limits and the exchanger capacity: strong, but not unlimited. A single-phase bath tolerates an open lid and routine service when its qualified fill level, temperature, fire classification and handling procedure allow it. A low fill that uncovers a hot component is not forgiving. Operationally it behaves like a very wet rack: contained, manageable, serviceable with gloves and a drip tray.

Two-phase immersion uses the latent heat of vaporization. The fluid is engineered to boil at a low temperature (typically ~49–60 °C) right at the hot components. The vapor carries the heat upward, condenses on a chilled coil at the top of a vapor-managed bath, and rains back down. Boiling moves heat far more aggressively than convection, so two-phase can buy heat-flux headroom with less circulation, provided the boiling surface, condenser and vapor system carry the duty. Count the total auxiliary power of each design before declaring an energy winner. Two-phase also demands a vapor-management enclosure, a fluid with a tightly tuned boiling point and tight vapor control, and until 2025 that fluid came almost exclusively from one supplier.

Bath physics sets the density limit. Immersion oils such as the coolant Vertiv specifies for CoolCenter (808 kg/m³ at 15 °C, flash point 198 °C) store roughly half the heat per kilogram that water does and conduct under a quarter as well. A GPU in an oil bath therefore keeps a finned heat sink, and the pump moves more fluid than a water loop would for the same duty.

A two-phase fluid carries its heat as vapor instead: Chemours’ October 2026 datasheet for Faizeon 500, the renamed Opteon 2P50, gives a 48.9 °C boiling point and 115 kJ/kg of latent heat. Boiling has a ceiling, the critical heat flux (CHF), past which a vapor film blankets the surface and the component temperature runs away. Zuber’s correlation with that datasheet’s properties puts CHF on a plain flat surface at about 15–20 W/cm². A Blackwell-class GPU pushes 1.2–1.4 kW through about 16 cm² of compute die. Dividing one by the other gives 75–90 W/cm², a scaling estimate rather than a measured flux (Chapter 5.1 explains the method). Spread over a 50 cm² boiler plate, the same power still comes to 24–28 W/cm². A two-phase bath for current GPUs therefore depends on an enhanced boiling surface, such as a porous coating or a finned boiler plate. Get that surface’s tested CHF, with your fluid at your operating pressure, in writing.

Chemours’ 24 February 2026 announcement reports qualification of Opteon 2P50 in 2CRSi’s Atlantis and Octopus servers, followed by a joint development agreement. That covers those named fluid-and-server combinations. The agreement’s wider direct-to-chip development goal does not certify any other cold plate, fluid formulation or delivered product. Chemours’ two 2025 milestones are smaller steps. In August 2025 Samsung Electronics approved the Opteon two-phase fluid for its fourth-generation SSDs, one component rather than a server. In May 2025 Chemours contracted Navin Fluorine to manufacture the fluid, establishing its first commercial-scale capacity from 2026. Neither gives you a second source, because the fluid remains Chemours’ proprietary product. The fluid now ships as Faizeon 500, so make sure the qualification letter and the purchase order refer to the same product.

The table can invert the old cooling scorecard. A bath that wins on heat capture can still lose once you price capex, serviceability, fluid supply, OEM warranty and insurability. Removing server fans does not remove pumps, condensers, fluid handling or GPUs idle during service. Direct-to-chip can win through tray access and common spares. Immersion can win when its heat capture, environmental isolation and service design repay the tank and fluid costs. The obstacle is often the operational and institutional friction around the cooling. Compare on one operating brief: an unmatched cooling-system PUE or an equipment-only price cannot settle it. GRC’s brochure, for example, prints 1.01–1.02 for two-phase and 1.02–1.03 for single-phase without saying what annual boundary either covers.

The most underestimated cost of immersion is the labor and downtime of touching the hardware, not the fluid. A direct-to-chip rack services like a conventional rack: pull a tray, the dripless quick-disconnects seal, swap the part, reinsert. Keeping an immersion tank powered does not make it hot-swap. What matters is which server you can isolate and lift while the rest of the bath stays in service. To replace a failed DIMM, NIC or (most commonly) a GPU, a technician lifts the server out of the bath, lets the fluid run off (single-phase) or manages vapor escape (two-phase), works on the board over a containment tray, and re-immerses it. Vertiv’s CoolCenter manual, for one, isolates and lifts a single server onto service rails without draining the tank, and allows maintenance right after runoff with no fixed wait. Other tank procedures may require draining, drying, refilling and reconditioning. Before you plan an SLA around the difference, time diagnosis, isolation, extraction, fluid recovery, checks and workload restoration on your own tank. Give the air-cooled comparison the same failed assembly, spare and crew. A task estimate is not a fleet MTTR, because the procedure decides whether one server or the whole bath goes out of service.

That tax compounds with the failure rate of the fleet being cooled. A GB200-class rack holds 72 accelerators each dissipating ~1 kW+, and at frontier scale GPU and optics failures are a daily fleet event, not an annual one. Every failure takes hardware capacity out of service for the length of the procedure. Whether it also interrupts the job depends on the training stack. A failed active node can stall a synchronous job or force a restart, but spare substitution or runtime fault recovery can resume training before the physical repair ends. Price the unavailable capacity and the job-level interruption separately (Chapter 9.4, Chapter 14.1). For a power-bound operator paying for every megawatt, idle accelerators in a tank are still the worst kind of stranded capacity, and the cooling that minimizes PUE can maximize the time a failed node spends out of service. → Chapter 5.4 on the serviceability advantage of cold plates.

Server-level spray cooling is a third option for fluid and service design. On 5 October 2026 Airsys and Shell announced a cassette-based dielectric spray system. The companies scheduled a live demonstration on a running server at the OCP Global Summit on 14 October 2026; as of 19 October neither had published measured results. Airsys claims it uses up to 80% less dielectric fluid than immersion. It does not say which immersion system that figure compares against, and it gives no production thermal rating. If you consider it, compare its fluid inventory and service boundary with your bath, and get hardware compatibility and pump and nozzle failure behavior in writing first.

Immersion inverts the floor-loading conversation. Direct-to-chip adds manifolds, tray loop assemblies and the in-rack fluid charge to an already dense rack, and it is the dense rack itself that forces the structural look (Chapter 5.10). Use the supplier’s loading record for the rack you buy; NVIDIA, for example, reports roughly 4,000 lb for the Vera Rubin NVL72 rack (March 2026). A flooded immersion tank is a different object. Vertiv’s CoolCenter immersion tanks have fully loaded masses of 1,732–2,095 kg, and the CDU and filtration skid adds a separate load. Vertiv’s table gives 1,721 kg/m² for the 42U tank and 1,679 kg/m² for the 52U tank, assuming 15 kg of server per U. Those are footprint-average screening pressures without support frames, not leg or frame reactions, so they cannot say whether a slab needs reinforcing. The manual gives separate support-frame configurations and foot reactions for that. Check the actual reactions, the move and installation states, and the load combinations against your floor system at scoping, not at install (Chapter 6.2).

Whether a tank fits an existing hall is therefore a project-specific question. Multi-ton equipment masses, low horizontal tanks that consume floor area differently from vertical racks, fluid storage and handling space, and suppression and containment requirements all compete for room in a brownfield hall designed for vertical air-cooled cabinets. The retrofit path past the cooling cliff often favors rear-door heat exchangers or direct-to-chip when a tank would need a new support frame, move route, handling area or fluid store. Keep the tank in the comparison where its actual reactions and service space fit, and price the enabling works where they do not. → Chapter 5.4.

Immersion is niche, not dead, and there are real workloads where it remains the right answer. Heat reuse is its strongest case, and the advantage is capture, not temperature. Because the whole board sits in the fluid, every watt leaves in one liquid stream that can feed a heat pump and a district-heating offtake, rather than splitting between a liquid loop and a residual air loop the way a DLC+RDHx rack does. That stream is not necessarily warmer. Oil convects heat poorly, so a single-phase bath keeps its bulk fluid cooler to hold component temperatures. Vertiv’s CoolCenter takes facility water into its heat exchanger at 12–35 °C. One loop further in, NVIDIA allows the secondary coolant reaching Rubin cold plates to run as warm as 45 °C. Compare the kWh delivered at the temperature the heat buyer needs, not the architecture label. Where a district-heating network or an industrial offtake can take the heat, immersion’s full capture can flip its economics. → Chapter 5.9 (engineering) and Chapter 15.5 (economics/district heating).

The other durable niches are extreme-density and harsh-environment deployments. Cryptocurrency mining fits where density per dollar still dominates after you count the actual service cost and the production lost during repair. Space- or dust-constrained edge sites favor a sealed tank over filtering and conditioning air. Specialized HPC belongs here where it needs thermal headroom past direct-to-chip today. Whether immersion offers useful headroom depends on component heat flux, allowable component and coolant temperatures, and demonstrated cold-plate duty, not on rack kW alone. NVIDIA’s June 2026 description of Rubin, for example, already has its cold plates running 75% water and 25% propylene glycol at up to 45 °C. If component duty outruns what cold plates demonstrate, immersion’s latent-heat headroom may re-enter the mainstream conversation, but only with a supply-secure, insurable two-phase fluid on a boiling surface rated past the GPU’s heat flux, or a sufficiently aggressive single-phase design. The bet immersion is making is that the density ramp eventually outruns cold plates. → Chapter 5.1 (the density wall) and Chapter 16.2 (subsystem roadmap).

A constraint that rarely appears in cooling comparisons but routinely decides them is whether the facility is insurable and code-compliant. To a fire authority and an insurer, a large tank of combustible or flammable dielectric fluid is a fuel load and an unusual suppression problem. Dielectric describes electrical behavior, not combustibility, so classify the fluid you are buying as described above. Some formulations are nonflammable: Chemours classes Opteon 2P50 as a nonflammable liquid. Even then, the system’s enclosure, electrical design, fluid handling and decomposition products still need review. Conventional rack-based suppression (clean-agent or pre-action sprinkler) does not map cleanly onto a tank farm, so immersion changes the fire model, the containment requirement and the spill-response plan. Two-phase requires a vapor-management enclosure. Whether that enclosure counts as a code pressure vessel depends on the specific system’s MAWP, construction, fluid and local code. PFAS handling is a separate diligence item for the fluid.

Immersion can therefore fail a project on diligence rather than engineering. An insurer (FM Global-class) or lender may decline to underwrite a novel cooling architecture with thin actuarial history, or the authority having jurisdiction may not sign off on the suppression scheme. Working technology does not settle risk transfer or life-safety approval; each is a separate gate. Engage the insurer and the AHJ before committing white space, not after. → Chapter 6.5 (fire detection, suppression and life-safety) and Chapter 2.6 (insurance and risk transfer).

The cold-plate allowance is 20 + 25 + 25 = 70 minutes, and immersion is 20 + 30 + 25 + 35 = 110 minutes. Both fit the 180-minute ceiling because the case grants an on-site qualified spare and a trained crew. Both stay in the running for the matched thermal and lifecycle-cost comparison. The shorter procedure alone says nothing about the cost of recovering a distributed job.

The immersion design has 70 minutes of allowance left. A missing qualified spare that adds 90 minutes produces 200 minutes and fails by 20. Stock the spare or change the service design. Get a timed demonstration and warranty confirmation before treating either allowance as an accepted restoration time. OCP qualification and service interfaces give background. Chapter 14.5 covers the operating procedure, and Chapter 13.5 the cooling acceptance handoff.

Choose immersion when its qualified hardware and fluid, its flooded support reactions and a demonstrated service process meet the same duty and restoration requirement as the alternative. A lower auxiliary load is worth nothing operationally if the supported replacement workflow misses the required restoration window.