Immersion cooling production lines for AI and HPC data centers are being engineered around future rack loads of up to 500 kW, with dielectric tanks sized to a 19-inch rack standard and a coolant distribution unit (CDU) bridging the dielectric loop to a chiller or dry cooler [S2]. Evonik positions high-performance dielectric fluids as the working medium inside these tanks, citing heat reuse into building loops as a sustainability lever for power-intensive sites [S1].
The line is laid out as a sealed primary loop (tank, CDU, filter, redundant pumps) decoupled from a secondary building loop through a brazed plate heat exchanger, so dielectric fluid never mixes with facility chilled water [S2]. For foundries and OEM planners reading this as a process analogy, the same loop discipline shows up in molding line heat-transfer skids and in automatic molding line coolant panels, where primary/secondary separation is the reliability gate.
Tank geometry, 19-inch rack standard, and component prep
Tanks are sealed vessels that house IT equipment submerged in dielectric liquid, with hardware inserted from the top and positioned vertically; fans and power supplies must be removed from servers before submersion, and each tank carries a power distribution unit (PDU) plus sensors for level, temperature, and pressure [S2][S4]. Vertiv's CoolCenter Immersion reference places the rack hardware horizontally in a tank that follows a typical 19-inch standard, so the line can be commissioned in steps without disturbing adjacent aisles [S2].
Submersion absorbs heat directly at the component, then the dielectric is pumped to the CDU; chilled water on the secondary side never contacts the electronics, which is the core safety argument versus spray or direct-to-chip designs [S2][S4]. The dielectric fluids used today are typically fluorinated oils and engineered fluids such as 3M's legacy Novec and Fluorinert product lines, chosen for their non-conductive behavior and high dielectric strength [S9]. For broader plant planning, this same closed-loop philosophy is detailed on the line frequency furnace reference, where inductive heating loops are isolated from mains water in an analogous way.
CDU, brazed plate heat exchanger, and pump redundancy
The CDU is the process heart of the immersion line: redundant pumps circulate dielectric from the tank through a brazed plate heat exchanger, then through a filter and back to the tank, while a flow modulating valve trims cooling capacity to the live heat load [S2]. Heat transfer happens across the plate pack, so dielectric pressure drop, plate material, and gasket compatibility with the chosen fluid are the first specs a process engineer should freeze on a P&ID before any pumps are ordered [S2].
Vertiv's reference design uses a secondary loop connected to a chiller, dry cooler, or waste heat recovery system, which lets operators in cold-climate sites dump heat to a dry cooler for part of the year and recover it for district or building heating for the rest [S1][S2]. The market data behind this layout is concrete: Dell'Oro Group reported worldwide data center liquid cooling revenue of $745 M in 2023 (up 52% year over year) and projected a 45% CAGR to over $4.8 B by 2028, with liquid cooling reaching 36% of total data center thermal management spend [S2].
Single-phase vs two-phase vs direct-to-chip: criteria comparison
Single-phase immersion keeps the dielectric below its boiling point and circulates it as a liquid, while two-phase immersion allows local boiling on hot components, with vapor condensing on a coil or tank lid for higher heat-transfer coefficients at the cost of fluid management complexity [S7]. Direct-to-chip (D2C) cooling, by contrast, targets cold plates on the highest-TDP parts (CPUs, GPUs, memory) and leaves the rest of the server to air, which is why Nvidia designed the Blackwell B200 and GB200 superchip specifically for direct liquid cooling paths rather than full submersion [S3][S7].
The decision matrix lines up clearly: single-phase immersion scores high on simplicity and fluid compatibility but caps heat flux at the dielectric's single-phase limit; two-phase immersion wins on peak kW per liter but adds vapor handling, condensate return, and stricter pressure-vessel review; direct-to-chip preserves the existing air-cooled chassis and is the lowest-risk retrofit for an automatic molding line-style brownfield upgrade where only the densest nodes are reworked [S3][S7]. The TDP jump frames the stakes: a typical server GPU sat at 150 to 300 W only a few years ago, while a single Nvidia Blackwell B200 Tensor Core reaches up to 1,200 W, and GB200 systems pair those accelerators, which is what drives the move away from air cooling toward any of these three liquid paths [S3].
Fluid selection, sustainability, and waste-heat reuse
Dielectric fluid choice defines both the safety envelope and the sustainability story: Evonik's immersion-cooling portfolio is explicitly built for data centers, HPC, and power-intensive applications such as energy storage, with the supplier highlighting that captured heat can be fed into building heating circuits to cut total energy and carbon footprint [S1]. Submer, a system integrator in the space, and Alfa Laval both stress the same heat-reuse argument: a dielectric bath enables direct heat transfer at the chip, and a downstream heat exchanger upgrades that low-grade thermal energy into usable hot water or process heat [S6][S7].
PUE is the metric operators watch: air cooling at high densities can push facility PUE above 1.4 and consume up to about 40% of total data center energy, whereas liquid and immersion paths are designed to bring PUE down and free that energy budget for IT load [S3]. The same heat-recovery logic shows up on the resin sand line reference, where exhaust heat from curing is captured to preheat resin, so the engineering instinct to reuse low-grade heat is not new to industrial process designers.
Failure modes, limitations, and who should NOT specify immersion
Immersion is not a universal answer: a fully populated air-cooled hall running legacy servers below 15 kW per rack is rarely worth a tank retrofit, and sites with strict structural floor-load limits must recheck slab ratings once the dielectric is filled, since a single 19-inch tank of fluid adds substantial dead load. Submer's own comparison and Supermicro's glossary both flag that fans and power supplies must be removed or re-engineered before submersion, which means the bill of materials for a "drop-in" immersion upgrade is never truly drop-in [S4][S7].
Two-phase designs in particular introduce pressure-vessel and flammability review overhead that single-phase dielectric baths avoid, and the legacy 3M Novec/Fluorinert product lines are being phased out by their original manufacturer, which complicates long-term fluid sourcing and forces new tank designs to qualify alternative chemistries [S9]. Operators also need to plan for dielectric maintenance: filters on the CDU loop, periodic fluid sampling for breakdown products, and tank-side leak detection are operational realities, not optional extras, and they share a maintenance culture with conveyor sorting line condition monitoring, where continuous sensor feedback is what keeps the line from drifting into a failure state [S2][S7].
Sourcing, standards, and verifiable signals to track next
Public standards activity on immersion cooling is still maturing, so most operator specifications are anchored in OEM reference designs, ASHRAE TC 9.9 liquid-cooling guidance, and supplier datasheets rather than a single binding immersion-specific code; teams should freeze their own internal spec on dielectric dielectric strength, kinematic viscosity at operating temperature, flash point, and material compatibility with tank seals and gaskets before signing a PO [S1][S4]. For comparable spec-first procurement logic on adjacent hardware, see the Industrial Refrigeration Industry 4.0 Adoption spec map for 2026 and the EV Traction Motor Manufacturing Equipment process map, both of which apply the same primary/secondary loop and component-level gate review to a different process.
Trackable signals to watch over the next two quarters: (1) OEM releases of 800 VDC and 1,200 W-class accelerator reference designs that confirm direct liquid cooling as the default path, putting more pressure on immersion to win on density per square meter; (2) supplier announcements of next-generation dielectric fluids with higher flash points and lower GWP as 3M-era chemistries exit the market; and (3) any operator disclosure of full-tank PUE below 1.1 at sustained 100 kW per rack or higher, which would be the first hard field data point that single-phase immersion can carry mainstream AI loads without a D2C fallback [S3][S7][S9].