Immersion cooling submerges IT hardware in a dielectric fluid, with rack densities in 2024 surveys already exceeding 30 kW at 17% of surveyed sites and AI workloads now pushing designs past 600 kW per rack [S2][S4].
The supply chain spans dielectric fluid blenders, brazed-plate heat exchanger OEMs, CDU integrators, and tank fabricators upstream; data centers, crypto miners, EV battery lines, and industrial process plants sit downstream of those vendors [S1][S5].
Where immersion cooling sits in the thermal management mix
Air cooling caps out around 20-30 kW per rack and fails above a chip TDP of 250-280 W, which is the threshold where forced-air heat transfer cannot hold junction temperature in spec [S2]. Direct liquid cooling splits into indirect (cold plates, heat pipes, thermosyphons) and direct (immersion, jet impingement, spray); immersion itself divides into single-phase and two-phase configurations, with the latter exploiting latent heat of vaporisation for higher heat-flux removal [S2][S3].
The dielectric fluid is the working medium, not water: mineral oil hydrocarbons, synthetic fluorocarbons, and silicone fluids are the three commercial families, and the fluid choice dictates whether the system runs single-phase or two-phase [S3]. dielectric cooling fluids anchor the upstream chemistry side of the chain, while process-side flow measurement governs how the secondary coolant loop is metered and balanced across the CDU and chiller interface.
Upstream: dielectric fluids, heat exchangers, and CDU components
Shell markets GTL-derived process oils specifically blended for immersion cooling duty in data center and blockchain applications, indicating a dedicated fluid grade rather than repurposed transformer oil [S6]. Two-phase systems use a low-boiling-point dielectric that flashes to vapour on the chip surface and condenses on a coil above the tank; the vapour-side heat transfer coefficient is materially higher than the liquid-side coefficient, which is why two-phase is reserved for the highest TDP devices [S3][S5].
The CDU is the integration node: each unit delivers 200-240 kW of cooling capacity and supports one to four immersion tanks in modular deployments, with a brazed-plate heat exchanger separating the dielectric loop from the building chilled-water primary [S4]. Alfa Laval and Vertiv both sell immersion tank skids as OEM products, with the tank itself treated as a 19-inch rack-equivalent enclosure holding servers vertically and instrumented with leak sensors and PDU feeds [S4][S5]. industrial valve selection matters here: flow-modulating valves on the CDU secondary loop set tank return temperature, and the wrong valve characteristic curve will hunt under variable AI load.
Downstream: data centers, crypto mining, EV, and industrial plants
Data centers are the anchor end-market, with hyperscalers such as Microsoft and Intel publicly named as operators alongside military and government installations [S1]. Rack power density has climbed 3.5x from 2010 to 2020, and a 2024 Uptime Institute survey of 721 facilities found 17% running maximum rack power above 30 kW, a band that air cooling cannot serve economically [S2]. Vertiv targets 600 kW per rack and beyond for AI training clusters, with each immersion tank absorbing 100% of rack heat while the room itself still needs air for humidity and particulate control [S4].
Cryptocurrency mining is the second-mover end market: Riot Blockchain operates 200 MW of immersion-cooled hash capacity supporting 46,000 mining devices, and Bitfury has installed multi-hundred-MW immersion systems in hyperscale mining sites [S1]. Electric vehicle battery lines and industrial manufacturing plants are identified as the next adoption wave, where immersion is sold on the same premise as in IT: higher power density, dust and corrosion exclusion, and heat reuse for adjacent processes [S1]. data center ceiling containment is a downstream spec that still has to be solved even when the rack is liquid-cooled, because the room air path is not eliminated.
Spec comparison: single-phase vs two-phase immersion
Single-phase immersive IT chassis and tub variants circulate dielectric without phase change, handle moderate-to-high workloads, and simplify integration into existing rack footprints; two-phase immersive tubs vaporise the dielectric on the chip, condense it on a coil, and are the only option for extreme AI and HPC heat flux [S5]. The trade-off is fluid cost and vapour management complexity against peak heat flux removal, with two-phase generally reserved for chips at or above the 300-1000 W per-device band that modern GPUs occupy [S2].
Across the three commercial fluid families, mineral oils sit in the lower-cost, higher-viscosity band suited to single-phase tubs; synthetic fluorocarbons span single and two-phase grades with tighter dielectric specs; silicone fluids target high-temperature stability and longer fluid life in sealed chassis [S3][S6]. CDU capacity (200-240 kW per unit), tank count per CDU (1-4), and primary-loop connection (chiller, dry cooler, or waste-heat recovery) are the three scaling knobs that drive total deployed cost on the data center side [S4].
Selection criteria, limits, and failure modes
Immersion cooling is not a drop-in for every site: retrofit requires reworked space layouts, upsized power feeds, and UPS-backed CDUs because pump failure is now a thermal event, not just a mechanical one [S4]. The dielectric fluid must remain chemically stable across thousands of thermal cycles, and any leak into the room air path defeats the sealed-tank premise; leak detection and dielectric-compatible seals are mandatory in tank design [S4][S5].
Because data centers account for 1-1.5% of global electricity consumption and over 40% of data center energy is consumed by thermal management systems, fleet-wide adoption of more efficient cooling methods can yield significant aggregate grid savings. Room cooling is not eliminated, even when rack heat is fully captured: air quality, humidity, and condensation control on tank exteriors still demand a residual CRAC or air-handler envelope around the immersion hall [S4]. Operators specifying immersion for AI workloads should validate CDU redundancy, dielectric fluid MSDS and reclamation logistics, and primary-loop heat-reuse economics before locking the tank count per CDU [S4][S5].
Standards, sourcing signals, and the next 6 months
No single IEC or ASHRAE standard yet pins immersion cooling, but emerging vendor consortia and data center operator working groups are publishing tank, fluid, and CDU interface specs as the install base grows [S1]. Shell's GTL-based immersion fluid line and Alfa Laval's and Vertiv's immersion tank skids are the most concrete OEM signals that the upstream supply chain has hardened beyond pilot scale [S4][S5][S6].
Track two signals through Q1 2027: the first hyperscaler AI training hall running 600 kW-class immersion racks in production, and the first published dielectric fluid reclamation/recycling standard from a major data center operator group. A third trackable node is whether EV battery cell formation lines begin quoting immersion-cooled buffer tanks, since the duty profile (high power density, dusty ambient, heat-reuse opportunity) maps almost directly onto the data center use case that is already shipping at scale [S1][S4].