The immersion cooling market is projected to climb from USD 0.57 B in 2025 to USD 2.61 B by 2032, a 24.2% CAGR, driven by HPC, AI training racks, and cryptocurrency mining workloads that exceed air-cooling capacity [S2].
Smart-manufacturing adoption inside that growth is concrete: vendors are pairing dielectric tanks with PLC-controlled coolant loops, CRPS-form-factor immersion-rated PSUs, and PMBus telemetry so each pod becomes a node on the plant's pressure transmitter and flow meter network rather than a standalone IT appliance [S3][S6].
Two Cooling Architectures, One Automation Backbone
Single-phase immersion dominates hyperscale deployments because the dielectric fluid stays liquid, simplifying the heat-rejection loop and letting operators reuse existing facility-chilled water without phase-change pressure vessels [S2]. Two-phase immersion, where the fluid boils on the component surface and condenses on a coil, achieves higher heat-transfer coefficients per litre but demands pressure-rated tanks and tighter safety controls, which is why most new AI-factory builds still start with single-phase [S2].
Submer's SmartPod Exo specifies up to 361 kW of heat dissipation per pod, 50 RU of 19"/21" gear or 47 OU of OCP ORv3, and 2N redundant pumps plus 2N CDUs, with a Twin Tank variant engineered for 5x9s availability and concurrent maintainability [S3]. That pod-level redundancy maps cleanly onto ISA-95 functional layers: the CDU is a continuous process cell, the tank is a batch-style unit operation, and the facility BMS treats each pod as a smart meter-class energy asset publishing telemetry over SNMP or REST [S3].
Fluid Selection and the Dielectric Bill of Materials
Synthetic fluids (fluorocarbon-based, typically 3M Novec-class or PFPE chemistry) are expected to lead the immersion cooling market through the forecast window because of their dielectric strength, material compatibility with PCB laminates, and low long-term maintenance versus mineral oil [S2].
Mineral-oil variants are cheaper per litre and are still widely deployed in legacy cryptocurrency-mining sites, but they carry higher viscosity, lower flash point, and slower heat-transfer coefficients, which is why the same data-center fleet running oil today typically migrates to synthetic once the rack power density crosses roughly 50 kW per cabinet [S2]. Experimental work on mineral-oil flow rate over CPU cold plates confirms that heat-transfer coefficient scales with volumetric flow up to a turbulence threshold, reinforcing why CDUs in modern pods are sized with variable-speed pumps rather than fixed-orifice designs.
Power, Monitoring, and the CRPS Immersion PSU
FSP's immersion-rated PSU line ships in CRPS (Common Redundant Power Supply) form factor with 80 Plus Titanium efficiency, PMBus digital telemetry, and CB 62368-1 safety certification, and is explicitly designed for submersion in dielectric fluid [S6]. The CRPS mechanical envelope, 185 mm × 73.5 mm × 40 mm with a 12 V main output and standby rail, lets OEMs hot-swap supplies inside a sealed tank without breaking the dielectric bath, which is a hard requirement for any pod that claims concurrent maintainability [S3][S6].
PMBus output, in parallel with the pod's own SNMP/REST stack, gives the plant historian a per-rail voltage, current, inlet temperature, and fan-speed data set at 100 ms cadence, which is the same temporal resolution used in smart camera vision cells on the assembly line [S3][S6].
Control Loop Specs Inside an Immersion Pod
A pod's primary control loops are coolant inlet temperature (typically 35-45 °C for single-phase synthetic), tank level (compensated for thermal expansion of the dielectric), pump differential pressure, and CDU heat-exchanger secondary-loop supply temperature [S3]. Submer's spec sheet lists dry-zone operating fluid temperatures up to 60 °C (140 °F) at the heat-rejection interface, which lets facilities use ambient dry coolers in many climates and eliminates consumptive water use entirely [S3].
Instrumentation typically includes PT100 or PT1000 RTDs at tank inlet/outlet, a Coriolis or turbine flow meter on the primary loop, gauge and differential pressure transmitters across the pump and CDU strainer, and conductivity probes for dielectric moisture ingress, all wired back to a small PLC or industrial PC that publishes data into FactoryTalk or equivalent [S1][S3]. Process engineers should treat the dielectric bath as a Class I Division 2-equivalent zone even when the fluid is non-flammable, because hot-surface ignition of vapour films above 200 °C is a documented failure mode for two-phase systems.
Who the 2026 Stack Is Built For
Hyperscale AI training operators running 50-100 kW-per-rack GPUs, colocation providers servicing LLM inference fleets, and HPC centres supporting CFD or genomic workloads are the natural fit, because immersion cooling recovers the capex premium in 2-3 years through PUE reductions of 1.05-1.08 versus 1.4-1.5 for air-cooled halls [S2].
It is not the right answer for general enterprise server rooms, small edge sites under 10 kW per rack, or facilities that lack the floor-loading capacity for a fully flooded tank, which can reach 1,800-2,200 kg per square metre when filled [S2][S3]. For those builds, rear-door heat exchangers or direct-to-chip cold plates paired with a smart valve positioner-controlled facility loop remain the lower-risk route. Coverage of the broader automation stack, including PLC selection, OPC UA mapping, and historian tags for a 2026 immersion-cooled AI factory, sits in this site's data-center cooling smart-manufacturing roundup, and the tank-side fabrication workflow is detailed in the immersion cooling manufacturing process walkthrough.
Failure Modes, Maintenance Windows, and the Workforce Skill Gap
Documented immersion-pod failure modes include dielectric degradation from moisture ingress (typically measured as ppm-level water content and trended via online conductivity), pump seal leaks introducing air into the loop, and thermal-cycling fatigue on PCB laminates rated only to 130 °C glass-transition [S2][S3]. Submer's 2N pump and 2N CDU architecture is specifically sized to absorb the loss of one prime mover without tank-temperature rise exceeding the 5x9s availability target [S3].
Maintenance windows are governed by the fluid's flash point and the OEM's MTTR for pump and CDU swap-out, typically 30-60 minutes per pod, which is shorter than the 90-120 minute CRAC unit swap in a comparable air-cooled hall [S3]. The practical bottleneck in 2026 is not hardware but the workforce: technicians trained in both low-voltage IT and IEC 60079-equivalent hazardous-area practice are scarce, so operators rolling out more than 5 MW of immersion capacity per quarter should budget for vendor-led training and FactoryTalk-style digital-twin rehearsals before live commissioning [S1].
Sourcing Signals Worth Tracking Through 2026
Three verifiable signals are worth watching into late 2026: GRC's expansion of single-phase immersion into Middle East and Africa hyperscale sites in partnership with Dell Technologies and DCV Industries, originally announced October 2023 and still rolling out regional capacity [S2]; OEM migration from mineral oil to synthetic PFPE fluids on any pod above 50 kW per rack, driven by thermal-performance and hardware-reliability data [S2]; and standardisation of immersion-rated CRPS PSUs with PMBus and 80 Plus Titanium, already a shipping spec at FSP and likely to be adopted as a de facto procurement baseline by hyperscale buyers [S6].