Immersion cooling in 2026 sits on a USD 520 million to USD 1.04 billion 2025 base across the two published taxonomies, expanding to USD 2.34 billion (phase-change/IT-equipment, 24.05% CAGR) or USD 4.88 billion (broader IT-equipment immersion, 24.75% CAGR) by 2032, with single-phase holding 62-70.7% of revenue and two-phase advancing at roughly 23-24% CAGR through the same window [S1][S4][S5].
Hyperscale AI training racks now operate at power densities that have collapsed the air-cooling envelope: NVIDIA GB200 NVL72 racks pull 120 kW versus the 15-20 kW air-cooling ceiling, and GPU clusters above 50 kW per rack are treated as the baseline procurement case rather than the exception [S2][S7]. That thermal shift is the single biggest reason immersion moved from a niche option to a 2026 build prerequisite.
Single-Phase vs Two-Phase: What the Segmentation Actually Says
Single-phase immersion holds 62-70.7% market share in the 2025 base across the PW Consulting and MarketsandMarkets taxonomies, favored for lower dielectric fluid cost, easier maintenance, and compatibility with standard server SKUs that do not require vapor-handling hardware [S1][S5]. Two-phase immersion, by contrast, is the fastest-growing technology segment at a 23.0% CAGR in one segmentation and absorbs the densest GPU/TPU racks because the latent heat of vaporization delivers heat-flux ceilings single-phase cannot match [S1][S3].
Open-bath immersion accounted for 65.2% of the 2025 phase-change-immersed market, and data center/cloud computing represented 69.9% of phase-change application share, confirming that hyperscale, not crypto-mining, is the dominant buying segment in 2026 [S4]. Two-phase achieves a Power Usage Effectiveness of 1.05 or less by exploiting natural evaporation and condensation, while single-phase D2C PUE typically sits at 1.2-1.3 and air cooling runs 1.4-1.6 [S2].
Fluid Chemistry and the PFAS Squeeze
PFAS phase-out rules in North America and Europe are pushing 2026-2032 fluid procurement toward PFAS-free hydrofluoroethers and alternative synthetic chemistries, with engineered fluorinated and hydrocarbon-based dielectric fluids dominating the qualified pool and legacy mineral-oil fluids still tolerated in some single-phase retrofits [S3][S4]. Chemours and 2CRSi completed Opteon two-phase fluid qualification in 2026, a concrete data point that the supply chain for low-GWP, PFAS-free two-phase dielectrics is no longer hypothetical [S4].
Buyers specifying dielectric fluids should track four concrete variables: dielectric breakdown voltage, kinematic viscosity at operating temperature, GWP, and fluid compatibility with connector polymers, because the 2025-2026 reformulation cycle is not just a regulatory checkbox but a chemical-stability gate that determines whether a tank survives 5+ years without fluid swap [S4]. The published segmentation logic maps these fluid variables directly to rack class: 50-100 kW AI training racks favor two-phase, mid-density HPC and edge favor single-phase [S3].
Vendor Map: Who Is Actually Shipping in 2026
LiquidStack, acquired by Trane Technologies in 2026, now offers single-phase and two-phase systems up to 252 kW per rack and is moving from product-only sales toward full data-center thermal management, a structural shift because channel access and end-to-end integration now matter as much as tank performance [S4][S5]. Asperitas (Netherlands) ships modular single-phase tanks built around natural convection for energy-efficient mid-density IT loads, and Iceotope (UK) focuses on chassis-level precision immersion that targets partial-immersion use cases rather than full-tank deployment [S5].
Submer Technologies, DCX, and LiquidCool Solutions round out the modular single- and two-phase shortlist that the 2026 PW Consulting market study names explicitly, and the 2026 vendor landscape is consolidating around factory-integrated stacks rather than component sales, because hyperscale buyers want shorter evaluation cycles and faster deployment [S5][S9]. Per industry projections referenced in [S4] and [S5], the immersion cooling market is growing at a 24.05% CAGR (2025–2032, per [S4]) and 24.75% CAGR (2025–2032, per [S5]).
Cost Stack: 120 kW Rack Economics and PUE
A 120 kW GB200 NVL72 rack on air cooling requires 48-72 kW of auxiliary cooling power, which the source quantifies as a 40-60% cooling overhead on IT load, and a 120 kW rack on air cooling can no longer be operated without thermal throttling that extends AI training times by 10-20% [S2]. Direct-to-Chip liquid cooling covers the 50-120 kW class and remains NVIDIA's reference method for GB200 NVL72, with over 95% of 2026 liquid-cooled racks estimated to use D2C rather than immersion [S2].
Two-phase immersion eliminates the CDUs and piping D2C requires, removes fans entirely, and holds pump power to a minimum because evaporation and condensation happen naturally at the heat source, with PUE of 1.05 or less confirmed in published 2026 guidance [S2]. Immersion also raises GPU MTBF by 20-30% over air cooling by suppressing oxidation, corrosion, and dust-induced failure mechanisms, a number that matters in any TCO model even if the upfront tank cost is higher than a D2C retrofit [S2].
Geography and Standards: Where the Money Sits in 2026
Asia Pacific held 36.9% of the broader data center cooling market revenue in 2025 and is the fastest-growing immersion region, with one segmentation pegging APAC at over 31% of the global immersion cooling market in 2026, roughly USD 213.12 million, driven by sovereign AI build-outs in China and Singapore and strict energy-efficiency policy [S3][S8]. North America led immersion revenue in 2025 at USD 231.9-420.9 million depending on taxonomy, and Europe trailed at USD 79.0-230.1 million [S4][S5].
Three standards bodies now govern the procurement conversation: UL Solutions' Immersion Cooling Program covers safety certification of systems and dielectric fluids for IT and server equipment, ASHRAE is publishing liquid cooling guidelines for high-density AI data centers in 2025-2026, and the Open Compute Project Immersion Sub-Project maintains open specifications and reference designs for both single- and two-phase immersion [S5]. The OCP-ASHRAE alliance formed in October 2025 to harmonize facility-readiness and interoperability requirements, which is the closest thing the industry has to a unified procurement spec in 2026 [S5].
Adjacent Stack and Selection Gate for Buyers
Immersion cooling is a server-room-scale play, not a utility-scale one, and the 2026 adjacent liquid-cooling taxonomy covers cold plates, CDUs, dielectric fluids, and tank-and-rack assemblies as separate procurement lines; the data center cooling 2026 vendor map lays out the broader stack where immersion sits, and the liquid cooling 2026 manufacturer tier map gives the cold-plate and D2C alternative that still ships over 95% of 2026 liquid-cooled racks [S2][S3].
Selection gate in 2026 reduces to four criteria: rack class (under 50 kW favors D2C, 50-100 kW favors two-phase immersion, above 100 kW with long idle duty cycles favors single-phase with heat-reuse), fluid regulatory status (PFAS-free hydrofluoroether for forward-looking 2030+ deployments, legacy fluorinated for 2026-2028 brownfield), warranty coverage from the server OEM (still patchy in 2026, a real adoption barrier), and total PUE including heat-reuse export to district heating networks where those networks exist [S4][S5].
Limits and Failure Modes Buyers Should Track
Two of the named 2026 barriers are upfront capex for specialized horizontal tanks and the lack of universal liquid-compatible warranty support from major server OEMs, and both slow retrofit adoption in legacy data centers even when the PUE math is favorable [S5]. Fluid management and ongoing maintenance protocols add operational complexity that single-phase simplifies but does not eliminate, and a poorly managed tank can degrade through fluid oxidation, polymer swelling, and non-condensable gas buildup in two-phase loops [S3][S4].
The 2017-2021 crypto-mining container pricing should not anchor 2026 HPC or AI procurement, because the high-fluidity dielectric fluids and higher-grade tanks used for AI training cost more than the open-air mining containers that dominated the earlier cycle, and the uptime expectations and build quality are not the same product class [S3]. The IEA's March 2026 report sized total data center capacity at 97 GW in 2024, projected to grow 2.3x to 225 GW by 2030, with most new capacity designed for liquid cooling from day one rather than retrofitted, which means the procurement decision in 2026 is really a 10-year infrastructure lock-in, not a one-year capex trade-off [S2].
Track the OCP-ASHRAE harmonized immersion spec and the 2CRSi-Chemours Opteon qualification expansion as the two nearest-term signals that determine whether a 2026 procurement file ages well or has to be re-specified by 2028, and confirm server OEM warranty language in writing before signing any single- or two-phase tank PO in the back half of 2026 [S4][S5].
Spec-level background on the components involved: pressure transmitter, flow meter, and industrial valve.