Immersion Cooling Hits $931M as AI Racks Exceed 100 kW
- Analysis Cooling & Heat Recovery Data Centers Efficiency & Optimization
Marin Holt
- 0
- 6 minutes read
- Immersion cooling market reached $931 million in 2026 revenue, projected to reach $4.9 billion by 2033 at 27.1% CAGR
- Technology delivers PUE of 1.03-1.10 versus industry average of 1.56, reducing cooling power overhead by 50-75%
- Single-phase systems hold 80.9% market share using synthetic hydrocarbon fluids at lower cost and complexity
- Two-phase fluorocarbon systems support rack densities exceeding 200 kW using fluids with 50-60C boiling points for frontier AI accelerators
- US market reached $454 million in 2025 with 16.9% CAGR, North America holding 44% global share
Immersion cooling reached $931 million in global revenue in 2026, up from negligible market share five years ago. The technology, which submerges servers directly in dielectric fluid, delivers PUE ratings between 1.03 and 1.10. That compares to the industry average of 1.56 reported by the Uptime Institute. For operators running AI workloads at rack densities above 100 kW, immersion cooling is no longer experimental. It is the baseline requirement.
Why Traditional Cooling Fails at Scale
Conventional air-cooled data centers handle rack densities of 17 to 30 kW. GPU-intensive AI training racks now exceed 100 kW, with next-generation chips pushing thermal design power past 1,000 watts per processor. Air cooling hits physical limits at these densities. Facilities that attempt it face exponentially rising fan energy, larger duct footprints, and PUE values stuck between 1.5 and 1.9.
Immersion cooling eliminates these constraints. Single-phase systems circulate dielectric fluid around components at ambient pressure. Two-phase systems boil the fluid at low temperature, capturing heat through phase change. Both approaches remove the need for raised floors, hot/cold aisle containment, and computer room air handlers.
Deployments Driving Adoption
Submer Technologies signed a memorandum of understanding with the government of Madhya Pradesh, India, in July 2025 to develop up to 1 GW of liquid-cooled AI data center capacity. The project targets two-phase immersion systems supporting rack densities above 100 kW.
In January 2026, Submer and Inspur deepened their collaboration to deploy two-phase immersion tanks for hyperscale facilities in China. The partnership focuses on frontier AI model training, where thermal loads peak during extended computation runs.
LiquidStack deployed a 40 MW immersion-cooled hyperscale facility in the Republic of Georgia. The project achieved 92.6% cooling energy savings and required 90% less floor space compared to air-cooled facilities of equivalent capacity. LiquidStack raised $20 million in a Series B extension in September 2024 to expand high-density AI cooling capacity.
KDDI in Japan recorded PUE values approaching 1.05 after deploying containerized single-phase immersion rigs, validating the technology for telecom edge applications.
The Economics of Conversion
Retrofit costs run $2 to $3 million per MW. A facility converting 75% of cooling from air to liquid reduces total power costs by 27% and overall energy consumption by 15.5%, according to BOYD Corporation data. For a 40 MW facility paying $0.08 per kWh, cooling energy savings of 40 to 50% translate to $1.4 to $1.75 million in annual operating cost reductions.
The immersion cooling market is projected to reach $4.9 billion by 2033, growing at 27.1% annually. Hyperscalers plan to invest $650 billion in AI infrastructure in 2026 alone. With GPU thermal envelopes expanding and power availability tightening, operators that delay the transition to liquid cooling face compounding efficiency penalties and stranded capacity.
Critical Perspective
Immersion cooling’s $931 million in 2026 revenue confirms the technology has crossed from experimental to operational, but the market projection assumes AI rack densities will continue growing beyond 100 kW, the same trajectory every hyperscaler was forecasting for air cooling in 2019 before workload consolidation flattened demand curves. The 1.03 PUE compares favorably against the 1.56 industry average, but PUE is a facility-level metric that excludes dielectric fluid procurement costs and the structural loading constraints of retrofitting existing data halls designed for 17 to 30 kW racks. The immersion cooling growth story is really two stories: greenfield AI campuses where the economics are compelling, and legacy retrofits where true implementation costs run 2 to 3 times higher. What the $931 million headline does not reveal: how much of that revenue is greenfield build versus retrofitted legacy facilities, because those are two entirely different investment theses.
