Immersion cooling remains behind direct-to-chip because a retrofit requires dedicated tanks, dielectric fluid, pumps, monitoring, redesigned cabling, and trained maintenance personnel, despite removing heat from entire servers [S2].
The practical market divides by deployment model: cold-plate systems can be introduced one cabinet or row at a time, while full immersion changes the rack, service process, and facility architecture [S2][S3].
Heat Loads Make Liquid Cooling Necessary
NVIDIA GPU power rose from 300 W for the V100 in 2017 to 400 W for the A100 in 2020, 700 W for the H100 in 2022, and 1,200 W for Blackwell in 2024 [S1].
Industry road maps cited in October 2025 pointed beyond 2,000 W per chip, with 5 kW chips being prepared for the foreseeable future [S1]. Rack density has moved from an average of about 8 kW toward 100 kW for AI applications, a tenfold increase that makes air removal increasingly inadequate [S1].
Water has roughly 4 times the specific heat of air, about 800 times its density, and 23.5 times its thermal conductivity, allowing comparable volumes to carry about 3,200 times as much heat [S1]. That advantage applies to liquid systems generally, but it does not determine whether servers should use cold plates or immersion tanks.
Architectural Difference Drives the Adoption Gap
Direct-to-chip systems mount sealed cold plates on high-heat CPUs and GPUs, with coolant circulating through closed channels without contacting the boards or chips [S3]. The system can retain air cooling for components such as DIMMs, making it suitable for targeted deployment rather than a facility-wide conversion [S5].
Single-phase immersion places complete server trays in a dielectric fluid, normally circulating through a coolant distribution unit, pump, and plate heat exchanger before the fluid returns for cooling [S5]. Open-bath vessels can operate at atmospheric pressure because little fluid evaporates, while two-phase immersion uses phase change at the heat-producing components [S1][S5].
The core comparison is therefore straightforward: direct-to-chip wins on retrofit scale, familiar service, and incremental expansion, while immersion wins on component coverage, removal of airflow constraints, and the theoretical density available in the tank [S2].
Cost and Facility Changes Restrict Immersion
Immersion carries materially higher initial capital expenditure than a direct-to-chip retrofit because operators must buy specialized tanks, dielectric fluids, pumps, and monitoring systems [S2]. Tank-based designs also require changes to rack layout, cable routing, and power distribution rather than only adding plates and fluid loops to existing cabinets [S2].
Operational training is a separate obstacle. Hardware compatibility, fluid handling, and tank maintenance differ from established air-cooling and hydronic procedures, so a new skill set is required [S2]. The technology also lacks one unified plant model, which increases procurement and commissioning uncertainty [S2].
Direct-to-chip deployment can start with the racks carrying the highest processor thermal design power and expand after operational validation, without replacing the rest of the data hall [S3]. That flexibility matters where power constraints are already 100 to 150 kW per rack and future designs may reach 600 kW to 1 MW [S4].
Thermal Capability Does Not Equal Commercial Suitability
Immersion can cool CPUs, GPUs, memory, power components, and other electronics rather than limiting liquid contact to selected processors [S2]. It also removes reliance on airflow, allowing denser rack packaging and reducing the cooling power needed to move air through the facility [S2].
Some immersion operators have recorded PUE values of 1.02 to 1.05, and full immersion can reduce partial PUE while shielding components from dust and vibration [S2][S4]. Two-phase systems improve heat transfer through phase change, while single-phase systems transfer heat without boiling the dielectric fluid [S1][S4].
Those benefits do not erase the retrofit burden. A broad cold-plate supply chain and industrial ecosystem already exists, and major AI systems have converged on liquid cooling, including NVIDIA Vera Rubin and Google TPUs [S4]. Once that equipment, maintenance capability, and vendor support are established, replacing the architecture carries cost without a guaranteed proportional thermal gain.
Decision Criteria Separate the Two Systems
For an existing data center with 100 to 150 kW racks, direct-to-chip offers the lower disruption path because it can be concentrated on the highest-density workloads and expanded gradually [S3][S4]. It is the stronger choice where facility teams value familiar procedures, limited redesign, and reversible deployment.
For a purpose-built facility, immersion becomes more credible because the tank, fluid circulation, cabling, power, and service process can be engineered together from the start [S2]. It fits operators that prioritize whole-server heat capture, very high density, or a new facility with no sunk conventional-rack architecture.
Single-phase immersion offers simpler phase management than a boiling system, while two-phase immersion provides stronger heat transfer through evaporation and condensation [S1][S5]. Neither option automatically delivers lower total cost: fluid replacement, component compatibility, maintenance access, facility conversion, and operator training must be priced alongside energy performance.
Immersion is also less suitable for organizations unable to change the supply chain, quality-control system, and maintenance model for large-scale deployment. Cold plates are better where the objective is to remove a thermal constraint without committing the whole facility to a different service architecture [S2][S4].
Adoption Will Stay Selective as Densities Rise
Immersion is used in cryptocurrency mining, experimental high-performance computing, and some edge-computing installations, but those cases have not displaced direct-to-chip in large-scale AI deployments [S2]. The market estimate of US$3,116.7 million by 2033 at a 24.1% compound annual growth rate indicates expansion from a smaller base, not proof that immersion will become the default [S8].
The practical signal is rack density. Direct-to-chip addresses the 100 to 150 kW generation with an established ecosystem, while immersion becomes harder to ignore as designs approach 600 kW to 1 MW and the cost of rebuilding conventional facilities rises [S4].
A second signal is whether new data centers treat fluid-filled servers as part of the base building design instead of as retrofit equipment [S2]. Until tanks, fluid logistics, maintenance, and service procedures become standardized, direct-to-chip will retain its deployment advantage despite immersion capturing more heat from each server [S2][S4].
The near-term decision is straightforward: specify direct-to-chip for incremental AI rack upgrades and risk-sensitive retrofits, while evaluating immersion for purpose-built, high-density facilities where tank architecture, trained operations, and whole-server cooling justify the conversion.