Benefits of Immersion Cooling for AI Workloads in 2026

The benefits of immersion cooling for AI workloads are measurable: partial PUE of 1.02 to 1.03 in shipping single-phase systems, a 10 to 11 percent reduction in server power from fan removal, zero evaporative water use when paired with dry coolers, and 100 to 360 kW of heat rejection per tank. Notably, the constraint is equally measurable: NVIDIA’s GB200 NVL72, GB300 NVL72, and Vera Rubin NVL72 racks are direct-to-chip reference designs, and no NVIDIA rack-scale system ships as an immersion design. Specifically, this article quantifies both sides and states which cooling architecture is standard practice at each rack density in 2026.

Why cooling is now the gating constraint for GPU density

Uptime Institute’s 2025 survey reports a weighted-average PUE of 1.54, the sixth consecutive year the figure has virtually stood still, with most operators adopting racks in the 10 to 30 kW range and few exceeding 30 kW ([1]; [2]).

AI racks sit far outside that band. In particular, NVIDIA states that hyperscale AI racks now exceed 135 kW compared with roughly 20 kW for a traditional rack, and that cooling has historically consumed up to 40 percent of data center electricity ([3]). For example, a single DGX GB200 NVL72 rack draws about 120 kW, weighs roughly 1.36 metric tons, and packs 18 1U compute nodes at 5.4 to 5.7 kW each ([4]).

Google’s OCP EMEA presentation states that machine learning will require more than 500 kW per IT rack before 2030 and describes a ±400 VDC architecture for 1 MW racks ([5]). NVIDIA’s Rubin Ultra NVL576 “Kyber” rack is expected to draw about 600 kW when it ships in the second half of 2027 ([6]). Importantly, air cannot carry that heat: water moves roughly 4,000 times more heat per unit volume than air for the same temperature change ([5]).

What immersion cooling is, and the two ways it is done

Immersion cooling submerges the entire server in a non-conductive dielectric fluid. As a result, the fluid contacts every component, so server fans are removed or disabled, and heat leaves the tank through a coolant distribution unit and a liquid-to-liquid heat exchanger to a facility water loop ([7]).

Single-phase immersion uses a hydrocarbon or synthetic fluid that stays liquid; pumps circulate it past the electronics and through the heat exchanger. By contrast, two-phase immersion uses a fluorinated fluid with a low boiling point, typically around 50°C, which boils on hot components; the vapor condenses on coils at the top of a sealed tank and drips back ([8]; [9]).

OCP’s immersion design guidelines list volumetric specific heat of roughly 4,182 J/L·K for water, 1,855 J/L·K for hydrocarbon fluids, 1,831 J/L·K for fluorocarbons, and 1.3 J/L·K for air ([10]). In addition, dielectric fluids carry over 1,400 times more heat per liter than air but less than half as much as water, which is why water-glycol cold plates reach higher heat flux than a bath.

The benefits of immersion cooling for AI workloads, quantified

Partial PUE of 1.02 to 1.03 in production systems

GRC’s ICEraQ Series 10 datasheet specifies a partial PUE below 1.03, 200 kW of capacity per system with 32°C chiller-free water, and 368 kW with 13°C water ([11]). For instance, at the Texas Advanced Computing Center, the Lonestar6 immersion deployment ran 280 kW across four tanks at 70 kW per rack with a 2 percent cooling overhead, for a pPUE of 1.02 ([12]).

LiquidStack’s two-phase DataTank 48U is rated for up to 252 kW with a pPUE of 1.02 to 1.03 and zero IT fan power ([13]). Similarly, Submer’s SmartPod platform lists a mechanical PUE of 1.015 with CDU power of 750 W per 50 kW tank ([7]). Notably, Alibaba’s large-scale immersion deployment reported a stabilized PUE of about 1.07 against 1.5 for its air-cooled facilities ([14]).

Facility-level numbers are higher because pPUE excludes distribution losses and the heat-rejection plant. Specifically, GRC’s TACC case study reports a facility PUE of about 1.1 for Lonestar6 ([15]). Firmus reports a facility PUE of 1.10 for its immersion-cooled H100 cloud and a design PUE below 1.05 for its 300 kW Cube300 containers ([16]; [17]).

Fan power eliminated: 10 to 11 percent of server power

Fan removal is the benefit that lands inside the IT load, where PUE does not count it. In particular, GRC’s datasheet states an industry-average 11 percent reduction in server power draw from fan removal, and the TACC deployment measured a 10 percent reduction ([11]; [12]). For example, microsoft’s two-phase pilot at Quincy, Washington reported a 5 to 15 percent server power reduction ([8]).

For GPU servers the effect is larger. Firmus measured an immersed H100 HGX system at 6.58 kW against 9 to 11 kW for the same platform air-cooled, a roughly 30 percent server-level saving, in its June 2024 MLPerf power submission ([16]). Importantly, that figure is vendor-published and includes power capping, not an independent audit.

Water: immersion is compatible with zero evaporative water, but so is direct-to-chip

Immersion tanks reject heat to a closed facility water loop; whether the site consumes water depends on what sits on the roof. As a result, GRC and Submer both specify chiller-free operation with inlet water up to 32°C, which lets a dry cooler do the job in most climates without evaporation ([11]; [7]). By contrast, Submer’s EXO tank accepts hot water up to 60°C ([18]).

US data centers directly consumed about 66 billion liters of water in 2023 at a site WUE of just over 0.36 L/kWh, and LBNL projects hyperscale direct water use of 60 to 124 billion liters in 2028 ([19]). In addition, NVIDIA states that evaporative cooling consumes millions of gallons of water per megawatt annually ([3]).

Immersion does not hold this advantage alone. Microsoft’s zero-water-for-cooling design, standard for all new sites since August 2024, uses chip-level liquid cooling in a closed loop and avoids more than 125 million liters per year per data center ([20]). For instance, the lever is warm-water liquid cooling plus dry coolers, and both immersion and direct-to-chip pull it.

Density per tank

Shipping single-phase tanks are rated at 100 kW per rack-equivalent for GRC’s ICEraQ SX, 140 kW and up for Submer’s EVO, over 110 kW with ASHRAE W3 water for LiquidStack’s single-phase system, and up to 361 kW for Submer’s EXO with 50 RU of capacity ([21]; [22]; [23]; [18]). Similarly, Vertiv’s CoolCenter Immersion, launched in EMEA in November 2025, spans 25 to 240 kW ([24]).

A tank is not a rack. Notably, LiquidStack’s single-phase tank occupies the footprint of four standard racks, so density per square meter is lower than the per-tank kW suggests ([23]).

Hardware reliability: one sourced figure

Vendors routinely claim lower failure rates; almost none publish data. Specifically, the one large operator figure comes from Alibaba, which reported a 50 percent lower failure rate for hard disk drives in its immersion-cooled storage compared with air cooling ([14]). In particular, iceotope’s chassis-level immersion trial with Meta held a 72-drive JBOD to a 3°C temperature spread at 40°C inlet water with cooling power below 5 percent of IT power ([25]). For example, no published, controlled GPU failure-rate comparison between immersion and direct-to-chip exists as of September 2026.

Immersion cooling benefits and challenges: the constraints operators actually hit

NVIDIA’s rack-scale systems are direct-to-chip designs

GB200 NVL72 is a liquid-cooled rack-scale system connecting 36 Grace CPUs and 72 Blackwell GPUs, and GB300 NVL72 is fully liquid-cooled ([26]). Importantly, the liquid is water-glycol on cold plates. The DGX GB200 NVL72 takes 25°C coolant at about 2 liters per second with a 20°C rise, while NICs and storage remain air-cooled by 40 mm fans ([4]).

NVIDIA’s Vera Rubin NVL72 MGX compute tray is 100 percent liquid-cooled and designed for 45°C liquid, with a liquid-cooled busbar and 800 VDC power; the same MGX rack footprint carries GB300 NVL72 and Vera Rubin NVL72, and Kyber is designed for 576 Rubin Ultra GPUs by 2027 ([27]). As a result, the ecosystem reference designs are direct-to-chip as well: Vertiv’s 7 MW GB200 NVL72 blueprint supports up to 132 kW per rack with XDU coolant distribution units, and CoolIT’s CHx2000 CDU serves 12 GB300 NVL72 racks at 2 MW ([28]; [29]).

Immersing an NVL72 rack is not a supported configuration. By contrast, immersion for NVIDIA GPUs in 2026 means HGX-form-factor servers rebuilt for tanks, such as the Dell XE9680-IR with eight 700 W H100 or H200 SXM5 GPUs, or Hypertec’s TRIDENT servers ([21]; [30]). In addition, SAVRN’s comparison of direct-to-chip and immersion covers the architectural split in more depth.

Warranty and support posture

OCP’s warranty guidelines state that coolants should be specified as compatible by the IT equipment manufacturer before testing so the warranty remains intact, that typical hardware warranties run 1 to 7 years, and that long-term public information on immersion product viability is “extraordinarily limited” ([31]).

Where explicit coverage exists, it is narrow and named. For instance, Intel introduced an immersion warranty rider for Xeon processors in 2022 and now certifies specific fluid-tank-server combinations for 4th and 5th Gen Xeon, including Shell fluids in Submer tanks with Supermicro servers, and ExxonMobil fluids in Baltimore Aircoil tanks with Hypertec servers ([32]; [33]; [34]). Dell has provided warranty coverage for Dell servers in GRC immersion systems under an OEM agreement since 2020 ([35]).

NVIDIA publishes no equivalent immersion warranty rider for data center GPUs, and its public liquid-cooling statements describe direct-to-chip cold plates, CDUs, and warm-water loops ([3]). Similarly, operators immersing GPUs rely on the server OEM or integrator to carry the warranty.

3M’s PFAS exit and the two-phase market

3M announced in December 2022 that it would exit all PFAS manufacturing, including the fluorinated fluids sold as Novec and Fluorinert, by the end of 2025; the business represented about $1.3 billion in annual sales ([36]). Notably, the exit completed on schedule at the end of 2025 ([37]).

Two-phase immersion depended on those fluids. LiquidStack, ZutaCore, and Wiwynn referenced Novec, while single-phase vendors such as GRC, Submer, and Iceotope were unaffected ([38]). Specifically, replacements such as Halocarbon’s HFMOP, boiling at 50°C, remain fluorinated chemistries, and Halocarbon estimates Novec held about 80 percent of the two-phase market ([39]). In particular, Asperitas rates the regulatory risk of two-phase fluids as high and single-phase hydrocarbons as low ([9]). For example, LiquidStack launched a single-phase product in November 2023 ([23]).

Fluid cost, tank weight, and floor loading

Shell Immersion Cooling Fluid S5 X lists at $3,365.72 per 208-liter drum from a US distributor as of August 2025, about $16 per liter or $61 per gallon ([40]). Importantly, a Submer SmartPodXL holds 1,186 liters of coolant, so filling one 50 kW tank at that list price is roughly $19,000 in fluid, or about $380 per kW before IT displacement is credited ([7]). As a result, fluid life offsets this: Hypertec lists a 15-year coolant life for single-phase against 2 years for two-phase ([30]).

Weight follows volume. By contrast, the SmartPodXL weighs 671 kg empty and 1,691 kg full before servers; the SmartPodX weighs 411 kg empty and 872 kg full ([7]). In addition, LiquidStack’s 48U two-phase DataTank weighs 2,500 kg ([13]). For instance, GRC specifies a floor loading of 822 kg/m², or 168 lb/ft², excluding IT equipment ([11]). Similarly, Castrol notes that immersion baths impose greater structural floor loading than air-cooled racks ([41]), which is why immersion retrofits gravitate to slab-on-grade halls.

Serviceability, optics, and material compatibility

Servicing a tank means lifting a wet server on a hoist and letting it drain. Notably, Submer specifies 2N pump and CDU redundancy so a tank keeps running during service ([18]). Specifically, Alibaba reports about 30 minutes of thermal inertia in a bath if cooling is lost, roughly 1,000 times that of air ([14]).

Optics are the sharpest compatibility problem for AI clusters. In particular, when an optical transceiver is immersed, fluid replaces the air gap between fiber ferrules and the changed refractive index causes reflection loss; OCP’s guidance is to use direct-attach copper, sealed transceivers, or port extenders that keep optics above the fluid line ([10]). For example, the same document warns that thermal pastes can dissolve and should be replaced with indium foil, that cable plasticizers leach into the fluid, that heat-shrink degrades above 60°C, and that spinning drives must be helium-sealed. NVIDIA’s NVL72 uses copper NVLink cabling partly because optics would add about 20 kW per rack, so the optics budget is already tight before immersion enters ([4]).

How do data centers cool 100kW+ AI racks without going full immersion?

The 2026 answer is direct-to-chip liquid cooling fed by coolant distribution units, with room air or a rear-door heat exchanger handling the residual heat. Importantly, Supermicro’s DLC-2 racks capture up to 98 percent of system heat on cold plates covering CPUs, GPUs, PCIe switches, DIMMs, VRMs, and power supplies, support up to 250 kW per rack, and run 45°C warm water to eliminate chillers ([42]).

Motivair’s MCDU-70, announced in January 2026, delivers 2.5 MW at 1.5 liters per minute per kW ([43]); Schneider Electric closed its 75 percent acquisition of Motivair in February 2025 at a Reuters-reported value of about $850 million ([44]; [45]). Google has run direct-to-chip cooling across more than 2,000 TPU pods over seven years at roughly 99.999 percent CDU availability since 2020 ([5]).

Motivair’s ChilledDoor rear-door heat exchanger is rated to 75 kW per rack and Supermicro’s RDHx to 120 kW, both neutralizing server exhaust with facility water without touching the server ([46]; [42]). As a result, aSHRAE’s liquid-cooling classes now embed the facility water upper limit in the class name, W17 through W45 and W+, replacing the old W1 through W5 scheme whose upper limits were 17, 27, 32, 45, and above 45°C ([47]; [48]).

Cooling architecture comparison

What are the key differences between single-phase and two-phase immersion cooling for AI GPU clusters?

Single-phase moves heat by pumping a stable hydrocarbon past the components; two-phase moves heat by boiling a fluorinated fluid on the components and condensing the vapor. The second mechanism carries more heat per unit of fluid and needs no pump, which is why two-phase tanks reached 252 kW earlier than single-phase tanks ([13]). By contrast, the cost is a hermetically sealed tank, vapor loss on every service event, and a fluid supply chain that lost its dominant producer at the end of 2025.

Is immersion cooling actually practical for GPU clusters exceeding 100kW per rack, or are CDUs the better path?

Both paths are practical at 100 kW; only one is the reference design. For NVIDIA rack-scale systems, CDUs feeding cold plates are the only supported path, because the NVL72 rack arrives built for 25 to 45°C water ([4]; [27]). In addition, for HGX-class 8-GPU servers, immersion is a working alternative with server-OEM warranty coverage, and it captures the fan power that cold plates leave behind ([21]).

The framework below states standard practice by density band as of 2026; SAVRN’s review of liquid cooling for AI covers the plant-side choices behind each row.

PUE 1.1 achievable: what does it actually take?

PUE 1.1 is achievable with any liquid architecture, and several operators publish it. For instance, LUMI in Finland runs at 1.04 on direct-to-chip cooling with heat sold to a district heating network ([49]). Similarly, google’s fleet trailing-twelve-month PUE was 1.09 through 2025 ([50]). TACC’s immersion deployment sits near 1.1, and Firmus reports 1.10 ([15]; [16]).

Three conditions recur at those sites. First, more than 80 percent of IT heat leaves in liquid, so CRAH fan energy collapses. Second, facility water runs warm enough, 32 to 45°C, that dry coolers or free cooling replace compressors for most hours ([11]; [27]). Third, electrical losses are minimized, which is where 800 VDC and eliminated 200 kg copper busbars enter the calculation ([27]).

PUE also misses the fan saving inside the servers: a site can hold 1.1 while immersion cuts IT load by 10 percent or more, which is why PUE alone is an incomplete metric for AI facilities. Notably, microsoft’s zero-water design is the mirror case, accepting a nominal PUE increase from mechanical cooling for a WUE near zero ([20]).

Water-free operation: immersion, direct-to-chip, and dry coolers

Neither immersion nor direct-to-chip consumes water by itself; the question is whether the facility loop terminates in an evaporative tower, a chiller, or a dry cooler. Specifically, dry coolers use no water but require IT equipment that tolerates warm water, which NVIDIA identifies as the condition for water-free AI factories ([3]).

GRC accepts inlet water from 5 to 32°C, Submer recommends inlet up to 32°C with a 37°C outlet, and Submer’s EXO handles up to 60°C ([11]; [7]; [18]). In particular, direct-to-chip has caught up: Vera Rubin NVL72 is designed for 45°C liquid, and Supermicro’s DLC-2 runs 45°C warm water with no chillers ([27]; [42]). LBNL’s modeling confirms the trade: air-cooled chillers and dry coolers use no water but more energy, while evaporative systems use less energy and more water ([19]).

The exception is hot climates. For example, a dry cooler cannot reject heat to 45°C ambient air with 45°C supply water, so W45 designs in Phoenix-class climates need trim chillers or adiabatic assist at peak hours, which reintroduces some water ([19]). Importantly, higher tank outlet temperatures widen the dry-cooler window, which is immersion’s remaining structural advantage on water.

How SAVRN builds for this

SAVRN’s Atom compute block is a factory-built, liquid-cooled unit designed around the direct-to-chip architecture that NVIDIA’s GB300 NVL72 and Vera Rubin NVL72 racks require, because that is the only path with a supported warranty and a published 45°C liquid specification ([27]). As a result, heat rejection is closed-loop to dry coolers, which is what makes the water-free AI data center design possible without evaporative towers.

Warm facility water is the design choice that ties the pieces together. Running the loop in the ASHRAE W45 class lets the same block serve cold-plate racks today and single-phase immersion tanks for HGX-class servers where an operator wants the fan-power saving and holds OEM warranty coverage, since both GRC and Submer accept 32°C water and Submer’s EXO accepts 60°C ([11]; [18]). Structural design follows tank-class loads rather than standard rack loads, because GRC’s published 822 kg/m² requirement excludes IT equipment and exceeds typical raised-floor ratings.

Power comes from SAVRN’s Electron family of on-site generation, so the block is self-powered and does not depend on grid interconnection timelines, and its distribution follows the 800 VDC direction NVIDIA and its MGX partners set at OCP 2025. Cooling, power, and structure are sized for the 130 to 600 kW band before the first rack arrives.

Chad Harris, Founder, SAVRN