Direct-to-Chip vs Immersion Cooling: Fluid Types, Chemistry, and Selection Rules

Liquid cooling is no longer a specialty deployment—it is the default path for any data center designed to host GPU-dense AI workloads. The question facility engineers are now forced to answer during early site planning is not whether to use liquid but which kind, and what that decision implies for fluid chemistry, procurement complexity, and long-run operating cost. The three main architectures—direct-to-chip cold plates, single-phase immersion, and two-phase immersion—are not interchangeable, and the wrong choice at the architectural stage is difficult and expensive to reverse once concrete is poured and racks are populated. This article works through the technical and chemical distinctions that matter to engineers making that call in 2026 and beyond, with enough specificity on fluid types, inhibitor chemistry, and selection criteria to support both technical procurement and facility design decisions.

The Cooling Architecture Landscape: From Air to Two-Phase

Five cooling architectures are actively deployed across the data center industry today, ranging from legacy forced-air configurations that still dominate enterprise refresh cycles to two-phase immersion systems occupying a small but intensely watched niche. Understanding where each sits on the density-versus-complexity curve is the prerequisite for any fluid chemistry discussion, because fluid selection is downstream of architecture, not the other way around.

Forced-air cooling—row-based CRAC/CRAH units with hot-aisle/cold-aisle containment—remains the installed base majority worldwide. It is fully retrofit-compatible, requires no special fluid management, and benefits from a mature service ecosystem. Its ceiling, however, is approximately 20–30 kW per rack under aggressive airflow conditions, and that ceiling is increasingly irrelevant to AI GPU workloads. A single H100 SXM5 node draws 10.2 kW; a standard 8-GPU DGX H100 system draws approximately 10.2 kW under sustained inference and up to 14 kW under training. Four such nodes saturate what a well-designed air-cooled rack can handle, and rack densities for hyperscaler AI clusters are already landing at 40–80 kW in production deployments. Air is architecturally dead for this workload class.

Rear-door heat exchangers (RDHx) represent an intermediate step: chilled facility water passes through a door-shaped heat exchanger bolted to the rear of a standard rack, cooling the exhaust air before it re-enters the room. RDHx is one of the more practical retrofit options because it requires no server modification, works with existing rack form factors, and handles densities up to roughly 40–50 kW per rack depending on chilled water supply temperature. The chemistry in this loop is conventional building HVAC chemistry—inhibited glycol blends are common—and the engineering is well understood. RDHx buys time for facilities that cannot immediately retrofit to cold plates, but it is not a long-term architecture for AI-scale density.

Direct-to-chip (D2C) cold-plate cooling is the current mainstream choice for GPU-dense builds. Cold plates are brazed or diffusion-bonded aluminum or copper heat exchangers that mount directly to the CPU, GPU, and memory packages, replacing the conventional heatsink. Liquid is circulated through the cold plate, picks up heat directly from the die package, and returns to a facility-side heat rejection system. The server still has fans for residual component cooling (VRMs, storage, NICs), making D2C a hybrid system—typically described as liquid-cooled with air-assisted residual cooling. Rack densities of 40–120 kW are achievable depending on chassis configuration, and OEM support from NVIDIA, AMD, Intel, and the major ODMs (Wiwynn, Quanta, Supermicro) is now mature. This is the architecture that dominates new builds in 2026.

Single-phase immersion submerges complete servers—without modification for some platforms, with modifications for others—in a dielectric fluid inside an open or sealed tank. The fluid does not change phase; it flows, absorbs heat by convection, and is pumped to an external heat exchanger. Densities above 100 kW per rack are achievable, PUE can approach 1.03, and the tankless, open-bath architecture makes hardware access straightforward. The tradeoffs are structural (tank weights run 2,000–4,000 kg fully loaded), fluid cost (dielectric fluids run $3–12 per liter versus $0.15–0.40 per liter for glycol-water blends), and the absence of a standard OEM ecosystem at scale.

Two-phase immersion submerges servers in a fluid that boils at low temperature at atmospheric or near-atmospheric pressure, transferring heat through the latent heat of vaporization rather than sensible heat. Theoretical thermal performance is excellent. The practical supply chain for two-phase fluids collapsed significantly after 3M's 2022 announcement of the planned discontinuation of its Novec family. This architecture remains in production at select hyperscaler research sites but is not a mainstream recommendation for new builds today.

The fluid CapEx index is normalized to untreated municipal water as 1×. Inhibited propylene glycol blends (D2C) run approximately 3–5× by volume cost; polyalphaolefin and mineral oil dielectrics (single-phase immersion) run 20–60×; two-phase fluorinated engineered fluids run 80–200×. These multipliers become material at hyperscale volumes—a 500-rack D2C deployment uses roughly 5,000–15,000 liters of coolant depending on loop sizing, while the equivalent immersion deployment would use 250,000–500,000 liters of dielectric at a minimum. The math is unambiguous.

D2C Cold-Plate Chemistry: What Is Actually in the Loop

The fluid inside a direct-to-chip cooling loop is almost never pure water, and it is never untreated tap water. The chemistry of this fluid has direct consequences for heat exchanger longevity, cold-plate corrosion rates, pump seal compatibility, and GPU die package reliability—and the OEMs responsible for those GPU packages publish explicit chemistry specifications that procurement teams need to treat as hard constraints, not suggestions.

The dominant fluid class for D2C loops is inhibited glycol-water blends, with propylene glycol (PG) preferred over ethylene glycol (EG) in most data center applications for two reasons: PG's lower acute toxicity reduces leak response burden under occupational safety standards, and PG is more compatible with incidental contact with aluminum heat exchangers without accelerating pitting corrosion. Ethylene glycol blends remain common in facility-side chilled water loops where human contact risk is lower and the existing infrastructure was built around EG-based chemistry. At the cold plate—the server-side loop—PG-based blends are the default recommendation from NVIDIA, AMD, and the major ODM ecosystem.

The typical D2C loop specification calls for a glycol concentration of 20–35% by volume, with the balance being deionized (DI) water. The glycol concentration is not primarily about freeze protection—most data center cold-plate loops operate at supply temperatures of 15–45°C and are not in environments where freezing is a credible risk—but about providing a stable carrier medium for corrosion inhibitors and maintaining predictable viscosity across the operating temperature range. At 25°C, a 30% PG blend has a viscosity of approximately 2.5–3.0 cP compared to 0.89 cP for pure water; this is an acceptable pump energy penalty for the corrosion protection the blend provides.

Corrosion inhibitor packages in D2C fluids serve multiple simultaneous functions: they protect copper cold plates from pitting and dealloying, protect aluminum manifolds from oxidation and crevice corrosion, passivate brazed joints, and suppress microbial growth in systems where loop temperatures fall within biological growth ranges. The inhibitor chemistry is typically a blend of azole compounds (tolyltriazole, benzotriazole, or both) for copper protection, molybdate or silicate compounds for aluminum passivation, and organic acid buffer systems for pH stability. pH is typically maintained in the 7.0–8.5 range; outside this range, inhibitor effectiveness degrades and both aluminum and copper corrosion rates accelerate nonlinearly.

The conductivity specification for D2C loop fluid is the single most frequently cited chemistry parameter in OEM documentation, and its rationale is worth explaining carefully. NVIDIA's cold-plate specifications for H100 and H200 SXM systems require coolant conductivity at 25°C to remain below 25 µS/cm. AMD's Instinct MI300X cold-plate documentation specifies a similar ceiling, as do specifications from Intel's Gaudi 3 documentation and the major ODM chassis manufacturers. The 25 µS/cm ceiling is not arbitrary—it represents the point above which electrochemical current flow through the coolant becomes sufficient to create measurable galvanic corrosion between the copper cold plate and the aluminum manifold and stainless steel fittings in a typical loop. Below 25 µS/cm, the coolant is resistive enough that ion migration rates stay below the threshold for accelerated material loss.

In practice, new DI water has a conductivity of 0.1–1.0 µS/cm. Once glycol and inhibitor package are added, loop conductivity typically rises to 5–15 µS/cm, well within spec. The risk is conductivity rise over time due to inhibitor degradation byproducts, corrosion products entering solution, and microbial contamination. A properly designed D2C loop should include conductivity monitoring with an alarm threshold at 20 µS/cm—leaving a 5 µS/cm margin before OEM spec breach—and an ion exchange or bypass filtration system to maintain conductivity below that threshold throughout the loop's service life. Particulate filtration is equally important: cold plates with microchannel or minichannel geometry can be occluded by particles in the 50–200 µm range. NVIDIA's DGX H100 documentation specifies particulate filtration to 50 µm; Supermicro's SYS-421GE-TNRT documentation specifies 100 µm. A 25–50 µm nominal filter on the loop return line is conservative practice.

Loop water analysis should be performed quarterly for the first year of operation and annually thereafter once chemistry stability is confirmed. The dissolved metals assay—copper and aluminum specifically—is more diagnostically useful than conductivity alone because it identifies which materials are actively corroding. A conductivity reading of 18 µS/cm with copper at 0.05 mg/L and aluminum at 0.3 mg/L indicates a different chemistry problem than the same conductivity reading with copper at 0.4 mg/L. Both are approaching alarm thresholds, but the remediation paths differ.

Why Direct-to-Chip Wins the 2026–2030 Build Cycle

The case for D2C as the default architecture for new AI compute builds over the next five years is not primarily a thermal argument—it is a systems argument. D2C delivers thermal performance that is adequate for current and near-term GPU TDP profiles, does so within a standard rack form factor that preserves interoperability with existing data center infrastructure, uses chemistry that is well-understood and supported by a deep supply chain, and benefits from an OEM ecosystem that has now fully committed to cold-plate support as a first-class product line rather than a premium option. The operational simplicity of D2C over immersion architectures is not incidental—it is the decisive factor for most operators.

On the thermal side, the H100 SXM5 is a 700W TDP package; the H200 is nominally similar; NVIDIA's next-generation Blackwell architecture (B200, GB200) carries TDPs in the 1,000–1,200W range. D2C cold plates operating with 15–25°C supply water can handle a 1,200W GPU package with junction-to-coolant thermal resistance in the 0.05–0.08°C/W range depending on cold-plate design and flow rate. At 1,200W, that means junction temperature rise above coolant supply temperature of 60–96°C—well within junction temperature ratings for current silicon. Single-phase immersion, by comparison, achieves similar or slightly better junction temperature performance through whole-system immersion, but the marginal thermal advantage at current TDP levels is not sufficient to justify the infrastructure complexity differential unless density is already a hard constraint.

The retrofit feasibility of D2C is a significant practical advantage. Existing raised-floor and slab-on-grade data centers with chilled water infrastructure can be adapted for D2C deployment by installing facility-side CDUs (coolant distribution units) that interface between the building chilled water loop and the server-side fluid loop. The rack footprint, power distribution architecture, and fire suppression systems all remain compatible. Cold-plate-equipped servers slot into standard 19-inch or 21-inch rack rails. This means a facility built in 2018 for 20 kW/rack air-cooled compute can be upgraded to 60–80 kW/rack D2C compute with a CDU installation, cold-plate server procurement, and loop chemistry management—no structural modifications, no tank installations, no fire suppression redesign.

Andre Taki, Lead Product Specialist — Cooling Chemistry:
"The engineers I talk to who've priced out immersion for 80 kW/rack builds almost always land back on D2C when they work through the full cost stack—structural reinforcement, tank procurement, fluid volume, and the three-year certification process some hyperscalers require before immersion is approved for live production. D2C chemistry is glycol-water with good inhibitors. We've been doing that in HVAC and industrial cooling for 60 years. The playbook exists. Immersion has a real place, but D2C wins the next five years by not asking operators to learn a new game."

The OEM ecosystem maturity argument is worth quantifying. As of early 2026, NVIDIA offers cold-plate liquid cooling as a standard configuration option on DGX H200 and MGX B200 systems. AMD's Instinct MI325X is available in liquid-cooled configurations from all major ODMs. Supermicro, Wiwynn, Quanta, and Foxconn all produce D2C chassis with published ASHRAE-style fluid specifications. CDU manufacturers including Airedale, Stulz, and multiple Asian ODMs ship interoperable systems. The liquid-cooled server market is not a specialty procurement challenge in 2026—it is a standard enterprise purchase with standard contracts, warranty terms, and service agreements. The same cannot be said for immersion at comparable scale.

The chemistry supply chain for D2C fluids reflects this maturity. Inhibited propylene glycol blends for D2C applications are manufactured to published specifications, available in tote and drum quantities from multiple qualified distributors, carry NSF-61 or equivalent certifications for incidental contact safety, and have ASTM D6130 or equivalent analytical methods for quality verification. Shelf life is typically 2–3 years in sealed containers. Replacement fluid for a 500-rack D2C deployment is a purchase order, not a supply chain event.

Where D2C reaches its limits is predictable and well-characterized: sustained rack densities above 120 kW with next-next-generation GPU TDPs, coupled with hard facility space constraints where density-per-square-foot is a binding constraint rather than a preference. For builds where those specific conditions apply—and they are real conditions at leading hyperscalers planning for post-Blackwell architectures—immersion begins to present a credible architectural alternative. But those builds represent a fraction of the total deployment volume through 2030, and the correct posture for the majority of operators is to default to D2C and revisit immersion when density requirements exceed what D2C can serve.

Single-Phase Immersion: Mineral Oil, Synthetic Hydrocarbons, and When the Math Works

Single-phase immersion is the most operationally approachable of the two immersion architectures, and the one that has seen the most serious production deployments outside of hyperscaler research programs. The basic premise is straightforward: servers, stripped of their fans, are submerged in a dielectric fluid inside an insulated tank. The fluid is circulated by a low-power pump to an external heat exchanger or dry cooler. No phase change occurs; heat is transferred entirely by the sensible heat capacity of the fluid and the convective flow rate the pump maintains. The system is electrically isolated from the server hardware by the dielectric properties of the fluid—bulk resistivity typically above 10^10 Ω·cm for commercial dielectric fluids—and the tanks are either fully enclosed or open-bath depending on vendor design philosophy.

Two fluid chemistry families dominate single-phase immersion deployments: mineral oils and synthetic hydrocarbon fluids, primarily polyalphaolefins (PAO) and alkylated benzenes. Mineral oil—specifically transformer-grade or inhibited white mineral oil—was the first fluid used in commercial single-phase immersion deployments and is still used by some operators due to its low cost, high availability, and long industry track record in electrical transformer cooling. The thermal properties of mineral oil are workable: thermal conductivity is approximately 0.13–0.15 W/(m·K) at 40°C, specific heat capacity is approximately 1.8–2.0 kJ/(kg·K), and viscosity at operating temperature (typically 40–60°C) is in the 10–30 cP range depending on grade and temperature. The viscosity is mineral oil's primary thermal performance liability—higher viscosity means reduced convective flow efficiency and lower effective heat transfer coefficient, requiring larger pump energy input to achieve equivalent heat removal compared to water-based fluids.

The comparison to water-glycol blends in the table above illustrates the thermal conductivity gap that is the fundamental engineering constraint of single-phase immersion: the best hydrocarbon dielectrics achieve thermal conductivities roughly one-third of inhibited glycol-water blends. This is not a solvable problem within the hydrocarbon chemistry class—it is a physical property of the molecular structure. The compensating mechanism is total surface area: in immersion, every surface of every server component is in contact with the dielectric fluid, not just the GPU die package. The aggregate heat transfer surface in an immersion tank is orders of magnitude larger than the cold-plate contact area in a D2C deployment, and this surface area advantage more than compensates for the lower thermal conductivity fluid in practice. Measured PUE figures from production single-phase immersion deployments consistently fall in the 1.02–1.10 range, with facilities in mild climates and waste heat recovery systems achieving the lower end of that range.

Synthetic PAO fluids have largely displaced mineral oil in new single-phase immersion deployments at operators with serious engineering programs. The reasons are: lower viscosity (better convective performance at lower pump energy), higher dielectric strength (greater electrical safety margin for high-voltage power delivery at the server level), better oxidative stability (longer service life before degradation products affect fluid properties), and more consistent quality control compared to natural petroleum-derived products. PAO fluids are synthesized from ethylene or alpha-olefin feedstocks under controlled polymerization conditions; the resulting product has a narrow molecular weight distribution and predictable, reproducible thermophysical properties. Mineral oil properties vary by crude source and refining approach.

The serviceability argument for single-phase immersion is genuinely differentiating. In an open-bath or hinged-lid tank design, hardware access requires draining fluid from the top of the tank, waiting for the server to drain, and lifting the dripping server out of the bath. This is not elegant, but it is possible without specialized tools, without facility downtime, and without purging a sealed loop. Fluid carryout—the volume of dielectric that adheres to the server hardware when removed from the tank—is the primary servicing cost factor: typical carryout rates run 0.5–2.0 liters per server removal event, requiring fluid makeup and adding to the total fluid loss accounting over the deployment lifecycle. Operators running high-availability immersion deployments budget for carryout explicitly in their fluid procurement contracts.

Single-phase immersion makes engineering sense in a narrow but real set of conditions: new-build facilities where structural capacity supports tank weights of 2,000–4,000 kg per tank position, rack density requirements exceeding 100 kW where D2C loop complexity at that density becomes unmanageable, and operators with the operational maturity to manage dielectric fluid programs including quality testing, disposal, and supplier relationships. It is not a retrofit architecture under any practical definition—the structural reinforcement alone typically costs $50–150 per square foot of floor space modified, and that cost must be amortized against the density and PUE benefits.

Two-Phase Immersion and the Post-Novec Transition

Two-phase immersion is the thermal engineering pinnacle of liquid cooling for electronics: servers submerged in a fluid that boils at a temperature low enough to absorb die heat through latent heat of vaporization, with the vapor condensed on a lid-mounted heat exchanger and returned as liquid to the bath. Boiling heat transfer coefficients—the rate at which heat can be transferred per unit area per degree of temperature difference—are dramatically higher than single-phase convective coefficients, which is why two-phase immersion offers the highest theoretical thermal performance per unit of hardware surface area of any cooling architecture. For GPU packages generating 1,000+ W/package at junction temperatures constrained to below 85°C, this matters.

The fluid class that enabled commercial two-phase immersion was 3M's Novec family of engineered hydrofluoroethers (HFEs) and perfluorocarbon (PFC) compounds. Novec fluids had boiling points in the 34–76°C range at atmospheric pressure, dielectric strengths above 40 kV/mm, low toxicity, and zero ozone depletion potential. 3M's Novec 7100, 7200, and 649 were the primary fluids in commercial two-phase deployment. In late 2022, 3M announced that it would discontinue the entire PFAS-containing product portfolio, including Novec, by end of 2025, citing regulatory pressure around per- and polyfluoroalkyl substance (PFAS) regulation in the EU, US, and other jurisdictions.

The post-Novec fluid landscape for two-phase immersion is fragmented and still evolving. Three research and commercial directions are active as of 2026. First, secondary-market engineered HFE fluids from suppliers including Solvay (Galden product line, PFAS-bearing), Daikin, and several Asian chemical manufacturers are available as near-drop-in Novec alternatives for existing deployments. These fluids carry similar PFAS regulatory risk and are not a long-term solution under current European Chemical Agency trajectory. Second, non-PFAS fluorinated fluids based on HFO (hydrofluoroolefin) chemistry are under active development and early commercial introduction; GWP characteristics are more favorable than traditional HFCs, but fluid property optimization for electronics boiling applications is still ongoing and thermal performance data at GPU-scale TDPs is limited. Third, hydrocarbon-based two-phase fluids—primarily neopentane and isopentane formulations with boiling points near 28–36°C—have been evaluated in laboratory settings for electronics cooling applications. The combustibility of light hydrocarbons creates significant facility safety classification challenges and has prevented commercial deployment at scale, though research programs continue.

Honesty requires acknowledging the current state directly: two-phase immersion is a niche architecture in 2026, and it will remain niche until the fluid supply chain stabilizes around a non-PFAS chemistry that achieves the boiling heat transfer performance of Novec fluids at commercial scale and competitive price. Some of the largest AI computing operators—including several hyperscalers with well-publicized sustainability commitments—have run two-phase immersion pilots, but the architecture is not in their standard deployment playbook for new builds. The thermal physics are compelling. The operational reality is not yet there. Engineers building data centers for delivery in 2026–2028 should not plan around two-phase immersion unless they have specific density requirements that neither D2C nor single-phase immersion can meet and have the organizational capacity to manage the fluid supply chain uncertainty this architecture currently carries.

Selection Decision Tree: From Rack Density to Architecture Choice

Architecture selection should follow a structured decision process that starts with actual and projected rack density, incorporates building structural constraints, and accounts for operational capability. The following framework is organized as a decision tree rather than a preference ranking, because the right answer depends on factors specific to each deployment scenario.

Step 1: Determine current and projected peak rack density. This means GPU TDP × GPUs per rack, plus ancillary load (switches, storage, power conversion losses). For a standard 8-GPU server at 1,200W per GPU, 8 servers per rack, plus 15% ancillary, peak rack load is approximately 88 kW. Use three-year projected peak, not current peak—cooling infrastructure is a five-to-ten-year asset and should be sized for the compute density the facility will host, not the compute density it hosts today.

Step 2: Assess retrofit vs. new-build status. If the building exists and has raised-floor or slab-on-grade construction without structural reinforcement capability above current loading, single-phase and two-phase immersion are eliminated from consideration immediately. The weight density of a fully loaded immersion tank—including fluid, hardware, and tank structure—typically runs 1,200–2,000 kg/m². Standard commercial data center slab loading is 500–1,000 kg/m². This is not a minor delta that can be resolved with engineering judgment; it requires structural reinforcement that may or may not be feasible given existing column and beam placement.

Step 3: Evaluate operational capability. Single-phase immersion requires a fluid management program—quality testing, carryout accounting, disposal pathways, and supplier relationships—that most enterprise data center operations teams do not currently maintain. D2C fluid management is substantially simpler: quarterly conductivity and pH checks, periodic filter replacements, and top-up fluid available from standard chemical distributors. The operational maturity gap between D2C and immersion is real and should be weighted heavily by operators who do not have in-house cooling chemistry expertise.

Step 4: Run the TCO model with realistic numbers. The fluid CapEx premium for immersion is significant but not the whole story—energy savings can be dramatic, and density advantages translate directly to reduced facility cost per GPU. However, operators routinely underestimate the total fluid volume required for immersion (tank dimensions, fill levels, and spare inventory requirements add up), the carryout loss rate over the fleet lifecycle, and the fluid disposal cost when tanks are drained for reconfiguration. These factors are addressed in detail in the following section.

For the majority of operators reading this article—those planning AI compute expansions in existing or near-term new-build facilities at densities between 40 and 100 kW per rack—the decision tree terminates at D2C in most branches. That is not a conservative recommendation; it reflects where the OEM ecosystem, operational tooling, and fluid supply chain sit in 2026.

Cost of Fluid TCO: Purchase Price Is a Fraction of the Real Number

Fluid total cost of ownership is consistently misframed in procurement conversations because the initial purchase price is the most visible line item and, for immersion fluids especially, a large absolute number. The framing that serves procurement teams better is to express fluid cost as a fraction of five-year facility cost-per-GPU—a metric that incorporates purchase price, change-out intervals, energy savings, density advantages, and the infrastructure capital that cooling architecture enables or requires.

For D2C deployments, the fluid cost model is relatively straightforward. A standard 500-rack D2C deployment using 30% inhibited PG-water blend requires approximately 8,000–15,000 liters of loop fluid at initial fill, depending on CDU sizing, manifold volumes, and loop lengths. At $0.35–0.60 per liter all-in (DI water + glycol + inhibitor package), initial fluid cost is $2,800–$9,000. Change-out intervals for well-maintained D2C loops with conductivity monitoring and bypass filtration are typically 3–5 years; without monitoring, inhibitor depletion can force replacement at 18–24 months. Annual fluid maintenance cost (top-up, replacement fluid, consumable filters) for a 500-rack D2C deployment runs $15,000–$40,000 depending on leak rates and monitoring rigor. Over five years, total fluid cost for a 500-rack D2C deployment is approximately $90,000–$250,000—a rounding error in the context of a facility that represents $50–150 million in total capital.

The table above uses a 500-rack deployment with 10 MW IT load as the basis, with energy cost assumptions of $0.06–0.08/kWh (wholesale data center rate, US market). The energy savings figures are the differential between the architecture's PUE and a baseline air-cooled PUE of 1.45—a reasonable reference for a well-operated air-cooled facility. At 10 MW IT load, dropping from 1.45 PUE to 1.08 PUE saves approximately 3.7 MW of facility cooling energy, worth roughly $1.8M per year at $0.06/kWh. Over five years, that is $9M in energy savings from fluid cost investment of $1.1–$3.5M—a compelling ROI, provided the structural CapEx and fluid procurement complexity are manageable for the operator.

The density premium is where immersion's TCO argument is strongest and most often understated. A single-phase immersion deployment that achieves 200 kW per rack position in a 1,000 m² data hall delivers 20,000 kW in a space that a 30 kW/rack air-cooled deployment would need 6,700 m² to match. At $800–$1,500 per square foot for purpose-built data center construction in US markets, the density advantage translates to $40M–$75M in avoided facility construction cost per 10,000 kW of IT load. That math is what drives hyperscaler interest in immersion for extreme-density applications: the fluid cost, even for PAO or engineered HFE fluids, is a small fraction of the facility construction savings when density is the binding constraint.

For operators whose density requirements are in the 40–85 kW range—which is the majority of the market through 2028—the density premium argument for immersion is less compelling. D2C at 80 kW per rack is three to four times denser than air-cooled compute; the marginal density gain from moving to immersion at that density is not worth the fluid cost, structural investment, and operational complexity differential. The crossover point where immersion's density advantage justifies its TCO premium is approximately 100–120 kW per rack for single-phase, based on current construction costs and energy pricing in US data center markets. Below that density, D2C wins the TCO model in most scenarios.

Procurement teams evaluating fluid suppliers should require certificate of analysis documentation for every lot, verify that DI water used in glycol blends meets ASTM D1193 Type II or better, and confirm that inhibitor packages are tested for compatibility with the specific metals present in their cold-plate loop—not just generic compatibility claims. For immersion fluids, dielectric strength testing per ASTM D877 or IEC 60156 and oxidative stability per rotating bomb oxidation test (RBOT/RPVOT, ASTM D2272) should be baseline acceptance criteria. These are standard industrial analytical tests, not specialty requirements, and suppliers unable to provide this documentation should not be on the shortlist.

The five-year view on cooling fluid procurement is fundamentally a strategic chemistry decision, not a commodities purchase. The architecture you choose determines the fluid class you need; the fluid class determines your supply chain, your maintenance program, and your long-run operating cost. Getting the chemistry right at commissioning—correct inhibitor package, correct glycol concentration, verified DI water quality, proper conductivity and pH baseline—is the single highest-leverage preventive maintenance action available to D2C facility operators. For immersion operators, establishing a fluid quality testing program before the first server is submerged is equally foundational. In both cases, the fluid cost is recoverable through energy savings and extended hardware life if managed correctly, and is a source of expensive, unplanned remediation costs if ignored.

For deeper dives into the specific fluid classes and chemistry protocols discussed in this article, see the companion sub-pillars on this site: OAT vs. NOAT vs. HOAT Inhibitor Packages for Data Center D2C Loops, which covers the specific corrosion inhibitor chemistry families and their applicability to mixed-metal cold-plate loops; and GPU Thermal Density, Coolant Flow Rate Optimization, and Cold-Plate Specification, which addresses the heat transfer engineering behind D2C cold-plate selection and flow rate sizing. All three articles are part of the Alliance Chemical AI & Data Center Cooling Chemicals resource center at /pages/chemicals-for-data-center-cooling, where you can also find product specifications for Alliance's inhibited propylene glycol blends, DI water formulations, and loop maintenance chemistry products qualified for data center D2C applications.