Two-Phase vs Single-Phase Cooling: Fluid Selection for Hyperscale Immersion

Immersion cooling has crossed from proof-of-concept to production procurement. Hyperscalers are writing RFPs, colocation operators are issuing RFIs, and fluid chemists are fielding calls from facility engineers who six months ago had never specified a dielectric fluid. The market is real — but it is also genuinely bifurcated. Two fundamentally different physical approaches compete for the same rack space, each demanding a different fluid chemistry, a different containment strategy, and a different relationship with the regulatory calendar. Getting the selection wrong at the specification stage means replumbing a tank farm in year two. This article is a structured comparison: heat-transfer physics, fluid supply chain, operational realities, and a frank view of where each technology belongs in a 2026 data center build plan.

The Two Architectures Explained

Single-phase immersion cooling places server hardware — boards, processors, power supplies, drives — directly into a bath of dielectric fluid. That fluid is kept liquid throughout the entire cooling cycle. It absorbs heat through convective and conductive transfer as it contacts hot surfaces, rises by thermosiphon or is pumped through a primary loop to a heat exchanger, sheds heat to facility cooling water, and returns cooled to the tank. The fluid never changes state. Temperature differentials across the loop typically run 20–35°C between fluid inlet and outlet. Pumping power is non-trivial because you are moving a dense, viscous liquid in volume. But the engineering is legible, the containment requirements are modest (open or lightly sealed tanks), and the fluid is chemically stable across years of operation.

Two-phase immersion cooling starts with the same premise — servers submerged in dielectric fluid — but exploits a different physical mechanism. The fluid is engineered to boil at 50–60°C, a temperature the processor junction can drive through the board substrate. Chip heat triggers nucleate boiling at the fluid-component interface; vapor rises to a condenser coil at the tank lid, surrenders its latent heat to facility water, condenses back to liquid, and rains onto the hardware below. The entire thermal transport vector is phase change, not sensible heat. The tank is sealed. The fluid vapor pressure must be managed. And because the boiling point is tightly specified, fluid chemistry is the primary design variable — not tank geometry or pump sizing.

Neither architecture is a drop-in retrofit. Both require purpose-built tank infrastructure, modified server configurations (fans removed, hydrophobic coatings on select components), and secondary cooling loops tied to the facility's cooling water plant. The question is not whether to immerse — it is which thermodynamic mechanism to bet the facility on.

Heat-Transfer Physics: Why Two-Phase Looks Better on Paper

The theoretical advantage of two-phase cooling is grounded in a fundamental property of matter: latent heat of vaporization is orders of magnitude larger than sensible heat capacity for the same mass of fluid. When water transitions from liquid to vapor at 100°C, it absorbs approximately 2,260 kJ/kg. A kilogram of water warming from 20°C to 80°C absorbs only 251 kJ. The ratio is roughly 9:1 for water. For the fluorinated fluids historically used in two-phase immersion, the ratio is lower — latent heats in the 100–180 kJ/kg range — but the boiling point is set at 50–60°C, positioning the phase change precisely where processor junction temperatures drive it.

Quantify this against a real workload. Consider a two-phase immersion tank designed for 80 kW of IT load. The condenser coil at the lid transfers heat at near-zero fluid temperature differential — the condensing fluid is always at its boiling point, so the approach temperature to facility water is extremely tight, typically 3–5°C. Fluid circulation is driven by gravity and vapor pressure. Pump parasitic load approaches zero on the fluid side. The tank can be compact because heat flux per unit volume is high.

Run the same 80 kW through a single-phase mineral oil tank. Mineral oil has a specific heat capacity of roughly 1.7–1.9 kJ/(kg·K) and a density around 860 kg/m³. To hold a 25°C temperature rise across the tank at 80 kW, you need a fluid mass flow rate of approximately 1.9 kg/s — about 130 liters per minute of viscous oil through a pump and heat exchanger. The pump power to move that volume against the pressure drop of a densely packed server tank is not trivial; plan for 3–6 kW of parasitic pump load. The tank volume required to maintain adequate dwell time for heat exchange is larger. The approach temperature to the heat exchanger secondary circuit is wider.

The chemistry that enables each approach is distinct. Single-phase fluids must maintain low viscosity across a wide temperature range, exhibit high dielectric strength to prevent electrical discharge, and resist oxidative degradation over multi-year service intervals. Two-phase fluids must be engineered to a precise boiling point, exhibit nucleate boiling behavior at realistic chip heat flux densities (typically 10–50 W/cm²), and recondense efficiently on copper or stainless condenser coils. The fluorinated chemistry that historically achieved this was elegant — but it came from a single dominant supplier whose 2022 announcement reshaped the market.

The Post-Novec Landscape

3M's Novec 649 and Novec 7100 were the defining fluids of commercial two-phase immersion cooling for nearly a decade. Novec 649 — chemically 1,1,1,2,2,3,3,3-octafluoro-4-(trifluoromethyl)pentane — offered a 49°C boiling point, near-zero GWP by fluorocarbon standards, and enough dielectric strength to be safe in open server environments. It anchored the product roadmaps of every major two-phase immersion vendor: Submer, LiquidStack, Green Revolution Cooling (in their two-phase line), and multiple hyperscaler internal development programs.

In December 2022, 3M announced a global exit from all PFAS manufacturing by the end of 2025. The decision was driven by mounting legal liability — 3M had already committed $10.3 billion to settle U.S. public water contamination claims — and by accelerating regulatory pressure from the EPA's PFAS National Primary Drinking Water Regulation and the EU's universal PFAS restriction proposal under REACH, which targets approximately 10,000 compounds. Novec fluids, while positioned by 3M as low-persistence variants, fall within the broad PFAS family and were swept into the exit decision.

The supplier scramble that followed was predictable but not clean. Solvay's Galden PFPE fluids (perfluoropolyethers) are a functional alternative for some applications, with boiling points from 57°C to above 200°C across the product line. AGC Chemicals offers Asahiklin AE-3000 (HFE-based) with a 56°C boiling point, positioned as a direct Novec 7100 analog. Chemours has secondary lines including Opteon products, though primary focus remains on refrigerant replacements. None of these suppliers has reached the volume, pricing, or logistics footprint that 3M maintained at the peak of Novec production.

The impact on existing two-phase deployments is operational. Facilities running Novec-filled tanks face three choices: secure long-term supply contracts at premium spot prices from distributor inventories, accelerate qualification of alternative fluorinated fluids (which requires vendor re-certification and in many cases modified condenser geometry to match alternative fluid condensation curves), or plan an accelerated tank rollover to a new fluid and possibly a new system architecture. None of these options is free. The operators who built two-phase immersion programs on the assumption of stable Novec supply are now managing a forced technology transition on timelines driven by regulatory calendars, not operational ones.

The regulatory calendar beyond Novec is not improving for fluorinated fluids. The EPA's final rule on PFAS reporting under TSCA Section 8, effective 2026, expands disclosure obligations. The EU PFAS restriction dossier, if finalized in its current form, would restrict virtually all fluorinated fluids in new applications within the EU by 2027–2028. Any two-phase fluid roadmap that depends on PFAS-class chemistry carries regulatory tail risk that single-phase hydrocarbon alternatives do not.

Single-Phase Fluid Chemistry

Single-phase immersion fluids divide into three primary chemical families: mineral oils, polyalphaolefins (PAOs), and synthetic esters. Each makes different trade-offs across the five variables that matter operationally: dielectric strength, viscosity profile, oxidative stability, materials compatibility, and end-of-life recyclability.

Mineral oil — specifically white mineral oil or naphthenic mineral oil refined to dielectric grade — is the incumbent single-phase fluid. It is widely available, inexpensive (roughly $3–6 per liter at volume), and well-characterized. Dielectric breakdown voltage for a well-refined mineral oil exceeds 30 kV per millimeter under ASTM D877 test conditions. Electrical conductivity runs below 1 pS/m, satisfying the isolation requirement for direct immersion of energized server boards. The primary liability is viscosity: at 25°C, a typical dielectric mineral oil runs 30–50 cSt. At cold ambient conditions — relevant during tank fill or after extended shutdown — viscosity can climb above 100 cSt, creating pump startup challenges and uneven flow distribution across densely packed server chassis.

Polyalphaolefins synthesized from 1-decene oligomers offer a flatter viscosity-temperature profile. A PAO-4 grade (4 cSt at 100°C) runs approximately 16–20 cSt at 40°C, substantially more pumpable than mineral oil at equivalent temperatures. PAO fluids are inherently more oxidatively stable than mineral oil — the synthetic backbone lacks the aromatic fractions that degrade under thermal stress — and they carry lower pour points, relevant for cold-climate facilities. The cost premium is real: PAO-grade synthetic hydrocarbon dielectric fluids run $12–20 per liter at volume, three to four times mineral oil. For a 40,000-liter tank system (a modest 8-tank immersion pod), that delta is $360,000–$560,000 in fluid cost alone.

Synthetic esters — primarily pentaerythritol esters derived from vegetable-oil fatty acids — occupy a specific niche. Their biodegradability profile makes them attractive in jurisdictions with strict spill regulations, and their higher flash points improve fire rating. However, esters are hygroscopic: they absorb water vapor over time, and dissolved water reduces dielectric strength measurably. Moisture management — sealed headspace, desiccant breathers, periodic dielectric testing — is a non-negotiable operational requirement with ester-based fluids.

For 2026 greenfield immersion deployments, the practical argument for single-phase is supply chain confidence. Mineral oil and PAO dielectric fluids are produced by ExxonMobil, Chevron Phillips, Nynas, and Shell in volumes that dwarf two-phase fluorinated fluid production. Lead times are weeks, not quarters. Specifications are standardized enough that competitive bidding is possible. End-of-life pathways — re-refining into base oil or blending into industrial lubricant streams — exist at scale. A facility engineer specifying single-phase immersion fluid in Q1 2026 can get three competitive quotes from domestic distributors within 48 hours. That is not currently true for two-phase fluorinated alternatives.

Two-Phase Fluid Chemistry

The engineered dielectric fluids suited for two-phase immersion are a narrow chemical family defined by their boiling points, not their brand names. The functional requirement is precise: the fluid must undergo nucleate boiling at the heat flux densities generated by server processors (10–50 W/cm² at the component surface), must condense on a water-cooled coil with minimal approach temperature, and must do both in a sealed environment without degrading materials — PCB substrates, component packaging, elastomer gaskets, metal wetted surfaces — over a multi-year service life.

Historically, Novec 649 (bp 49°C) and Novec 7100 (bp 61°C) satisfied these requirements. Their successors in the post-3M market are imperfect substitutes. Solvay's Galden HT55 (bp 55°C) and HT70 (bp 70°C) are PFPE-class fluids with excellent thermal and chemical stability, but PFPE fluids have higher GWP than Novec variants, their production volumes are constrained to Solvay's capacity, and pricing reflects that scarcity — Galden fluids typically run $150–250 per kilogram at low volume, compared to Novec 649 at approximately $80–100 per kilogram before supply disruption. AGC's Asahiklin AE-3000HT is positioned as a direct HFE replacement with a 56°C boiling point, but qualification data from two-phase system vendors is still accumulating as of early 2026.

Boiling point tuning is the core engineering variable. Chip junction temperatures for current AI accelerators — NVIDIA H100, H200, GB200 — run 80–95°C at TDP. The thermal resistance path from junction through package, board, and fluid boundary layer means the fluid bulk temperature during nucleate boiling must stay below 65°C to maintain adequate temperature differential. A fluid with a boiling point of 50°C provides a 30–45°C margin; a fluid at 60°C provides a tighter 20–35°C margin. Tighter margins reduce the system's tolerance for thermal excursions during transient load spikes. Fluid selection is, in this sense, a reliability specification as much as a cooling specification.

Sealed-tank operation is not optional for two-phase systems; it is inherent to the physics. The fluid vapor must be contained and condensed — not vented to atmosphere, both for economic reasons (fluid is expensive) and for occupational safety (fluorinated vapor at elevated concentration presents inhalation risk). Maintenance access requires vapor recovery equipment: a recovery cart that captures vapor before the tank is opened, condenses it, and stores it while service is performed. This adds capital cost, operational complexity, and a dependency on vendor-specific recovery tooling that has no analog in single-phase operations. Technicians require specific training. Fluid losses through imperfect vapor recovery accumulate cost over the facility lifecycle.

Selection Logic: When Each Architecture Wins

The architecture decision is not a performance optimization — it is a risk portfolio decision. Single-phase and two-phase immersion are not interchangeable; they optimize for different constraints, and the right answer for a 2026 build depends on which constraints dominate the program.

Single-phase wins when procurement certainty is a hard requirement. For greenfield hyperscale builds committing to immersion in 2026 with production timelines 18–36 months out, the single-phase fluid supply chain is the only one that can be contracted with confidence. Mineral oil and PAO dielectric fluids are available from multiple domestic and international suppliers on standard commercial terms. Tank infrastructure is simpler — open-top or lightly covered tanks, standard plumbing, no pressure vessel certification. Service is performed by technicians who can be trained in a day, using PPE available at any industrial safety supplier. The performance ceiling is lower than two-phase on a per-volume basis, but for GPU rack densities below 150 kW per tank — which describes the majority of 2026 production AI inference deployments — single-phase delivers adequate thermal performance at materially lower operational risk.

Two-phase still earns its place in a specific scenario: hyperscale AI training clusters where rack density exceeds 150–200 kW, where floor space is genuinely constrained, and where the operator can absorb the fluid CapEx and operational complexity as a trade against reduced building footprint. These deployments exist — large GPU pod builds for foundation model training, where the economics of dense compute per square foot are compelling enough to justify premium fluid costs and specialized maintenance programs. They are not the typical case.

Direct liquid cooling — specifically cold plate systems circulating water or water-glycol through manifolds attached directly to processor heat spreaders — dominates current high-density AI accelerator deployments. NVIDIA's GB200 NVL72 rack design, the highest-density commercially available AI compute platform at the time of this writing, ships with integrated direct liquid cooling as the primary thermal solution. D2C bypasses the dielectric fluid question entirely: the coolant is water, the supply chain is municipal infrastructure, and the thermal resistance path from junction to coolant is shorter than either immersion approach.

"For most 2026 greenfield deployments, single-phase immersion is the pragmatic answer. The fluid is proven, the supply chain is intact, and the serviceability story is straightforward. Two-phase earns its place in the handful of hyperscale AI training clusters where you are stacking extreme density and every degree of junction temperature margin matters — but that is not most data centers. Direct liquid cooling is still doing the heavy lifting where GPU density is highest, and immersion in either form is complementary to that, not a replacement for it."

— Andre Taki, Director of Products & Sales · Cooling Chemistry Practice Leader, Alliance Chemical

The selection matrix is therefore sequential: evaluate D2C first for AI accelerator workloads. If immersion is the specified architecture, evaluate single-phase unless density requirements and space constraints specifically justify two-phase complexity and cost. If two-phase is specified, plan for Novec-free fluid sourcing from day one — not as a future migration, but as the initial design constraint.

Operational Considerations

Operational planning for immersion — whether single or two-phase — begins with weight. Dielectric mineral oil runs 850–880 kg/m³, or approximately 7.1–7.4 pounds per gallon. A standard single-phase immersion tank sized for 12 servers at 1U–2U density holds approximately 1,000–1,500 liters of fluid. At 880 kg/m³, that is 880–1,320 kg of fluid mass alone, before accounting for the tank structure, servers, and support hardware. A fully loaded single-phase immersion tank can exceed 3,000 kg — 6,600 pounds — per unit. Floor loading specifications for immersion pods require structural engineering review that is categorically different from traditional raised-floor data center planning, which is designed for 1,500–2,500 kg per rack equivalent.

PAO fluids run slightly lighter at 820–840 kg/m³, and most fluorinated two-phase fluids are denser: Novec 649 at 1,600 kg/m³, Galden HT55 at approximately 1,700 kg/m³. A two-phase tank at equivalent volume holds more mass in fluid alone — a counterintuitive result of fluorocarbon density. Tank structural and floor loading calculations must account for this.

Serviceability is a sustained operational cost, not a day-one specification. In single-phase immersion, server removal is mechanically similar to pulling from a conventional rack: lift the chassis vertically from the fluid, allow residual fluid to drain into the tank or a drip tray, and transfer to a maintenance bench. Total hot-swap time for a trained technician runs 15–25 minutes per server including fluid management. Component replacement — DIMMs, NVMe drives, power supplies — is performed on a bench after removal and can be done without specialized tooling or fluid recovery equipment.

In two-phase immersion, tank access begins with vapor recovery. A recovery cart connects to the tank vapor port, draws down the headspace vapor, and stores it in liquid form. This process takes 20–45 minutes depending on tank size and vapor load. Only after vapor recovery is complete can the tank lid be safely opened. Server extraction proceeds similarly to single-phase, but technicians must work within a defined timeframe before vapor accumulation in the headspace requires recovery restart. Total time from work order to component access is typically 60–90 minutes before the first tool touches the hardware. For high-availability workloads where mean time to repair is a contractual SLA, this is a material operational difference.

Retrofit feasibility is a simple answer for both architectures: neither is a retrofit. Immersion cooling requires purpose-built tank infrastructure, modified server configurations, secondary cooling loops tied to facility cooling water, and structural floors rated for fluid weight. The premise that existing air-cooled data halls can be converted to immersion by installing tanks in existing raised-floor space is not operationally realistic. Immersion is a greenfield architecture. The economic case is built on the full lifecycle cost of new construction — not on recapturing depreciated air-cooling infrastructure. Any capital plan that assumes immersion retrofit of existing air-cooled space should be revised before it reaches an investment committee.

Fluid management programs — periodic dielectric testing, viscosity monitoring, particulate analysis, water content measurement for ester-based fluids — are the operational analog of lubricant analysis programs in industrial machinery. Single-phase immersion fluids typically warrant quarterly dielectric testing (ASTM D877), annual viscosity check, and particulate count after major maintenance events. Fluid service life with proper management runs 5–10 years for mineral oil, longer for PAO. Budget for fluid replacement as a capital line item at the 7-year mark in financial models; assume earlier if contamination events occur.

As hyperscale operators formalize immersion procurement programs, fluid specification, supply chain qualification, and operational protocol development belong in the pre-construction planning phase — not as afterthoughts during commissioning. The decisions made in the specification document determine whether a facility's cooling chemistry is an operational asset or a source of unplanned downtime. For a deeper look at the full landscape of cooling chemicals for data center applications — from corrosion inhibitors in D2C secondary loops to glycol blends for free-cooling economizers — see the AI & Data Center Cooling Chemicals pillar at Alliance Chemical: Chemicals for Data Center Cooling. Related sub-pillars cover dielectric fluid specifications for immersion cooling, thermal interface material selection for direct liquid cooling, and cooling water treatment chemistry for hyperscale facilities — each written for the engineers and procurement teams who are building the infrastructure that AI workloads run on.