Two-phase cooling walked into 2026 carrying two reputations at once: the physics that could save a data hall enormous energy, and the operational baggage that keeps chief engineers hesitating before they sign off. Dielectric fluid boiling directly against a chip surface pulls heat away faster than any single-phase loop, and the latent heat released during vaporization dwarfs what convection alone can move. Vendor decks repeat that thermal advantage constantly, so the harder conversation rarely gets equal airtime. Once fluid turns to vapor, it has to go somewhere, and where it goes determines whether a facility runs quietly or fights pressure swings for years. Fluid chemistry has also shifted underneath much of the industry, as 3M’s exit from PFAS manufacturing prompted many operators using PFAS-based dielectric fluids to requalify coolant supply chains and evaluate alternative chemistries.

Boiling Is the Easy Part — Containing the Vapor Is the Real Job

Engineers understand nucleate boiling reasonably well, and a single immersed chip on a lab bench demonstrates the effect predictably within a few degrees. The real engineering challenge starts the moment vapor bubbles detach from the board and rise through the fluid column toward open headspace. Rise rate depends on bubble size, fluid viscosity, and rack density, and none of those variables stay constant as workload shifts from idle to full inference load. A tank with insufficient headspace lets vapor reach the lid before it fully condenses, and that mismatch shows up as pressure spikes rather than smooth heat rejection. Pressure stability, not boiling point alone, decides whether the condenser coil keeps pace with vapor generation during a sudden compute burst. Tank geometry, condenser placement, and headspace volume together shape steady performance far more than the specific fluid poured into the bath.

Some condenser sizing approaches begin with average rack wattage, although robust designs also account for transient peak heat loads that place the greatest stress on vapor containment systems. GPU thermal design power has climbed sharply across recent hardware generations, with newer accelerators drawing well over a kilowatt each under sustained load. When several boards ramp together, vapor generation can outpace the condenser’s instantaneous capacity even though the daily average stays within design limits. That gap forces excess vapor into the headspace, raises internal pressure, and can push a relief valve to release working fluid straight into the room. Condenser sizing based solely on nameplate average heat loads can lead to insufficient vapor-handling capacity under transient AI workloads, with operational issues sometimes becoming apparent only after sustained production use.

Fluid Loss Without a Leak: The Carryover Problem

No cracked seal or dripping fitting needs to exist for a two-phase tank to lose fluid steadily across its operating life. Every lid lift for a routine service call exposes the headspace to room air, and whatever vapor sits above the fluid line escapes before the lid closes again. Technicians pulling a failed drive or swapping a network card rarely spend more than a few minutes with the tank open, and that short window still allows meaningful vapor carryover. Published industry case studies and immersion cooling vendors have reported annual fluid loss in the range of approximately one to two percent of tank volume under normal operating conditions, although actual rates vary with fluid chemistry, enclosure design, maintenance practices, and operating conditions. Poorly designed enclosures, with wide lid openings or condensers positioned far from the fluid surface, run considerably higher than that baseline figure.

During heavy compute load, internal pressure rises and a relief valve opens to prevent over-pressurization, venting a mix of vapor and air into the surrounding room. Two-phase tanks also lose fluid through a quieter mechanism tied to pressure cycling rather than physical access. During idle periods, condensation drops internal pressure below ambient, and the same valve opens in reverse to let outside air back in. This cyclical exchange between tank and room repeats many times a day across a large deployment, and each cycle carries away a small quantity of working fluid. Operations teams tracking replenishment volume against compute utilization typically find loss rates track workload variability more closely than raw uptime hours. Building a recovery loop that captures escaping fluid before it reaches the relief valve turns this compounding cost into a manageable, budgetable line item instead of a recurring surprise.

Altitude Rewrites Your Boiling Point

Saturation temperature depends on ambient pressure, and ambient pressure drops steadily as elevation increases above sea level. Water boils at 100 degrees Celsius at sea level, but the same water boils near 95 degrees Celsius at Denver’s elevation of roughly 1,600 meters. Dielectric fluids used in two-phase tanks follow the identical physical principle, so their saturation point shifts downward at any site built above sea level. A tank engineered around a fluid’s sea-level boiling point behaves differently once installed in Denver, Mexico City, or Johannesburg, all of which sit well above the elevations most cooling specs assume. Mexico City sits near 2,240 meters and Johannesburg near 1,750 meters, and both cities host growing data center markets that many fluid datasheets never anticipated.

Lower saturation temperature at altitude changes more than the number printed on a spec sheet, because it directly affects how much energy the condenser needs to remove per unit of vapor. Vapor density falls as saturation temperature drops, and lower vapor density reduces the heat transfer coefficient available during condensation at a given surface area. That reduction means the same condenser performing well at sea level may need larger surface area or higher airflow to hit equivalent thermal targets at elevation. Site teams using condenser designs developed solely from sea-level reference data should evaluate altitude effects during engineering, as higher elevations can influence thermal performance depending on system design and operating conditions. Altitude derating tables already exist for standard HVAC equipment, and two-phase condenser sizing needs the same category of adjustment before a tank design gets finalized.

When One Rack Boils Harder Than the Next: Cascade Boiling Inside Shared Tanks

Not every board in a shared immersion tank generates heat at the same rate, and that difference in heat flux creates uneven boiling patterns across the bath. A GPU-dense board running a full inference workload boils far more aggressively than a lightly loaded storage node sitting a few slots away in the same fluid volume. Bubbles rising from the hotter board pass upward through the fluid column and can interfere with heat transfer on components positioned above or nearby. Research on rising bubble interference has documented measurable cooling degradation on upper components, even when those components carry a lower thermal load themselves. This preferential boiling effect means neighboring hardware absorbs a performance penalty caused entirely by tank layout rather than its own workload. Component placement inside a shared tank becomes a genuine thermal design decision, not simply a mechanical or space-driven one.

When heat flux from one board climbs high enough, it risks crossing critical heat flux, the threshold where nucleate boiling collapses into an insulating vapor film against the surface. That vapor film blocks direct liquid contact, drops the heat transfer coefficient sharply, and can send local temperatures climbing within seconds rather than minutes. Dryout on one board rarely stays isolated, because the vapor plume it generates changes local fluid circulation and temperature gradients for everything sharing that tank. Operators running mixed hardware generations in a single bath face a harder version of this problem, since newer accelerators push heat flux well beyond older board designs. Tank zoning, grouping similar heat-flux hardware together and separating high-density boards from sensitive neighbors, reduces the odds of one board’s boiling crisis degrading another’s performance.

Two-Phase Doesn’t Lose on Thermal. It Loses on Control

Every argument against two-phase cooling eventually traces back to control, not thermodynamics, once the surface-level objections get stripped away. The latent heat advantage is real, the heat transfer coefficients are real, and no single-phase system matches the raw thermal ceiling two-phase boiling can reach. What separates a successful deployment from a troubled one is whether pressure stays predictable, whether vapor stays contained during service, and whether condenser sizing accounts for actual site conditions rather than generic averages. Facilities that incorporate containment, fluid-loss management, and altitude considerations into system design are generally better positioned to reduce operational risks than designs focused primarily on fluid selection and tank sizing alone. The PFAS supply shock added another layer of complexity, forcing operators to requalify fluid chemistry while they were still learning to manage vapor behavior day to day.