Two-phase cooling field guide for AI data centers

Two-phase cooling, and why boiling moves so much heat

Two-phase cooling removes heat by boiling a dielectric fluid right where the heat is made and condensing the vapor somewhere it can give the heat back up. The fluid changes state, liquid to vapor at the hot surface and vapor to liquid at a cooler one, and that change of state is the whole trick. A fluid soaks up a large amount of heat to boil, far more than it soaks up just warming a few degrees, so a small flow of a boiling fluid carries heat that a much larger flow of warm liquid would struggle to move.

That is the appeal at the extreme end of AI and HPC density. When a single accelerator concentrates more than a kilowatt into a package the size of a deck of cards, and a rack pulls past 120 kW, you want the densest possible way to lift heat off the silicon. Boiling a fluid at the surface is about as dense as it gets.

It is not a free win, and that is the honest framing this guide keeps. The fluids that boil where you need them have been costly, they ask for sealed systems and careful containment, and the ones that worked best raise a serious regulatory question about PFAS. So two-phase sits alongside single-phase liquid cooling, not automatically ahead of it. The single-phase direct-to-chip guide and the single-phase immersion acceptance guide cover the more common choices, and most large jobs today are still making one of those.

Latent heat versus sensible heat: the physics that does the work

The reason two-phase moves so much heat comes down to latent heat versus sensible heat, and it is worth getting straight because everything else follows from it. Sensible heat is the heat that changes a fluid's temperature. Warm a liquid one degree and it absorbs a certain amount of energy, its specific heat. Single-phase cooling lives entirely on sensible heat: the coolant goes in cool, comes out warm, and the temperature rise is the heat it carried.

Latent heat is the heat a fluid absorbs to change state without changing temperature. Boil a liquid and it takes in a large slug of energy to break into vapor, all at the boiling point, with no rise in temperature while it does. For a typical engineered dielectric the latent heat of vaporization can be on the order of a hundred times the heat the same fluid picks up warming a single degree. The exact ratio belongs to the fluid's data sheet, so treat that as the shape of it, not a number to quote, but the gap is large no matter which fluid you pick.

Two practical consequences fall out. You move the same heat with far less fluid flow, because each kilogram carries so much more. And the chip sits at a more stable temperature, because the fluid pulls heat at its boiling point instead of climbing as it goes. That second one matters for hardware that hates a moving junction temperature. The catch is that you are now managing a boiling fluid and its vapor, which is harder than pushing warm water, and the rest of this guide is mostly about that cost.

What is the difference between single-phase and two-phase cooling?

Single-phase and two-phase cooling differ in whether the coolant changes state. In single-phase, the liquid stays liquid the whole way through. It flows past the heat, warms up, carries that heat away as a temperature rise, and gets cooled back down at a heat exchanger. Nothing boils. This is the mature, default form of liquid cooling, whether it is a cold plate on a chip or a tank full of oil, and it is what most deployments run.

In two-phase, the liquid boils to vapor at the hot surface and condenses back to liquid at a cooler one. The heat rides in the phase change, the latent heat, rather than in a temperature rise. That single difference cascades into everything: the fluid has to be one that boils at a useful temperature, the system has to manage vapor and pressure, it has to be sealed against vapor loss, and servicing it is harder. The single-phase direct-to-chip guide and the single-phase immersion guide cover those mature forms in depth.

Do not assume a liquid-cooled deployment is one or the other. A cold-plate rack can be single-phase or two-phase, and so can an immersion tank, and they look similar from across the room. Confirm which it is from the equipment submittal before you plan anything, because the coolant, the pressures, the leak behavior, the containment, and the service procedure all change with the answer.

The two forms: two-phase direct-to-chip and two-phase immersion

Two-phase cooling shows up in two shapes, and they map onto the two single-phase forms its sibling guides cover. Two-phase direct-to-chip puts an evaporator cold plate on the processor and lets the fluid boil inside that plate. Two-phase immersion drops the whole server into a sealed tank and lets the fluid boil off the hardware directly. Both use latent heat. They differ in where the boiling happens and how much of the server the fluid touches.

Direct-to-chip keeps a mostly conventional server. The plate boils the fluid on the hot chips, the vapor leaves through a line, and air or a second loop still handles the memory, drives, and power components the plate does not reach. Two-phase immersion is the opposite extreme: the fluid contacts everything, boils across the whole board, and the vapor rises to a condenser in the top of the sealed tank.

The rest of this guide treats the shared pieces, the condenser, the fluid, the PFAS question, the sealing and serviceability, once, and notes where the two forms diverge. The two sections that follow walk each form on its own first.

Two-phase direct-to-chip: the evaporator cold plate and the vapor line

Two-phase direct-to-chip mounts a cold plate on the chip the same way single-phase does, but here the plate is an evaporator: the fluid enters as liquid, boils inside the plate's channels against the hot die, and leaves as vapor or a vapor-and-liquid mix. Because the fluid carries heat as latent heat, the flow through the plate is small for the heat it removes, and the chip sits at a tighter temperature than a single-phase plate holds.

The fluid is a dielectric or a refrigerant chosen to boil at a useful temperature. Designs in this space use lower-GWP refrigerants such as R-1233zd(E) and R-1336mzz(Z), both near a GWP of one, and the moderate-GWP R-515B at roughly 290, often in a pumped arrangement where a pump moves liquid out to the plates and the vapor returns to a coolant distribution unit that condenses it. Vendors report pumped two-phase systems carrying well past 160 kW per rack at high efficiency, with the fluid and the design setting the real limit, so hedge those figures to the manufacturer.

The vapor side is the part that is genuinely different from single-phase, and it drives the plumbing. Vapor takes far more volume than the liquid it came from, so the return piping runs much larger than the supply, by several times for the low-pressure refrigerants, and the system has to manage pressure and any non-condensable gas that creeps in. None of that exists on a single-phase loop. Read the single-phase direct-to-chip guide for the cold-plate, manifold, and quick-disconnect fundamentals the two share, then layer the vapor management on top.

Two-phase immersion: boiling off the hardware in a sealed tank

Two-phase immersion submerges the server in a low-boiling dielectric fluid and lets the fluid boil directly off the hot components. There is no cold plate and no pump driving the boil. Heat off a processor turns the fluid to vapor right at the surface, the vapor rises through the bath, and it collects in the vapor space above the liquid.

A condenser coil sits in the top of the sealed tank, fed with facility water. The vapor hits that cooler coil, gives up its latent heat, condenses back to liquid, and rains back down into the bath to boil again. The cycle is passive in the sense that the boil and the rise are driven by the heat itself, which is part of the elegance: a tank of boiling fluid with a condenser lid and no fans is a quiet, dense way to cool. The trade is that the tank has to be sealed to keep the vapor in, because the vapor is expensive and, in the fluids that worked best, an environmental concern.

That sealing is the line between the two immersion families. A single-phase immersion tank can run open or loosely covered, because nothing is evaporating, and the single-phase immersion acceptance guide covers it in full. A two-phase tank cannot. The sealed lid, the vapor space, and the condenser are the system, and you cannot just lift the cover to pull a server. Confirm which immersion type a job is before you write a procedure, because almost everything about handling and service changes between them.

The condenser: where the vapor gives the heat back

The condenser is where two-phase cooling closes the loop, and it is the component single-phase does not have. Its job is to take the vapor the chip made, cool it below its boiling point, and turn it back to liquid so it can go boil again. In two-phase immersion the condenser is a coil in the top of the sealed tank. In two-phase direct-to-chip it usually lives in the coolant distribution unit, a heat exchanger that condenses the returning vapor against facility water.

Whatever its shape, the condenser rejects the heat to a facility water loop, the same downstream path single-phase liquid cooling uses. The vapor condenses against the cooler water, the heat crosses into the facility loop, and the building rejects it at a tower, a dry cooler, or a chiller. The two-loop separation is the same idea the single-phase guides describe: the fluid the hardware sees never mixes with the building water.

Two things make the condenser its own discipline. It has to be sized and fed so it actually condenses all the vapor the load produces, because vapor that does not condense raises pressure and backs up the boil. And the seal around it matters, since this is a boundary where vapor can escape and where non-condensable gas can leak in and blanket the coil, which quietly kills its performance. Vapor management, keeping the right amount of vapor, condensing it fully, and keeping air out, is most of what running a two-phase system well comes down to.

The engineered dielectric fluid

The fluid is the heart of two-phase cooling, and it is an engineered dielectric, not something you choose on the jobsite. It has to be non-conductive, so a boil right on a live board does not short anything, and it has to boil at a temperature that keeps the chip in band, commonly somewhere in the rough range of 30 to 60 C depending on the fluid and the design. That boiling point is tuned to the hardware: too low and it boils off too easily and pressurizes, too high and it does not boil at the chip temperature you want to hold.

It also has to be stable for years against hot silicon, compatible with everything it touches, and clean enough not to foul. Those demands narrow the field hard, which is why the fluids that worked best for two-phase were specialty fluorinated liquids. They were also expensive, and the cost is not a footnote when a tank or a loop holds a large standing inventory of the stuff.

Every real number here belongs to the fluid manufacturer's data sheet: the boiling point, the latent heat, the dielectric strength, the density, the compatibility list. Do not carry them in your head or quote a generic figure on a submittal. Confirm the delivered fluid matches the specified fluid by name and grade, and pull a baseline sample so there is a documented starting point to trend against, the same discipline the immersion acceptance guide spells out. The fluid is the single thing that defines a two-phase system, and it is also the thing the next section says you have to think hardest about before you commit.

Are two-phase cooling fluids PFAS?

Yes, the fluids that made two-phase cooling work have been PFAS, and this is the central question hanging over the whole technology. The low-boiling dielectrics two-phase relied on, the engineered fluorocarbons sold under names like Novec and Fluorinert, along with the PFPE and HFE and FK-5-1-12 chemistries used elsewhere, are per- and polyfluoroalkyl substances. They are often called forever chemicals because the carbon-fluorine bond does not break down in the environment. That same bond is part of what made them good coolants and a long-term liability at the same time.

The supply side already moved. 3M, the dominant maker, announced in December 2022 that it would exit all PFAS manufacturing by the end of 2025, and industry reporting put the last orders for the Novec fluids around the end of March 2025 with production winding down after. That decision came alongside thousands of lawsuits and a multi-billion-dollar settlement with public water systems over PFAS contamination, and on top of tightening EPA and international rules. Major hyperscalers pulled back from two-phase immersion research as a direct result. Interim and reportedly PFAS-free replacement fluids have been discussed, but whether a clean fluid that boils where you need it, stays non-conductive, and is not a regulated PFAS has truly arrived at scale is unsettled as of this 2026 review.

So be blunt with an owner about this. A two-phase design ties you to a fluid class under active regulatory and supply pressure, and a system you cannot get fluid for in five years is a stranded asset. There is no blanket ban in force everywhere as of this writing, and the rules differ by jurisdiction and keep moving, so verify the current PFAS regulations where the project sits and the actual availability of the specified fluid before the design is locked. This is exactly the kind of question that should be answered before the capital is committed, not discovered after the tank is filled.

Fluid cost, vapor loss, and the inventory you carry

The fluid is expensive, and in a two-phase system it is also a thing you can lose to the air, which makes the economics harder than single-phase. A tank or a loop holds a large standing inventory, and that inventory is a real line item on the capital cost, not a consumable you top off cheaply. The PFAS supply squeeze in the section above has pushed those costs the wrong way and made the fluid harder to source at all.

Vapor loss is the running cost. Any vapor that escapes a seal, a fitting, a service opening, or a poorly condensing condenser is fluid gone, and unlike a water leak you may not see it pool. It leaves as vapor and the inventory drops over time. That is why the seal integrity and the vapor management matter to the budget, not just to performance: a system that weeps vapor is a system quietly spending money and, with a fluorinated fluid, venting a regulated substance.

Plan for makeup fluid and for an inventory you track. Budget the initial fill at the fluid's real price, plan a makeup allowance for the loss the design will have, and meter the inventory so a falling level shows up as a number, not a surprise at the next sample. The fluid cost is one of the reasons many teams weighing two-phase against single-phase land on single-phase: water or a water-glycol mix is cheap, and you do not lose it to the air.

Sealing and containment for vapor

Two-phase systems have to be sealed against vapor, and that is a higher bar than the leak-tightness a single-phase loop needs. A single-phase loop has to hold liquid. A two-phase system has to hold vapor, keep its operating pressure, and keep outside air from leaking in, because air that gets in becomes non-condensable gas that blankets the condenser and degrades it. Every seam, lid seal, fitting, and penetration is a place vapor can leave or air can enter.

Containment is the second layer, and it follows from the fluid being valuable and, in the fluorinated case, regulated. A two-phase immersion tank carries spill containment sized to its fluid volume the same way a single-phase tank does, but it also has to contain vapor under the sealed lid and manage the vapor space, which a single-phase tank never deals with. A two-phase direct-to-chip loop carries the vapor return piping, the pressure, and leak detection at the points vapor and liquid can escape.

The honest framing is that sealing is not a detail you finish at the end. It is a property of the whole system that has to be designed in, proven at commissioning, and maintained, because a slow vapor leak does not announce itself. It shows up months later as a dropping fluid level and a condenser that is not keeping up, and by then you have spent fluid and money chasing a problem that a tighter seal would have prevented.

Serviceability of a sealed system

Servicing a two-phase system is harder than servicing air or single-phase, and underrating that is a common planning mistake. The system is sealed and holds a boiling fluid under its own vapor pressure, so you cannot just pop it open and pull a part. Opening a two-phase immersion tank means managing the vapor space and the seal, letting the system come to a state where it is safe to open, and recovering or containing the vapor rather than letting it escape. A two-phase direct-to-chip node may need the loop brought to a condition where a connection can be broken without losing fluid or admitting air.

Compare that to the single-phase forms. A single-phase direct-to-chip server pulls on dripless quick-disconnects in a minute. A single-phase immersion node lifts out of an open bath and drips back on a rail. A two-phase system asks for more: vapor handling, fluid recovery, and a procedure that protects both the expensive fluid and the people near it. That is slower, and it needs the right equipment and training to do safely.

Plan the service reality into the design, not into the first outage. Confirm there is a defined procedure for opening the system, recovering vapor, and reclosing it to seal, and that the access and equipment to do it are actually there. A sealed system that is a misery to service gets serviced badly, and badly servicing a vapor system near energized hardware is exactly the risk you want designed out.

Why two-phase cooling is used for AI

Two-phase cooling is aimed at the highest heat densities, which is why it comes up in AI and HPC and almost nowhere else. The driver is heat flux at the package. A modern AI accelerator concentrates well over a kilowatt into a tiny die, and a dense rack pulls past 100 to 120 kW, with reference platforms landing around 120 kW in a single rack. Air ran out long ago, and even single-phase liquid has a ceiling per plate as the watts climb.

Boiling a fluid right at the die is about the densest heat removal there is, because the phase change pulls a large amount of heat off a small, hot area while holding the surface near the fluid's boiling point. That is the case for two-phase: when the chip is dumping more heat into a smaller spot than single-phase can comfortably hold, the latent-heat approach has headroom that warming a liquid does not. Vendors point to pumped two-phase racks carrying well past 160 kW as evidence of where it reaches.

Be careful with the timing, though, because this is where hype gets ahead of the field. As of this 2026 review, current top-end accelerators are still being cooled with single-phase direct-to-chip at warm supply temperatures, so the densities that strictly require two-phase are largely ahead of us rather than here. Hedge the where-it-shines case to the specific chip and platform. Two-phase is the technology to watch as thermal design power keeps climbing, not a box every AI hall needs to check today.

Efficiency and the PUE case

Two-phase cooling has a real efficiency case, and it comes from the same physics that moves the heat. Because the fluid boils at the chip, it holds a small temperature difference and lets the system reject heat at a warm condensing temperature, which means the facility loop can run warm and lean on free cooling for much of the year instead of mechanical chillers. The pumping energy is low too, since latent heat moves the load with little flow, so you spend far less moving heat than air ever did with fans.

Vendors cite large cooling-energy reductions for pumped two-phase, in some cases reductions on the order of 80 percent against legacy air, and correspondingly low PUE. Treat those as design targets tied to a specific site and a specific product, not numbers you inherit, because the real result depends on the climate, the supply temperature, and how much of the year free cooling is available. The single-phase guides make the same point about warm-water operation, and the efficiency advantage of liquid over air is shared across both, not unique to two-phase.

The honest read is that two-phase can be the most efficient of the options at the extreme densities, but the gap over a well-run single-phase warm-water system is not always large enough to outweigh the fluid cost, the sealing, and the PFAS risk. The efficiency comes from running warm and rejecting heat the cheap way, and a single-phase system designed to do the same captures most of it.

Maturity and adoption

Two-phase cooling is the less mature of the liquid options, and that is the plain status as of this 2026 review. Single-phase direct-to-chip is the mainstream choice for AI racks and single-phase immersion is the more deployed of the two immersion families. Two-phase, in both its cold-plate and immersion forms, is less proven at scale, with fewer large production deployments and a thinner operating track record behind it.

The PFAS overhang is a large part of why. The fluid supply that two-phase immersion was built around contracted hard, hyperscalers pulled back, and that slowed the whole category while the industry waits to see whether a clean replacement fluid and stable regulation arrive. Two-phase direct-to-chip with low-GWP refrigerants is a more active development path, with reference work from the Open Compute Project and several vendors, but it is still emerging rather than the default. The vendors are real and the engineering is real; the question is scale and longevity.

What that means for a project is restraint. Treat a two-phase proposal as the newer, less-proven option, weigh the vendor's track record and the fluid's future honestly, and do not let a strong efficiency pitch paper over the maturity gap. The technology may well grow as chip power climbs past what single-phase holds, but that is a bet on the future, and the next section is about how to make that call today.

When two-phase versus single-phase?

The choice between two-phase and single-phase is a density-versus-risk call, and for most jobs today single-phase wins it. Single-phase direct-to-chip and single-phase immersion are mature, run on cheap or well-understood fluids, are easier to seal and service, and carry no PFAS exposure on the single-phase hydrocarbon and water-glycol fluids. They cover the densities the current generation of AI hardware actually runs, which is why the bulk of new large deployments are single-phase.

Two-phase earns a look when the heat flux genuinely exceeds what single-phase can hold at the plate, when the efficiency at warm rejection is worth the complexity, and when the fluid question can be answered with confidence for the life of the asset. That is a narrower case than the marketing suggests. It is most defensible at the leading edge of density, with a vendor that has a real track record and a fluid with a clear regulatory and supply future.

The framing to give an owner is straightforward. Single-phase is the lower-risk default that covers most needs now; two-phase is the higher-density, higher-complexity option to consider when the density forces it and the fluid risk is acceptable. The direct-to-chip and immersion guides cover the single-phase forms in full. Choose against the building, the density, and the fluid's future, not against a preference for the newest thing.

The facility side behind the loop

Whatever happens at the chip, the heat still has to leave the building, and the facility side of a two-phase system is the same job single-phase faces. The condenser, whether a coil in a tank or a heat exchanger in a coolant distribution unit, rejects the heat into a facility water loop, and that loop ends at a cooling tower, a dry cooler, or a chiller depending on the climate and the supply temperature the design targets. The fluid the hardware sees never mixes with the building water.

The warm-rejection advantage carries over and is arguably stronger here. Because the condensing temperature can be warm, the facility loop often runs warm enough to reject heat with free cooling for much of the year, skipping chiller compressor energy. ASHRAE TC 9.9 liquid-cooling guidance describes the warmer supply ranges these systems target, and the warmer the loop the more free-cooling hours the site earns. The cooling guides cover the plant, the tower, and the chilled-water architecture in depth, since that downstream side is shared with single-phase.

Commission the facility side alongside the cooling system, because a condenser starved of facility water cannot reject its load no matter how well the chip side boils. The point to carry is the same one the single-phase guides make: the loop at the chip does not work without the loop in the yard, and the two get proven together.

Material and seal compatibility

Every material the fluid touches has to be compatible with that fluid, and a two-phase system has the added wrinkle that vapor reaches places liquid might not. The wetted and vapor-exposed set is the cold plates or tank, the seals and gaskets, the hoses and fittings, the condenser, and the pump components where there is a pump. A dielectric fluid is a solvent, and over years it can swell or soften the wrong elastomer, lift an adhesive, or extract a plasticizer, and the seal that fails is the one letting vapor out.

The compatibility list belongs to the fluid and the equipment vendors, and it is specific to the chemistry. A seal material that is fine in one fluid can degrade in another, so the seals, the gaskets, and the wetted metals have to be chosen against the actual fluid, not a generic chart. Optical components and labels are known watch items in immersion, since fluid can wick where you do not want it. On the IT side, two-phase immersion needs the same server preparation single-phase immersion does, replacing standard thermal grease and removing parts that do not belong in fluid.

This ties straight to the warranty. The chip vendor, the cooling-system manufacturer, and the fluid maker each specify what materials and fluids are approved, and substituting an off-list hose or seal because it was on the shelf can both seed a failure and void the support you paid for. Honor the approved-materials and approved-fluid lists together, because a single incompatible seal is enough to start a vapor leak across the system.

Fluid safety, vapor, and decomposition

Two-phase fluids carry safety considerations that single-phase liquids mostly do not, and they center on the vapor. The dielectric fluids are generally non-flammable and electronics-safe by design, which is part of their appeal, but vapor is still vapor. A heavy vapor can displace air in a low or enclosed space, so ventilation and the room treatment matter, and the safety data sheet sets the handling, the personal protection, and the exposure limits. Read the SDS for the specific fluid and follow it, because the guidance is fluid-specific and not something to generalize.

The decomposition products are the part worth being blunt about. Fluorinated fluids are stable in normal operation, but heated far past their working range, on the order of several hundred degrees C as you would see in a fire or a severe fault, they can break down into hydrogen fluoride and other halogenated products that are hazardous to breathe. That is not an operating-temperature concern, since the fluid boils far below that, but it is a fire and severe-fault concern that belongs in the life-safety review and the emergency procedure.

Handle it the way the SDS and the manufacturer say. Ventilate where vapor can collect, protect skin and eyes per the data sheet, plan the fire and fault response around the decomposition risk, and train the people who service the system on the vapor handling. The fluid is safe to run and dangerous to abuse, and the procedures are what keep it on the right side of that line.

Commissioning and acceptance

A two-phase system gets proven before it carries real load, and the sequence layers vapor concerns on top of the usual liquid-cooling checks. Confirm the fluid identity against the spec and pull a baseline sample. Leak-test the system, which for two-phase means proving it holds vapor and pressure and keeps air out, not just that it holds liquid. Fill to the right level with the vapor space and any expansion volume the manufacturer specifies. Prove the condenser actually condenses the full load and that the facility-water interface delivers design flow and temperature.

The vapor space and the seal are the acceptance points that do not exist on a single-phase job. Confirm the system seals, that non-condensable gas is purged or within the design limit, and that the pressure holds where it should under load. Then run the system under real or simulated heat load through the failure cases, the same integrated-test discipline the immersion acceptance guide details, so a system that holds on a calm afternoon is proven against the day something breaks.

Because two-phase immersion is an immersion tank at heart, most of the acceptance discipline is shared with the single-phase immersion acceptance guide: the floor load under a filled tank, the fluid baseline, the leak integrity, the heat rejection, the fire-code position from the authority having jurisdiction. Read that guide for the tank-acceptance sequence in full, then add the vapor-space, sealing, and condenser proofs that are specific to two-phase. Do not write a two-phase procedure from the single-phase one without those additions.

Records: fluid inventory, leaks, and the acceptance baseline

A two-phase system that is not documented hands the next operator a sealed box they cannot run with confidence, and the fluid is the thing they most need a record of. Capture the fluid identity and grade, the baseline chemistry, the fill volume, and the as-commissioned readings, because the fluid is an inventory and a maintenance item the owner inherits, not a fill-and-forget. A falling inventory is the first sign of a vapor leak, and you can only see it falling if someone wrote down where it started.

Keep the record live, not just a turnover document. Track the fluid inventory over time, log every leak and the vapor recovered during service, and trend the condenser approach and the fluid chemistry against the acceptance baseline. A field tool like FieldOS is a reasonable place to hold the fluid inventory, the leak log, the sample results, and the acceptance readings so they live with the asset instead of in a binder nobody opens. The table below is the shape of what to capture; the real values come from the fluid data sheet and the equipment submittal.

Common mistakes

• Treating a two-phase system like a single-phase one, and missing the vapor, pressure, and sealing it demands.
• Ignoring the PFAS and regulatory risk on the fluid, and locking a design to a fluid class that may not be available for the asset's life.
• Vapor leaks and fluid loss from poor sealing that drain an expensive inventory quietly over months.
• Not planning the serviceability of a sealed system, so opening it for service becomes an improvised, unsafe event.
• Material incompatibility, a seal or gasket the fluid degrades, that starts a vapor leak no one designed for.
• Underestimating the fluid cost, both the initial inventory and the makeup for vapor lost in operation.

Field checklist

Standards and references

Two-phase cooling sits inside a thin but growing reference set, and the cooling-system manufacturer's design is usually the controlling document. ASHRAE Technical Committee 9.9 publishes the thermal guidelines and liquid-cooling guidance that set the facility-water supply ranges and the water-quality and compatibility framework the heat-rejection side is built around, the same guidance the single-phase liquid guides draw on. Treat the specific class temperatures as the published ranges and confirm them against the current edition.

The Open Compute Project publishes liquid-cooling and immersion work that much of the industry aligns to, including reference material on two-phase refrigerant-based direct-to-chip and on immersion fluids and tanks. Name the OCP document by topic and confirm the current revision, since the documents revise on their own cycle. The cooling-system and fluid vendors publish the fluid spec, the boiling point and latent heat, the compatibility list, and the supply conditions their equipment needs, and where the immersion form is used, the fire and life-safety path runs through NFPA 75 and NFPA 30 as the single-phase immersion acceptance guide describes.

The fluid's regulatory status is its own reference, and it is live. The PFAS rules under the EPA's TSCA and the international equivalents are tightening and the reporting timelines have moved, so the regulatory status of any fluorinated fluid has to be confirmed against current regulation in the project's jurisdiction, not a remembered snapshot. Hedge the boiling points, densities, and latent-heat figures to the fluid manufacturer, the system to the cooling vendor, and the PFAS status to current regulation. Three things hold across all of it: two-phase uses latent heat to reach the highest densities, the dielectric fluid and its PFAS question are central to the decision, and the system has to be sealed, contained, and planned for service.

Units and terms

Two-phase cooling borrows vocabulary from refrigeration, the fluids industry, and the data center side, so the same idea can read differently across a vendor sheet, a fluid data sheet, and a spec. The terms below are the ones that cause the most confusion on a first job.

Two-phase cooling

Cooling that removes heat by boiling a dielectric fluid at the heat source and condensing the vapor, carrying heat as latent heat

Single-phase cooling

Cooling where the liquid stays liquid and carries heat as a temperature rise (sensible heat), the mature default

Latent heat

The heat a fluid absorbs to change state from liquid to vapor with no change in temperature; far larger than sensible heat per unit of fluid

Sensible heat

The heat that changes a fluid's temperature; how single-phase cooling carries its load

Dielectric fluid

A non-conductive engineered fluid the hardware can boil safely; its boiling point is tuned to the chip

Evaporator

The cold plate or surface where the fluid boils and takes heat off the chip

Condenser

The coil or heat exchanger where the vapor gives up its latent heat and condenses back to liquid, rejecting heat to facility water

Boiling point

The temperature at which the fluid changes to vapor; chosen to hold the chip in band, per the fluid data sheet

PFAS

Per- and polyfluoroalkyl substances, the fluorinated chemistry behind most two-phase fluids, under heavy regulatory and supply pressure

Non-condensable gas

Air or other gas that leaks into a two-phase system and blankets the condenser, degrading heat rejection

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is direct-to-chip liquid cooling?
• What is immersion cooling for data centers?
• What is in-row cooling?
• What is direct-to-chip liquid cooling?
• What is an as-built drawing?