2028: The Two-Phase Turning Point? Why AI Data Center Cooling May Be Ready for Its Next Shift
Today, the liquid-cooling story is largely a single-phase story.
Water or water-glycol moves through cold plates attached to GPUs and CPUs, carries heat through hoses and rack manifolds, and transfers that heat through a CDU into the facility cooling system.
For current AI infrastructure, this architecture makes sense.
It is understandable.
It is serviceable.
It fits relatively well into the rack and facility infrastructure that the data-center industry already knows.
But AI power density is not standing still.
And that raises an increasingly important question:
What happens when single-phase DLC itself begins to approach its practical limits?
The Thermal Roadmap Is Moving Faster Than the Cooling Roadmap
An OCP educational webinar presented in May 2026 by IDTechEx and ZutaCore illustrates the problem clearly.
Its thermal-density roadmap places the GB200-era rack around the 130 kW level, followed by substantially higher rack densities for Rubin-class systems and a projected transition toward roughly 600 kW-class and eventually 900 kW-class rack architectures later in the decade.
The exact future products and power levels will continue to evolve.
But the direction is difficult to miss:
Compute density is increasing much faster than traditional cooling infrastructure was originally designed to accommodate.
The same presentation describes a widening “thermal gap” in which heat generation begins to outpace conventional heat-removal capability.
Single-Phase DLC Is the Right Solution Today
This should not be misunderstood as an argument against single-phase liquid cooling.
Quite the opposite.
Single-phase direct-to-chip cooling is likely to remain the dominant transition technology for the current generation of AI data centers.
The architecture is already becoming familiar:
Cold Plate → Hose → Quick Disconnect → Rack Manifold → CDU → Facility Water
It offers a practical way to move a large portion of server heat directly into liquid while maintaining conventional rack-based service models.
The OCP webinar example shows DLC removing approximately 90 percent of the thermal load from a high-density GPU server.
That is an enormous improvement over trying to move all that heat through air.
But Single Phase Has a Scaling Problem
Single-phase cooling removes heat by increasing the temperature of a liquid.
As the heat load rises, designers generally need some combination of:
- higher liquid flow,
- larger channels,
- greater pump capacity,
- larger manifolds,
- more pressure-drop management,
- and increasingly sophisticated flow balancing.
The OCP presentation identifies exactly these concerns: high flow rates, pumping power, increased rack-level complexity and reduced efficiency at very high heat flux.
At moderate rack density, these are engineering challenges.
At extreme rack density, they may become architectural challenges.
Two Phase Changes the Physics
Two-phase direct liquid cooling uses a different mechanism.
Instead of relying only on the sensible heating of a liquid, the working fluid changes phase.
Liquid reaches the hot surface.
It boils.
The phase change absorbs a large amount of heat.
The vapor then travels away from the processor and is condensed back into liquid.
The thermal loop becomes:
Liquid → Evaporator / Vapor Cold Plate → Vapor → Condenser → Liquid
This ability to move heat through latent heat can significantly reduce the mass flow required for a given thermal load.
Why Lower Flow Matters
The OCP presentation provides one vendor comparison that is particularly interesting.
For the same processor heat load, it compares approximately:
Single Phase: 1.5 LPM/kW
Flow Boiling: 0.7 LPM/kW
Pool Boiling: 0.3 LPM/kW
These figures come from ZutaCore's technology presentation and should therefore be viewed as vendor-specific performance claims rather than universal industry values.
But the underlying engineering direction is important.
Two phase potentially allows more heat to move with less circulating liquid.
That could mean smaller pumping requirements, lower liquid mass, and a different approach to the thermal architecture of ultra-dense AI racks.
2028 Could Become an Inflection Point
DATAAD does not believe that the entire AI data-center industry will suddenly switch to two-phase cooling in 2028.
Technology transitions do not happen that way.
Single-phase systems will remain in service for many years.
Air cooling will remain in lower-density applications.
Hybrid architectures will continue.
But 2028 may still become an important inflection point.
Why?
Because the processor and rack roadmap begins to move into a range where simply increasing water flow becomes less attractive.
At the same time, two-phase systems are moving from laboratory demonstrations toward real deployments.
The OCP webinar cites an Oregon State University HPC and AI deployment using two-phase direct-to-chip cooling in existing infrastructure.
And an audience poll during the webinar gives another clue.
Although 60 percent of respondents said they had no plan for large-scale two-phase deployment, 15 percent expected implementation within 24 months and another 10 percent within 36 months.
For a technology that is still early in commercialization, that is meaningful.
The Most Important Point: The Industry Does Not Start From Zero
Perhaps the biggest reason a two-phase transition could accelerate is that the data-center industry has already begun building the physical infrastructure required for liquid cooling.
Today's single-phase DLC investments create experience with:
- cold plates,
- server liquid connections,
- rack manifolds,
- hoses,
- quick disconnects,
- CDUs,
- leak monitoring,
- and facility-side heat rejection.
Two-phase DLC does not necessarily discard this entire architecture.
It changes it.
This distinction could define the next cooling market.
Two Phase Is Not a Drop-In Replacement
Infrastructure inheritance does not mean component interchangeability.
A water-glycol hose system cannot simply be filled with an arbitrary refrigerant.
A standard single-phase QD may not be appropriate for vapor containment.
A seal selected for water service may behave differently with a fluorinated working fluid.
Two-phase systems introduce new requirements involving:
- working-fluid compatibility,
- permeation,
- vapor leakage,
- pressure control,
- liquid and vapor return architecture,
- internal volume,
- seal materials,
- and refrigerant recovery.
So the transition is not:
Single Phase → Replace the coolant → Two Phase
It is closer to:
Existing DLC Architecture → Re-engineered Two-Phase Fluid System
The Cold Plate Will Change First
The most obvious change occurs at the processor.
Single-phase microchannel plates are designed to move liquid continuously across a heated surface.
Two-phase evaporators must control boiling.
The OCP presentation shows microstructured surfaces, wick structures and controlled nucleation as important elements in pool-boiling cold plates.
The problem is no longer simply maximizing flow.
The challenge is controlling where and how the fluid boils.
This creates a new market for:
vapor cold plates, evaporators, microstructured surfaces and advanced thermal materials.
Then Comes the Working Fluid
The fluid itself becomes strategically important.
Water is inexpensive, widely available and thermally effective for single-phase DLC.
Two-phase systems require working fluids with a very different combination of properties.
The industry will have to consider:
- boiling temperature,
- latent heat,
- dielectric behavior,
- material compatibility,
- pressure characteristics,
- flammability,
- toxicity,
- and global warming potential.
Honeywell and Daikin Are Companies to Watch
This makes fluorochemical and refrigerant companies increasingly relevant to AI infrastructure.
Honeywell's former advanced-materials business — now Solstice Advanced Materials — has extensive low-GWP refrigerant capabilities and has already positioned its refrigerant and liquid-cooling technologies around more efficient data-center thermal management.
Daikin is another company worth watching.
Its DAISAVE family includes low-GWP fluorinated liquids intended for heat-transfer applications and data-center cooling, including grades with low boiling points and dielectric properties.
This does not mean that Honeywell or Daikin have already “won” the future two-phase DLC fluid market.
No universal two-phase direct-to-chip fluid standard has emerged.
But their fluorochemical expertise makes them strategically interesting as this ecosystem develops.
The Next Big Interface Market: QDs, Hoses and Fittings
If the working fluid changes, the interfaces must change with it.
This may create one of the least discussed opportunities in two-phase cooling.
Today's OCP liquid-cooling ecosystem has spent years developing common interfaces around single-phase UQD, UQDB and related connection concepts.
Two-phase DLC raises another set of questions:
How much vapor leakage is acceptable?
How much fluid loss can occur during disconnection?
What level of permeation is acceptable through a hose?
Which elastomers remain stable over the expected service life?
Should the liquid line and vapor return use the same coupling architecture?
How should technicians — or eventually robots — distinguish supply and vapor-return connections?
These are not small component questions.
They determine whether two-phase cooling can become a serviceable mass-market architecture.
The future two-phase market may be won as much by fluid-interface engineering as by the cold plate itself.
Hoses May Become More Sophisticated
Single-phase DLC hoses primarily have to move coolant reliably with low pressure loss and low leakage.
Two-phase operation can introduce additional requirements.
Refrigerant permeation may matter more.
Material compatibility may become more restrictive.
Internal volume becomes more important because working-fluid charge matters.
Vapor and liquid lines may require different diameters and routing strategies.
This could push the hose market from a relatively conventional fluid-transfer component toward an engineered part of the refrigeration circuit.
The CDU May Also Become Something Different
Today's CDU primarily controls and transfers heat between the technology cooling system and facility water.
In a two-phase architecture, that equipment may increasingly perform functions associated with refrigerant management:
separation + condensation + fluid storage + circulation + monitoring + pressure control.
Some vendors may continue calling these systems CDUs.
Others may describe them as heat-rejection units, refrigerant distribution units or two-phase cooling systems.
The terminology may change.
The function becomes more sophisticated.
And This Is Where HVAC Enters the Story
Two-phase DLC could also blur another boundary:
the boundary between IT cooling and HVAC.
Single-phase liquid cooling created a relatively clear division.
The server loop collected heat.
The CDU transferred it.
The facility chilled-water system rejected it.
Two-phase systems bring refrigeration physics deeper into the IT side of the data center.
This creates an obvious question:
Who is best positioned to integrate the refrigerant loop with the facility heat-rejection system?
LG Is Already Moving From Chip to Chiller
LG Electronics is an interesting company to watch because its current data-center portfolio already spans both sides of this boundary.
It is developing cold plates and CDUs on the IT side while maintaining major capabilities in chillers, CRAH systems and facility HVAC.
In 2026, LG publicly presented a 1.4 MW CDU as part of its AI data-center cooling portfolio.
That combination — chip-side liquid cooling plus facility-side HVAC — becomes particularly valuable if future thermal systems require deeper integration between the two.
Samsung Could Enter Through a Different Door
Samsung's route is different but equally interesting.
Its acquisition of FläktGroup gave it a much larger position in applied HVAC and mission-critical data-center cooling.
Samsung has said it intends to combine its chiller technologies with FläktGroup's cooling capabilities for data-center applications.
If two-phase DLC grows, the opportunity will not necessarily be limited to whoever makes the evaporator on top of the GPU.
The much larger opportunity may be designing the complete thermal system from:
silicon → refrigerant → rack → heat rejection → building.
The Market Could Become Much Broader Than Today's DLC Market
This is why the next generation of liquid cooling could be commercially fascinating.
The single-phase market created demand for:
Cold Plates + QDs + Hose Kits + Manifolds + CDUs.
The two-phase market could add another layer:
Refrigerants + Vapor Interfaces + Advanced Seals + Refrigerant Hoses + Condensers + Recovery Systems + HVAC Integration.
What looks today like a specialized server-cooling technology could therefore pull together several industries that historically operated separately.
Semiconductor thermal engineering.
Fluid connectors.
Industrial hoses.
Fluorochemicals.
Refrigeration.
HVAC.
Data-center controls.
2028 Is Not a Deadline — It Is a Window to Watch
There are still significant obstacles.
Standards are immature.
Working-fluid economics matter.
Environmental regulations matter.
Leakage and service procedures matter.
Operators must trust the technology.
Supply chains must develop.
And single-phase DLC will continue improving.
For those reasons, DATAAD does not view 2028 as a universal conversion date.
We view it as a possible market window.
The combination of rapidly increasing processor heat flux, rising rack density and several years of accumulated single-phase DLC experience could make the 2028–2030 period particularly important for two-phase direct-to-chip commercialization.
The Most Interesting Competition May Be Just Beginning
The first AI cooling race was straightforward.
Air versus liquid.
Liquid is increasingly winning the high-density segment.
The next race may be more complex.
Single-phase liquid versus two-phase liquid.
And this time, the winners may not come from one industry.
They may come from an ecosystem:
Cold Plate + Fluid + QD + Hose + Fitting + CDU + HVAC + Controls.
DATAAD Insight
The most important lesson from the current liquid-cooling transition may be that infrastructure changes more slowly than processors.
That is precisely why the spread of single-phase DLC matters so much.
It is teaching the data-center industry how to bring liquid into the rack.
Once that infrastructure, service knowledge and supply chain exist, the next thermal technology does not have to start from zero.
Two-phase cooling may inherit much of the architecture — while replacing the physics inside it.
Single-phase DLC is building the liquid infrastructure.
Two-phase DLC may be the technology that eventually pushes that infrastructure to its next level.
And if rack power continues to rise as expected, the period around 2028 may be when that transition becomes difficult to ignore.
DATAAD Special Analysis
Source basis includes the May 19, 2026 OCP Educational Webinar “Beyond Conventional Two-Phase Cooling: A New Approach for Scalable AI Infrastructure,” presented with IDTechEx and ZutaCore. Performance figures attributed to ZutaCore should be understood as vendor-reported examples. The 2028 transition timing is a DATAAD editorial thesis based on thermal-density trajectories and current deployment signals, not an official OCP or NVIDIA transition schedule.