Single Phase vs Two Phase DLC and What Latent Heat Buys

A Korean consortium building direct liquid cooling for 300kW-per-rack AI data centers is not choosing between single phase and two phase cooling. It is assigning them to different buildings. New construction gets a two-phase cold plate rated for a 3kW chip and a target power usage effectiveness of 1.0x. Retrofits of existing air-cooled halls get a single-phase cold plate rated for a 1.4kW chip and a PUE target of 1.1x. The split is written into a 19.24 billion won national research and development project that runs through the end of 2029.

The program splits by building type, not by cooling philosophy

The Korean government’s direct liquid cooling initiative treats single phase and two phase as two products for two markets, not two rungs on one upgrade ladder. The Ministry of Climate, Energy and Environment and the Korea Institute of Energy Technology Evaluation and Planning (KETEP) are funding “Development of a Direct Liquid Cooling System to Reduce Data Center Power Consumption,” led by Sinsung Engineering with ten partner institutions including the Korea Institute of Machinery and Materials and Korea Institute of Industrial Technology. The project runs from April 2026 through December 2029, with a total budget of 19.24 billion won, 13.0 billion of it from the government.

New-build data centers rated above 300kW per rack get a two-phase cold plate paired with a 2MW in-row cooling distribution unit (CDU) and a 350kW dry cooler, sized for chips with a thermal design power near 3kW. Existing air-cooled facilities get a single-phase cold plate paired with a 100kW in-rack CDU and a rear-door heat exchanger, sized for chips near 1.4kW and rack loads near 100kW, built to reuse the retrofit site’s existing floor space and air-handling equipment rather than replace it. Source: KHARN, “Sinsung Engineering Begins Developing DLC for 300kW/rack AI Data Centers” (August 2, 2026)

The dividing line the consortium drew is building type, not cooling performance. A project with a 3.9-year runway is attaching the higher-capacity coolant strategy to buildings that do not exist yet, and the lower-capacity one to buildings that already have poured concrete, existing chillers, and a fixed floor plan. That is a constraint on what can be installed, not a judgment about which physics performs better.

Table 1. Two DLC tracks in one national programKETEP-funded project, 2026-2029

* Targets stated in an active R&D program proposal, not field-verified results

Source: KHARN, “Sinsung Engineering Begins Developing DLC for 300kW/rack AI Data Centers” (August 2, 2026)

Two phase cooling trades sensible heat for latent heat, and the math is checkable

Two-phase cooling removes a given amount of heat with a smaller volume of coolant because it uses the heat absorbed during boiling rather than the heat absorbed by warming a liquid. A single-phase loop obeys Q = ṁ · cp · ΔT: heat removed equals mass flow rate times specific heat times temperature rise. A two-phase loop obeys Q = ṁ · h_fg: heat removed equals mass flow rate times the latent heat of vaporization, with no temperature change required at all once boiling starts.

Counterpoint Research’s February 2026 analysis of ZutaCore’s product line quotes a worked example for a 150kW rack: 220 liters per minute of water-glycol for a single-phase cold plate loop against 45 liters per minute of dielectric refrigerant for a two-phase loop, and describes two-phase cooling as removing the same heat “using only about 10% of the flow rate” of single phase. Source: Counterpoint Research, “AI Data Centers – ZutaCore In Pole Position To Benefit From Transition To Two-Phase Cooling” (February 20, 2026)

The single-phase side of that example checks out under the sensible-heat equation. Treating the 220 L/min flow as roughly 227 kg/min of a 25% propylene glycol mixture (specific heat about 3.8 kJ/kg·K) and solving ΔT = Q / (ṁ · cp) for a 150kW load gives a temperature rise of about 10.4°C, a plausible design point for a high-density facility water loop. The two-phase side can be checked against Honeywell’s published data for R1233zd(E), a refrigerant used in commercial two-phase cold plates: latent heat of vaporization of 194 kJ/kg at its 19°C boiling point and a liquid density of about 1,296 kg/m³. Solving ṁ = Q / h_fg for the same 150kW load gives a mass flow of about 46.4 kg/min, which converts to roughly 36 L/min of liquid refrigerant, a plausible figure given that a working system runs its boiling point warmer than 19°C, which lowers both the latent heat and the liquid density somewhat. Source: Honeywell Solstice Performance Fluid Technical Bulletin (accessed 2026)

That recomputed 36 L/min sits closer to 220 L/min at a ratio of about one-sixth than the vendor’s own worked example of 45 against 220, which comes to about one-fifth. Neither figure supports the “10% of the flow rate” claim in the same document’s prose. A flow ratio somewhere between one-sixth and one-fifth is still a genuine engineering win, but it is not the order-of-magnitude number the marketing implies, and a reader sizing pipe diameter or pump head off the 10% figure would undersize the system.

Lower flow does not remove the engineering burden, it relocates it

The KETEP project’s own thermal resistance targets show that two-phase cooling is not simply single-phase cooling with a smaller pipe. The consortium’s stated targets are a cold plate thermal resistance of 0.01 K/W for the two-phase design and 0.02 K/W for the single-phase design, against a current commercial baseline the consortium describes as 0.04 to 0.07 K/W. The two-phase cold plate is also targeted to sustain a heat flux above 100 W/cm² by combining a Tesla-valve channel structure at the inlet, meant to suppress vapor backflow, with a micropillar structure downstream, meant to keep a liquid film present on the surface through capillary action and re-wet the plate after local dry-out. Source: KHARN, “Sinsung Engineering Begins Developing DLC for 300kW/rack AI Data Centers” (August 2, 2026)

Dry-out and critical heat flux (CHF) reduction are specific to boiling: once vapor forms fast enough to blanket part of the surface, that patch stops receiving liquid contact and local temperature spikes, a failure mode single-phase cooling does not have because its coolant never changes state. The consortium’s single-phase cold plate, by contrast, is being redesigned around jet impingement combined with a micro-fin structure, aimed at cutting chip-surface temperature variation by as much as 50% and addressing a different problem: single-phase cooling’s difficulty steering flow to a hot spot on the die once the plate’s channel layout is fixed.

I read the choice of failure mode each design is built to solve as the real dividing line between the two approaches, more than the headline flow-rate savings. A single-phase plate fails by leaving one region of the chip undercooled while flow goes elsewhere. A two-phase plate fails by drying out a region once vapor generation outruns liquid replenishment. Neither failure mode disappears at lower flow; the two-phase design spends its flow-rate savings on managing a new one.

Table 2. Single-phase vs two-phase cooling, 150 kW rackexample values

* Vendor-published figures for one two-phase product line; not an industry average. Recomputed flow rate uses R1233zd(E) latent heat at its 19°C boiling point as a stand-in fluid

Source: Counterpoint Research (February 2026); Honeywell Solstice Performance Fluid Technical Bulletin

The retrofit path chose single phase for reasons beyond thermal performance

A parallel, KETEP-linked pilot project run by Samhwa Ace illustrates why a retrofit-focused developer would choose single phase even where two phase cools better. Samhwa Ace built a 120kW mock-up facility at its Anseong plant in 2024 and originally began development with two-phase cooling for its superior performance, then switched to single phase because of environmental regulatory risk tied to fluorinated dielectric refrigerants. The switch surfaced its own problems: the mock heat source did not match a real GPU’s thermal profile closely enough, and the coolant discolored within a month from oxidation; a purpose-built heat source matching the real thermal profile solved that. A second problem, micro-leaks from threaded fittings and seal tape unsuited to a low-surface-tension fluid, was resolved only after switching to sanitary-grade piping joined by welding. Source: KHARN, “[Kharn Report] Coolant Reliability and Efficiency Set to Drive Liquid Cooling Era” (undated, accessed 2026)

Reading the retrofit specification next to Samhwa Ace’s own account of abandoning two phase, the choice looks less like an engineering preference and more like a regulatory and liability one: a fluorinated dielectric fluid is a harder commitment to write into a decade-long retrofit contract than a water-glycol loop that any existing chilled-water plant already knows how to service. A retrofit customer is also, by definition, keeping most of the building’s existing air-handling equipment, and a rear-door heat exchanger bolted onto a single-phase loop reuses that equipment far more directly than a sealed two-phase vapor loop would.

What the efficiency numbers promise, and what they still need to prove

None of the PUE figures attached to this program are measured yet. They are targets a consortium wrote into a funding proposal, to be tested against real hardware between 2027 and 2029. The consortium’s own energy model, run against a hypothetical 40MW data center at a global-average baseline PUE of 1.54, projects total facility draw falling from 61.6MW to 42.2MW and annual electricity use falling by about 169.9GWh a year if PUE reaches 1.0x, worth an estimated 34.1 billion won in annual electricity cost and about 78,000 tCO2e in annual carbon reduction. Source: KHARN, “Sinsung Engineering Begins Developing DLC for 300kW/rack AI Data Centers” (August 2, 2026)

The demonstration itself is scheduled for a 10MW data center in Yeongdeungpo, Seoul, covering 9,733 m² across ten above-ground floors and four basement levels, where the two-phase system will be tested at 300kW per rack using a thermal test vehicle rather than production hardware, while the single-phase retrofit system will be tested at 100kW per rack using actual NPU servers. The plan calls for six months or more of operation including the summer months, over 1,000 total operating hours, across twelve combinations of load (25%, 50%, 75%, 100%) and season, evaluated against ASHRAE 90.4’s mechanical load component (MLC) criteria and JEDEC standards. Source: KHARN, “Sinsung Engineering Begins Developing DLC for 300kW/rack AI Data Centers” (August 2, 2026)

The difference between a 1.0x and a 1.1x PUE target on paper is the exact quantity these projects exist to test with a thermal test vehicle instead of a spreadsheet, and a program that tests its highest-density design on simulated heat rather than production GPUs is telling a reader something about how far from commercial deployment that design still sits.

What the flow-rate math assumes

This analysis does not cover whether the KETEP program’s PUE and thermal-resistance targets will be met; those are stated goals for work running through 2029, not measured outcomes, and program targets in Korean national R&D funding proposals are set to be ambitious enough to justify the award. It does not cover coolant costs, connector standardization, or field failure rates for either cooling family, none of which the sources here quantify. The flow-rate recalculation in this piece uses R1233zd(E) property data at its listed 19°C boiling point as a stand-in for the KETEP project’s actual refrigerant and operating pressure, which are not yet public; a real system’s saturation temperature, and therefore its exact flow rate, will differ. It does not cover immersion cooling, which also splits into single-phase and two-phase variants but is a separate cooling architecture from the cold-plate DLC covered here. Any use of these cooling systems in a real facility requires engineering review specific to the site, the refrigerant, and local fire and refrigerant-handling code.

Choosing a phase by building type and refrigerant rules

• Retrofitting an existing air-cooled facility without a fixed timeline to demolish it favors single-phase DLC, which was designed in this program specifically to reuse rear-door heat exchangers and existing air paths.
• A new build targeting 300kW per rack or higher, where the floor plan and mechanical room are not yet fixed, is the only scenario this program’s two-phase track is designed for.
• A vendor’s flow-rate reduction claim for two-phase cooling is worth recomputing against the coolant’s own published latent heat and density figures rather than accepted as stated, since even a vendor’s own worked example can differ from its own headline percentage.
• Regulatory exposure on fluorinated refrigerants is a real design constraint on two-phase adoption, independent of thermal performance, and should be checked against current and pending regulation in the country of deployment before committing to a fluid.
• A cold plate’s stated thermal resistance is only comparable across single-phase and two-phase designs when the test heat flux and coolant temperature are also stated, since the failure modes each design manages (localized dry-out versus localized undercooling) are different problems with different acceptable operating margins.

References

KHARN, “신성엔지니어링, 300kW/랙 AI DC용 ‘DLC’ 개발 착수” [Sinsung Engineering Begins Developing DLC for 300kW/rack AI Data Centers], August 2, 2026, https://www.kharn.kr/news/article.html?no=31504 KHARN, “[KHARN칸] 냉각액, 데이터센터 신뢰성·효율 좌우 리퀴드쿨링시대 핵심 인프라 급부상” [Coolant Reliability and Efficiency Set to Drive the Liquid Cooling Era], accessed 2026, https://www.kharn.kr/news/article.html?no=31639 Counterpoint Research, “AI Data Centers – Zutacore In Pole Position To Benefit From Transition To Two-Phase Cooling,” February 20, 2026, https://counterpointresearch.com/en/insights/AI-Data-Centers-Zutacore-In-Pole-Position-To-Benefit-From-Transition-To-Two-Phase-Cooling Honeywell, “Solstice Performance Fluid Technical Bulletin,” accessed 2026, https://advancedmaterials.honeywell.com/content/dam/sustainability/en/documents/document-lists/technical/SolsticePf-Specifications-TechBulletin.pdf ASHRAE, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers,” ASHRAE TC9.9 white paper, 2021, https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf Korea Institute of Energy Technology Evaluation and Planning (KETEP), 에너지수요관리핵심기술개발사업 [Energy Demand Management Core Technology Development Program], program overview, https://www.ketep.re.kr