# Two-phase cold plates run in desert heat on warm water loops

By MarketScale Newsroom · Published 2026-10-23 · Software & Technology on MarketScale
Canonical: https://www.marketscale.com/industries/software-and-technology/two-phase-cooling-runs-ai-chips-in-45-degree-desert-heat-without-water-1273523

> UNICOM's Austin Hypes on why two-phase cold plates and immersion gain ground, and the PFAS rules that complicate it.

## Key points

- A Motivair by Schneider Electric engineer puts single-phase cold plates at an estimated 55% of the 2026 liquid cooling market and expects them to dominate for five to 10 years.

Austin Hypes starts with a rack that was engineered for 20 kilowatts of air-cooled servers and now has to hold 50 kilowatts of AI gear. The building's air handlers were sized for the first number. Speaking on a MarketScale-produced episode of UNICOM Engineering's Flawless Execution podcast with host Daniel Litwin, the company's chief technologist walks through five ways an operator can get the extra heat out of the room, and why he thinks the two newest are gaining ground.

The timing is set by chip power. In a May 2026 post on its own site, GPU cloud provider Cudo Compute put an Nvidia GB200 NVL72 rack at 120 to 132 kilowatts and said the next generation will exceed 240 kilowatts. Also in May 2026, Greg Alexander, a senior thermal engineer at Motivair by Schneider Electric, wrote on the company's own blog that single-phase cold plates remain the default and, citing Persistence Market Research, hold an estimated 55% of the 2026 liquid cooling market.

Hypes agrees single-phase will be the most widely deployed option for a few more years. His argument is about what happens at the margin as pressures, connection counts and outdoor temperatures climb.

## The rear door is the cheapest first step and it tops out fast

Hypes calls the rear door heat exchanger the foundational step for a room built for air. The servers stay air-cooled and draw cold air at the front. A liquid-fed radiator mounted on the back door of the rack catches the hot exhaust before it reaches the room's air conditioning.

He describes it as usually one of the lowest-cost ways into liquid cooling because the IT equipment does not change, only the rack supporting it. The limit is capacity. Cudo Compute puts the rear door ceiling at roughly 40 to 70 kilowatts a rack, depending on the facility.

In an April 2023 whitepaper, the Open Compute Project, or OCP, wrote that a traditional door designed to minimize approach temperature can enable a 30 degree Celsius room on 27 degree Celsius facility water, and gave an illustrative example, not a measured result, of a hybrid design pairing cold plates with the door in which a 30 kilowatt rack is cooled on 30 degree water with fan power at 1.5% of rack IT capacity. For an operator with a stranded air-cooled hall, that is the cheapest bridge. It is not where the GB200-class racks in Cudo's figures end up.

## Single-phase glycol wins on availability and loses on pressure

Direct liquid cooling, or DLC, puts a small sealed heat exchanger called a cold plate on each CPU and GPU, with a manifold at the back of the rack feeding cold fluid in and hot fluid out. In single-phase DLC the fluid stays liquid the whole way round. Hypes says the usual fill is a propylene glycol solution the industry calls 25, a cousin of car radiator fluid, which Alexander's post describes as 75% water and 25% glycol.

Hypes calls it the bread and butter of the industry, available from most every tier one server maker, and says it will be around for a long time. The limit is physics. Thermal design power, or TDP, is the heat a chip is rated to produce, and as TDP climbs the cold plate needs more flow, which means bigger pipes or higher pressure.

Pipes cannot grow without the server growing. So the industry raises pressure.

> Well, as anyone who works with any mechanical system can tell you, higher pressure ultimately leads to more failures. — Austin Hypes, Chief Technologist, UNICOM Engineering

The connection count is the other side of that risk.

> So it is not uncommon in a typical size rack to have greater than 90 individual fluid connections between the ones inside the servers and the servers themselves connecting to the manifolds on the racks. — Austin Hypes, Chief Technologist, UNICOM Engineering

He says that count is growing and in many cases can double, and that every joint is a potential point of failure that can drip or spray a conductive fluid across live boards. Warning systems catch many small leaks, he says, but nothing is perfect, and when a single-phase outage happens it tends to be large.

The industry's own target sets the scale. Dober, a coolant supplier, writes in its own material, citing 2023 OCP liquid cooling guidelines, that OCP targets an annual cold plate and coolant loop failure rate of 0.3% or lower. For a fleet of 10,000 cold plates, that target allows 30 loop failures a year, arithmetic on the target rather than anyone's measured result.

Dober adds, citing Uptime Institute data relayed in a 2026 analysis, that water and coolant events sit persistently among the top five root causes of unplanned data center outages. For the operator, that makes leak detection a line item of the same order as the pumps.

## Two-phase cold plates move milliliters per chip and run in desert heat

Two-phase DLC uses the same cold plate position with a different fluid. A refrigerant, calibrated to boil at a set temperature, enters the plate, boils on contact with the hot chip, and leaves as vapor for a condenser tied to a water loop. Alexander's post describes the same mechanism: heat is carried away by the latent heat of vaporization, the energy a liquid absorbs as it turns to gas, so the fluid picks up heat with little temperature rise and at a constant temperature.

Hypes draws three advantages from that. Hypes says two-phase cold plates run on milliliters of refrigerant per chip where single-phase needs liters a minute. That lets the plumbing shrink to small-diameter pipes while chip density keeps climbing, on a technology the refrigeration industry has run for decades.

Second, the refrigerants are dielectric, meaning they do not conduct electricity. A minor leak is unlikely to damage the server it is in and less likely still to damage its neighbors. That blast radius, how much equipment goes offline or gets damaged in a failure, is usually a primary or secondary reason buyers look at two-phase, Hypes says.

Third is temperature. Because the refrigerant does its work by boiling, the coolant can enter the chip warm, and everything downstream, including the facility water loop, can run warmer too.

For an operator siting in a hot, dry region, the water point is the decision. A loop that condenses vapor at desert ambient temperatures needs no chiller compressors and no evaporative towers, and Hypes says it often runs with somewhat better power efficiency as well.

## Schneider's engineer says single-phase keeps the field for five to 10 years

Alexander makes the case for staying put. In the May 2026 post he wrote that two-phase is more complex, needs a bigger infrastructure investment and more deployment time, is seldom used today, and that single-phase is likely to dominate for the next five to 10 years because it is easier to deploy, costs less and aligns with silicon roadmaps. He named two conditions that would favor two-phase, chips too dense to pump enough coolant through and workloads that need a constant chip surface temperature, and wrote that neither is occurring yet.

His sharpest example is Nvidia's newly announced Vera Rubin NVL72, which he wrote can be cooled with single-phase at 45 degrees Celsius. That is a coolant temperature, not an outdoor one, but it sets a high bar for the density argument.

Hypes does not dispute the share. He says single-phase will likely still be the most widely deployed technology for new deployments for the next few years. Where he parts from Alexander is on the cost of a failure and on the refresh cycle: each single-phase generation, he says, tends to be unique and vendor specific, validated as a reference design, so next year's server is not a slide-out, slide-in swap and some infrastructure gets ripped and replaced.

Cost cuts both ways in the two accounts. Hypes puts the rear door at the low end and Alexander puts two-phase at the high end. Neither company publishes a cost premium for two-phase cold plates or immersion against single-phase at scale.

## The refrigerant question runs through PFAS rules and the Novec phase-out

The dielectric fluids carry a regulatory question the interview does not reach. David McGlocklin, a cooling engineering manager at Schneider Electric, wrote on the company's blog in May 2026 that per- and polyfluoroalkyl substances, the group of chemicals known as PFAS, are used in two-phase immersion and two-phase direct-to-chip cooling. He wrote that reporting on PFAS use under the Toxic Substances Control Act, or TSCA, has been pushed by the Environmental Protection Agency, after three extensions, to January 31, 2027.

McGlocklin added that data center operators are among the organizations that must report once the rule takes effect, that TSCA does not ban PFAS outright, and that New Jersey, Maine, Massachusetts, Minnesota and Washington have or are planning their own requirements. The reporting is designed to control use and, eventually, phase the chemicals out.

The supply side moved first. McGlocklin wrote that 3M, which he describes as the biggest PFAS producer, has phased out production of Novec, a fluid that had been used extensively in two-phase immersion and two-phase cold plate systems, and that the combination of regulation and that discontinuation has halted the growth of two-phase immersion cooling. Research into alternatives is under way but slow, by his account.

The Environmental and Energy Study Institute, a policy research group, wrote in an article citing October 2025 press reporting that PFAS accumulate in the environment and the body, that cooling systems are one of three PFAS sources in a data center alongside semiconductors and fire suppressants, and that direct releases from closed cooling loops are likely limited, with the larger emissions historically coming from the plants that make the fluids. For the buyer, Hypes' physics holds: the fluids are nonconductive and well understood. The chemistry behind them is what an operator may have to report in 2027.

## Immersion buys an hour when the chilled water stops

Immersion is the approach where the fluid touches the electronics. The server goes into a dielectric bath, either an open tank or a sealed chassis that becomes its own enclosure, and the fluid flows over memory, add-in cards and SSDs rather than the hottest chips alone. Hypes' frame is that air is a fluid too, and immersion is air cooling with a far denser medium that flows slower and holds far more heat.

That heat capacity, which he calls thermal mass, is the operational point. When a pump or the building water loop drops out, the fluid keeps absorbing heat, so Hypes says equipment can stay up for ten or fifteen minutes and sometimes an hour or more, enough time to move workloads elsewhere or shut down gracefully. A cold plate loop has no such buffer.

DatacenterDynamics reported in January 2026 that Australian HPC provider DUG runs tens of thousands of immersion-cooled servers across Houston, Kuala Lumpur and Perth, that its standard tanks started near 50 kilowatts and were pushed to 60 to 70, and that its Nomad modules reach about 100 kilowatts with a second heat exchanger. The same report said providers commonly offer tank densities of 100 kilowatts or more, some reaching 200 to 300, that NTT and Firmus run immersion deployments, and that some research suggests higher uptime and more resilient hardware, while crypto miners remain the biggest buyers.

The catch Hypes names is the hottest chip. Immersion fluid still needs heat sinks on the GPU to pull heat into copper or aluminum fins, and he says it is not as efficient as a cold plate at the highest densities. Because the servers do not plug into a manifold, though, a tank can hold gear from several vendors and several generations, the flexibility single-phase racks lack.

The sealed-chassis variant is the one Hypes points at edge sites. A technician pulls the chassis from the rack like a cold plate server and never touches the fluid.

## OCP has connector specs, and the market still has many sizes

Hypes answers the standards question with a caveat. There are OCP specifications for cold plate connectors, he says, but there are many to choose from, in many sizes and several generations, which produces incompatibilities between vendors and between generations.

OCP's own sub-project page shows the work in progress. The Open Compute Project's Cold Plate Sub-Project lists a Universal Quick Disconnect, or UQD, version 2.0 workstream revising the hand-mate and blind-mate connector spec to close gaps in version 1, with the draft still in beta. A quick disconnect is a coupling that seals both halves when unplugged. Alongside it sit an Open Rack V3 blind-mate quick connector whose hose evaluation summary is dated January 14, 2026, and a pivoting blind-mate coupling meant to carry the higher flow rates of higher power systems.

OCP announced a published guideline for dielectric fluids in two-phase cold plate racks in July 2024, with a water-based single-phase guideline and a propylene glycol guideline alongside it. The sub-project's stated scope runs from the cold plate to the coolant distribution unit, or CDU, the pump and heat exchanger cabinet that keeps the rack loop separate from facility water.

What the page does not deliver is a single connector every OEM ships, which is the gap Hypes describes. Server makers have not laid out publicly which GB200 or GB300 class systems they validate for two-phase cold plates or immersion. Hypes says single-phase is the configuration available from most tier one vendors.

ASHRAE, the engineering society that sets thermal guidelines for IT equipment, adds the operating envelope. Cudo Compute writes that ASHRAE Technical Committee 9.9 introduced an H1 environmental envelope for high-density systems that need tighter temperature bands, and published liquid cooling guidelines in January 2026 covering CDU design, transient modelling and coolant quality monitoring.

## Hybrids pair cold plates with a tank or a door

Hybrid, in Hypes' usage, means any two of the above. The pairing he expects most is two-phase cold plates on the GPUs inside an immersion tank, with the dielectric bath cooling the low-power parts and carrying the bulk heat from the plates to a heat exchanger. The second is cold plates with a rear door.

Most cold plate servers still need fans for the parts a plate does not touch. Hypes gives the example of a server that rejects 2,000 watts to air after the plates take the rest, and says a liquid-fed door can bring that remainder back to room temperature so the facility needs no air handler. The measure he gives operators is heat capture, the share of a rack's heat that leaves in liquid.

Two signals will tell operators whether the ground is shifting. Alexander's is chip wattage: he puts the industry flow standard at 1.5 liters per minute per kilowatt, and the day a GPU needs more than a server-sized pipe can pump, pumping power alone favors two-phase, a day he says has not come. McGlocklin's is January 31, 2027, when PFAS reporting under TSCA takes effect for operators. The question Hypes hears from buyers weighing two-phase is the blast radius: how much equipment goes offline, or gets damaged, when one of a rack's connections lets go.

## Sources

- [Cold Plate (Cooling Environments Project)](https://www.opencompute.org/projects/cooling-environments/cold-plate/) (Open Compute Project)
- [Evaluating the Limits of Door Heat Exchanger Solutions Optimized for Open Rack V3](https://www.opencompute.org/documents/acs-door-hx-whitepaper-final-230419-pdf) (Open Compute Project)
- [Two-phase vs single-phase direct-to-chip liquid cooling](https://blog.se.com/datacenter/2026/05/05/two-phase-vs-single-phase-in-direct-liquid-cooling/) (Schneider Electric (Motivair))
- [PFAS phase-out & data center liquid cooling: US 2026 guide](https://blog.se.com/datacenter/2026/05/22/pfas-phase-out-liquid-cooling-us-data-center-operators-must-do/) (Schneider Electric)
- [Immersion cooling and IT reliability: What do the numbers say?](https://www.datacenterdynamics.com/en/analysis/immersion-cooling-and-it-reliability-what-do-the-numbers-say/) (DatacenterDynamics)

Tags: UNICOM Engineering, liquid cooling, two-phase cooling, immersion cooling, AI data centers, data center operations

---
Source: MarketScale, https://www.marketscale.com/industries/software-and-technology/two-phase-cooling-runs-ai-chips-in-45-degree-desert-heat-without-water-1273523. Published for AI indexing and citation; cite the canonical URL. Site guide for agents: https://www.marketscale.com/llms.txt