Cooling the AI datacenter (4/4): immersion and the megawatt horizon

Contents

We reach the last rung of the spectrum laid out in the first article. If the cold plate brings liquid up to the processor package, immersion removes the last intermediary: it submerges the whole board in a fluid that does not conduct electricity. On paper it is the definitive solution, with the best physics in the catalogue. In practice, 2026 has left it in a secondary role, and the reasons behind that distance between thermodynamics and deployment are the interesting part of this article.

This also closes the series, so the final section looks ahead: to the 600 kW rack announced for 2027, to the megawatt per cabinet, and to a technology that could make the whole argument obsolete.

The single-phase tank

In single-phase immersion the server lives inside a bath of dielectric fluid that never changes state. The fluid absorbs heat by convection, natural or pump-forced, and hands it to a heat exchanger that passes it to facility water. The fluid families admitted by OCP’s base specification are four: hydrocarbons, fluorochemicals, natural esters and synthetic esters.

That specification is useful because it puts hard numbers where marketing puts adjectives. An immersion fluid must maintain dielectric strength above 6 kV/mm throughout its service life, a flash point of at least 150 °C, a boiling point above 155 °C in single-phase, volume resistivity above \( 10^{11} \) ohm·cm and dynamic viscosity below 0.015 N·s/m². That last figure is the one governing operating cost: a fluid too viscous forces so much pumping that it eats the efficiency gain. And there is a requirement almost nobody expects to find on a cooling datasheet: dielectric constant below 2.3 and loss tangent below 0.05 between 20 MHz and 40 GHz, because the fluid is physically inside the path of high-speed signals.

Commercial capacities per tank have grown fast. Submer declares up to 361 kW per unit in its SmartPod EXO, with redundant pumps and CDU and operation on hot water up to 60 °C. GRC offers 368 kW CDUs with 13 °C water and a range going down to 13 kW for edge installations. Asperitas illustrates the design axis well: its natural convection model runs without pumps, with zero watts of non-IT consumption and around 32 kW per tank, while its forced convection version reaches 60 kW per tank with vertical cassettes. Academic reviews place commercialised single-phase around 100 kW per cabinet, with vendor-declared pPUE below 1.08 and in some cases close to 1.03.

Those pPUE figures need careful reading: they are partial, covering only the cooling loop and not total facility consumption. GRC itself, which sells single-phase, puts the efficiency difference between single-phase and two-phase at around 1 %, which already suggests the choice between them is not settled on thermal performance.

What has to change in the server

Here is the part commercial presentations tiptoe around, and the one that decides a deployment’s operational viability. A standard server is not submerged as it comes. OCP’s design guidelines for immersion equipment list the modifications, and the list is long.

Fans are not supported in most immersion strategies, so they have to be disconnected. The practical detail is that many firmwares refuse to boot without a tachometer signal, so a retrofit needs an emulator supplying the pulses the BMC expects to see. Conventional thermal paste can dissolve into the fluid, with a double effect: heat transfer at the chip is lost and the bath is contaminated; the recommendation is to replace it with a solid interface material such as indium foil, or to remove the heatsink if bath cooling is enough. Mechanical hard drives are excluded except helium-sealed ones, and in practice storage becomes SSD or NVMe.

Optics present a little-known and rather elegant problem: submerging a transceiver fills the ferrule’s air gap with fluid, changes the refractive index and introduces reflection losses. The ways out are copper cables with sealed connectors, silicon photonics, or extenders keeping the optics out of the bath. Labels are erased, because adhesives and inks dissolve, and with them serial numbers and MACs, which forces engraving or external documentation. Elastomers also suffer: EPDM swells, and in swelling can bend capacitor leads. Cable jackets with plasticisers, chlorine or sulfur dissolve and contaminate the fluid. And heat-shrink materials degrade above 60 °C.

None of this is insurmountable, and server vendors certified for immersion have it solved at the factory. But it explains why the ecosystem took so long to start: for years the real blocker was not thermal, it was contractual, because nobody warranted a submerged server. That changed in 2025, when Intel certified single-phase immersion fluids for the first time, Shell’s, and introduced a specific warranty extension for Xeon processors in immersion. Supermicro certified the first complete system under that scheme, and ExxonMobil added a second fluid family late in the year. Immersion has been viable in the lab for a decade; it has been purchasable with a warranty for eighteen months.

Boiling: the two-phase case

Two-phase immersion changes the mechanism. The fluid, a fluorochemical with a low boiling point, boils on contact with the hot chip surface; the vapour rises and condenses on a cooled coil mounted in the tank lid, and the condensate falls back into the bath. The cycle needs no pumps for the primary fluid, and phase-change heat transfer is an order of magnitude apart.

The numbers make it clear. Published reviews place the heat flux limit of forced air around 1.6 W/cm² for room air and up to a few tens with an optimised heatsink; water in a cold plate reaches the 16 to 200 W/cm² band depending on conditions; boiling reaches 500 W/cm² and remains effective above 1,000 W/cm². Critical heat fluxes measured with FC-72 on smooth copper are around 48 W/cm², and double with structured surfaces. Typical boiling points sit between 49 and 61 °C depending on fluid. Microsoft deployed the first two-phase system in production at a cloud provider in Quincy in 2021, with a fluid boiling at 50 °C, and measured per-server consumption reductions of 5 to 15 %. Commercially, LiquidStack declares up to 250 kW per 4U tank and 1.5 MW in a 48U format.

Two-phase’s problem is not thermal. It is the fluid, along two routes. The first is economic and operational: the vapour escapes. One system documented in the literature lost 102 litres a year on a total volume of 595, around 17 % annually, with a reference fluid at 75 dollars a litre. That is a recurring expense no business model digests well. The second is regulatory, and it is the one that has settled the contest.

The PFAS wall

In December 2022, 3M announced its complete exit from PFAS manufacturing by the end of 2025, including the fluorinated fluids that underpinned almost all two-phase immersion on the market. It was a business of some 1.3 billion dollars a year, abandoned under regulatory and litigation pressure.

At the same time, the universal PFAS restriction under REACH was advancing in Europe. Submitted in January 2023, it covers more than ten thousand substances and affects manufacture, placing on the market and use. ECHA’s risk assessment committee issued a final opinion favouring a broad restriction in March 2026, with the socio-economic committee’s opinion in public consultation and expected by year end. There are temporary derogations of up to twelve years for medical devices and semiconductors, and intermediate derogations for other sectors, but datacenter heat transfer fluids do not appear with specific treatment, and that is the risk for anyone who has built their room around a fluorochemical.

Alternatives exist and are nascent. Chemours is developing Opteon 2P50, a hydrofluoroolefin with a GWP of 10 that it trialled at full scale in Japan in March 2025 alongside NTT DATA and Hibiya, still in development and without a firm commercial date. Other suppliers cited as substitutes continue to offer fluorinated chemistry, and that is the nuance usually lost in the debate: a low GWP solves the fluid’s climate problem, not necessarily its fit within a regulatory definition of PFAS that rests on molecular structure. Until that uncertainty closes, an operator signing up to fifteen years of service life on a fluorinated fluid is accepting a risk they do not control.

Single-phase immersion with hydrocarbons or esters sits outside this problem, and it is one of the reasons it holds more than 70 % of the immersion segment today.

The third way: precision immersion

Between the open tank and the cold plate there is an intermediate approach worth mentioning. So-called chassis-level precision immersion encapsulates each server in its own sealed enclosure and floods only that volume, with flow directed at the components that need it. It captures practically 100 % of system heat, memory, storage and power supplies included, removes the fans and operates at noise levels below 50 dB, all inside a standard rack and without rebuilding the facility.

Its advocates’ commercial argument is direct and describes the tank’s problem well: classic immersion demands facility re-engineering, is heavy, hard to maintain and hard to deploy outside hyperscale environments. That the argument comes from inside the immersion field itself says a good deal.

Why direct liquid to chip is winning

With better physics, published standards and silicon vendor certifications, immersion ought to be eating the market. It is not, and the 2025 and 2026 figures show it from several angles.

Uptime’s 2025 cooling systems survey puts direct liquid cooling adoption at 22 % of installations. Immersion does not appear broken out in the report, which indicates its share is below the reporting threshold. Dell’Oro, which estimates the global liquid cooling market at some 3 billion dollars in 2025 and projects close to 7 billion by 2029, states that single-phase DLC accounts for the great majority of capacity entering service and describes immersion as a technology that “finds adoption selectively, where architectural trade-offs justify it”. Estimates of immersion market size range between 349 and 570 million dollars for 2025 depending on the analyst house, with annual growth of 24 to 25 %, healthy but on a small base and below the rate at which the total liquid market is doubling. An appreciable share of that figure also corresponds to cryptocurrency mining and edge installations, which is where immersion found its first mass market.

The reasons appear in that same Uptime survey, in the question about what makes liquid viable. The two most-voted factors are ease of retrofit onto existing infrastructure (46 %) and ease of maintenance (35 %), which are precisely the two axes where the tank loses by definition. A liquid-to-air sidecar is placed next to an existing rack in an afternoon. An immersion tank is construction work: it changes room layout, floor loading, the server handling procedure and staff training.

Add to that the gravity of the ecosystem. NVIDIA’s reference architectures for GB300 NVL72 and for Vera Rubin are cold plate, with manifolds, UQD connectors and liquid-cooled busbars. Google, with roughly half its global footprint on liquid deployed or enabled and around a gigawatt of capacity across two thousand AI chip pods, does it with cold plates and has contributed its 2 MW CDU to OCP. The capital of 2025 went the same way: Eaton’s purchase of Boyd Thermal, Daikin’s entry into Chilldyne, Motivair’s integration into Schneider. And the most eloquent indicator comes from two-phase’s own standard-bearer: LiquidStack launched a one-megawatt CDU for direct liquid to chip in 2024 and keeps extending that line. No company abandons its founding thesis, but all of them add the product that sells.

There is a final argument immersion has lost along the way: water. For years the tank was sold as the only route to a datacenter with no water consumption. Microsoft reached a fleet WUE of 0.30 litres per kWh in 2024, against 0.49 in 2021, and its next-generation design with closed-loop chip-level cooling avoids more than 125 million litres a year per facility with a WUE close to zero. Every design of theirs since August 2024 uses it, with the first sites operational in 2026. The water advantage no longer distinguishes immersion.

What the tank does keep, and it is real, is outlet temperature. In both single-phase and two-phase, the fluid returns heat above 60 °C stably, which is exactly what a district heating network needs. With heat reuse quotas entering European regulation, that is the niche where immersion has an argument the cold plate does not match as comfortably.

The horizon: from the 142 kW rack to the megawatt

The roadmap NVIDIA has published sets the pace for all physical infrastructure:

GenerationPower per rackYearType of figure
Hopper~40 kW2023deployed
GB200 NVL72120-130 kW2024-25deployed
GB300 NVL72up to 142 kW2025specification
Vera Rubin NVL144190-230 kW2026roadmap
Rubin Ultra NVL576 “Kyber”600 kWH2 2027roadmap
1 MW rack1,000 kWfrom 2027roadmap

The Rubin Ultra announced for the second half of 2027 declares 576 GPUs, 2.5 million parts, 365 TB of fast memory and mandatory liquid cooling, with projected TDPs on the order of 3.6 kW per GPU. Vera Rubin reuses the Blackwell generation chassis with minor changes; Rubin Ultra demands a new rack and a new cooling system.

At that scale electrical distribution changes too, and here the series closes on itself. The 800 V DC architecture NVIDIA promotes for these racks carries 85 % more power through the same conductor than a 415 V AC distribution, with 45 % less copper and up to five points of improvement in end-to-end efficiency. In physical terms: it avoids having to run two-hundred-kilo copper bars to a single cabinet. It is the same trend anticipated in the power series (in Spanish), now with real facilities operating under that scheme and with the power vendor ecosystem lined up behind it.

None of those platforms contemplates immersion. All assume cold plate, warm water and CDU.

Microfluidics: attacking heat at the source

One line of work remains that could redraw the whole map, and it comes not from the facility side but from chip packaging.

The starting point is a paper published in Nature in 2020 by the EPFL group: microchannels fabricated monolithically inside the semiconductor substrate itself, co-designed with the electronics rather than bolted on top. The results were of another scale: heat fluxes above 1.7 kW/cm², with coefficients of performance above 10,000 and pumping powers of barely 0.57 W/cm², fifty times better than conventional straight microchannels.

Six years later that is leaving the lab along two paths. Microsoft announced silicon-etched microfluidics in September 2025, with channels the thickness of a hair and leaf-vein-inspired patterns designed with AI help to follow the chip’s thermal signatures; it declares up to three times better heat extraction than a cold plate and a 65 % reduction in maximum temperature rise inside a GPU, after four design iterations in a year. And TSMC presented direct-to-silicon liquid cooling integrated into its CoWoS packaging platform at ECTC 2025: a silicon microcooler fusion-bonded to the back of the chip, with no thermal interface layer, thermal resistance of 0.055 °C/W and demonstrated sustained TDP above 2.6 kW while keeping junction temperature rise below 63 °C.

If that line matures, the tank versus cold plate debate loses much of its point, because the bottleneck stops being how the board is bathed and moves inside the package. And one detail decides where the industry migrates: once heat leaves the chip in water at 45 or 60 °C, the infrastructure needed downstream is exactly the one direct liquid to chip already builds. Manifolds, CDUs, dry coolers. The tank has no natural path there.

Takeaways, and closing the series

Immersion has the best physics in the catalogue and a modest market position. That apparent contradiction sums up well what physical infrastructure teaches: the winning technology is not always the one that cools best, but the one that fits into rooms that already exist, with teams that already know how to operate them, with warranties somebody signs and with standards that allow changing supplier. Immersion still has its place, in new facilities built around it, in geographies with aggressive regulatory PUE caps, at the silent edge, and in deployments where high-temperature heat recovery is part of the business model. But anyone building an inference factory in Europe in 2026 is going to fit cold plates, and will be right to.

Three ideas run through all four articles.

The first is that cooling is a co-design, not an add-on. Rack density decides electrical distribution, floor loading, thermal method and water temperature at the same time, and those decisions are taken together or they are not taken well. A room designed for 40 kW per cabinet that then tries to host 130 ends up paying twice.

The second is that operating temperature is the most profitable efficiency lever there is. Raising the setpoint, reducing CDU approach and working with warm water is what allows the compressor to be switched off, the evaporative tower to be dropped and, along the way, waste heat to be delivered at a useful temperature. Energy, water and reuse move together, and all depend on the same parameter.

The third is that cooling is availability. A 130 kW rack without flow reaches dangerous temperatures in seconds, faster than it would fall over from a power cut. Pump and CDU redundancy, early detection, a bounded failure domain and real load testing belong to the same discipline as the UPS and the generator set of the power series (in Spanish). A thermal system nobody has exercised at full load is, once again, an unverified promise.

This closes the physical infrastructure vertical opened with the power chain. Between the two series the watt’s full journey is described: it enters through the utility feed, crosses switching and the UPS, reaches the silicon, converts entirely into heat and leaves through water toward the atmosphere or toward a heating network. What happens in the middle, the compute, is what usually takes the attention. Without the rest, it does not happen.

See also

Sources