Liquid-Cooling on lo0 — Tech Blog

Cooling the AI datacenter (1/4): the thermal challenge

Sat, 27 Jun 2026 20:00:00 +0200

The power series (in Spanish) followed the watt from the utility feed all the way to the chip. But there is an implacable physical law that closes the loop: almost every watt that goes into a server comes back out as heat. An inference factory does not only have to deliver megawatts to its GPUs; it has to remove those same megawatts as heat, and remove them fast enough that the silicon does not cook itself. That is the other half of physical infrastructure, and it is the half the AI era has turned upside down. This article opens the second series in the vertical: cooling.

The approach is the same one used in the power series: universal concepts, valid for any datacenter, read throughout against the question of why AI, with its 130 kW racks and its roadmap toward the megawatt, is forcing a historic transition from air to liquid.

The law that governs everything: power is heat

Start with the physical principle, because everything else follows from it. The electrical energy a server consumes does not disappear: by conservation of energy, practically all of it turns into heat that has to be evacuated. For practical purposes, the thermal power to dissipate equals the electrical power consumed. A rack that swallows 130 kW is a 130 kW radiator, the equivalent of several dozen domestic heaters switched on at once inside a single cabinet. Cooling means moving that heat from the chip to, eventually, the atmosphere, crossing a chain of media (air, water, refrigerant) as real as the electrical chain that fed it.

The ability of a fluid to carry heat obeys a simple but revealing relation: the heat a flow can remove is proportional to its mass flow rate, its heat capacity and the temperature rise it accepts,

$$\dot{Q} = \dot{m} \cdot c_p \cdot \Delta T$$

That formula alone explains why air is in trouble. Air has a low heat capacity and a minimal density, so removing a lot of heat means moving enormous quantities of it. Water, for the same volume, carries on the order of a thousand times more heat than air. When rack density spikes, sticking with air is like emptying a swimming pool with a hairdryer: technically conceivable, practically absurd.

Put numbers on it. Solving the formula above for flow rate, removing a given power with a fixed temperature rise requires $ \dot{m} = \dot{Q} / (c_p \Delta T) $. For a 50 kW rack cooled by air with a reasonable rise of about 11 °C (the usual 20 °F), the required airflow is around 7,800 cubic feet per minute: literally a gale blowing through the cabinet. Doubling density to 100 kW doubles the gale, and there is no way to contain that much moving air without the turbulence, noise and recirculation that ruin efficiency. Water does the same job with a ridiculously smaller flow and volume, because its $ \rho \cdot c_p $ product, the capacity to carry heat per unit of volume, is overwhelmingly higher. Physics leaves no room for nostalgia: above a certain density, air simply cannot.

AI density breaks the model

For decades, air was more than enough. A traditional rack dissipated between 5 and 15 kW, and a well-designed room cooled it without breaking a sweat. Air is comfortable: cheap, safe, no pipes to drip onto the electronics. The problem is that AI has multiplied density by ten.

The root of it is the chip. An H100 (Hopper) GPU dissipated about 700 W; a B200 (Blackwell) goes up to 1,000 W, and as far as 1,200 W in its full-specification version. That increase in power per chip is the first cause of the whole problem: it is not only that there are more chips, it is that each one dissipates almost twice what the previous generation did, on a silicon area that barely grows, so heat density per square centimetre spikes. An air heatsink has a physical limit to how much heat it can pull off a small surface before the surrounding air saturates; past that point, no matter how many fans are added, the chip throttles thermally and drops its clock so as not to burn, which means wasting very expensive silicon. Liquid, with direct contact and far greater capacity, pulls that heat away without the chip having to back off.

And those chips do not travel alone: the NVIDIA GB200 NVL72 packs 72 GPUs and 36 CPUs into a single rack rated at 120 kW, with real deployments measuring 130-132 kW at full load. The roadmap does not let up: the Vera Rubin generation points at 190-230 kW per rack, Rubin Ultra “Kyber” at around 600 kW in 2027, and there is already talk of the 1 MW rack toward 2028. The distant figures should be taken for what they are, vendor roadmap rather than deployed installations, but the slope is unmistakable.

Where is the limit of air? Industry figures place it with fair consensus: air is efficient up to about 10-20 kW per rack, room-level cooling reaches about 27 kW with effort, and row-level cooling about 30 kW. With advanced techniques such as rear-door heat exchangers it can be stretched to 40-60 kW. Above that, thermodynamics stops cooperating: cooling a 50 kW rack with air would take on the order of 7,800 cubic feet per minute, a gale impossible to contain inside a cabinet. Between 50 and 100 kW, liquid stops being an option and becomes the only viable physics. AI racks sit well above that boundary. That is why 2025-2026 is the moment of the great transition.

The spectrum of solutions

It helps to see cooling not as one technology but as a spectrum ordered by density. As heat per rack rises, you climb from one rung to the next:

Method	Typical density supported	Medium that touches the heat
Room air (CRAC/CRAH) + containment	up to ~30 kW/rack	Air
Rear-door heat exchanger (RDHx)	up to ~40-60 kW/rack	Air → water
Direct liquid to chip (DLC)	~50 kW to >130 kW/rack	Water/refrigerant in a cold plate
Immersion	100-250+ kW/rack	Dielectric fluid

Each rung brings the coolant closer to the hot spot. Air cools the whole room; the rear door intercepts hot air at the rack outlet; direct liquid to chip brings water to a cold plate pressed against the processor itself; and immersion submerges the entire electronics in a fluid that does not conduct electricity. The closer the liquid gets to the chip, the more heat is captured and the higher the density that can be operated. The next three articles in this series walk that spectrum in detail: air and its limits, direct liquid to chip, and immersion.

The reality of 2026, though, is not a clean replacement but coexistence. Even a direct-liquid rack still evacuates a fraction of its heat through air, between 2 % and 20 % depending on how close the liquid gets to each component, so the hybrid room, liquid for the GPUs and air for the rest, is the norm during the transition. Vendors have understood this: the most widely deployed solutions combine both worlds.

The journey of heat, end to end

Just as the power series followed the watt from the utility feed to the chip, it is worth following the heat in the opposite direction, from chip to atmosphere, because cooling is a chain with as many links as the electrical one. Heat is born in the transistor and conducts to the chip package; from there it passes to a heatsink or a cold plate; that first medium, air or liquid, carries it to a heat exchanger; a second loop, the facility water loop, takes it out of the room; and a final rejection stage, a cooling tower, a dry cooler or a chiller, hands it to the atmosphere. Every hop has a temperature difference and a loss, and the art of cooling consists of minimising the number of hops and maximising the temperature at which you work, so that less energy (and less water) is needed for the final rejection.

Here a concept appears that runs through the whole series: the coolant distribution unit (CDU), the thermal equivalent of a transformer. The CDU separates the loop that touches the equipment, clean and controlled at precise flow and temperature, from the facility loop, and exchanges heat between the two without mixing them. It is the boundary between the server vendor’s world and the building operator’s, and where a good part of the efficiency is decided. We will look at it in detail when we get to direct liquid to chip.

The metrics of heat

To govern cooling you have to measure it, and several metrics belong in an architect’s toolkit. The first is already familiar from the power series: PUE (Power Usage Effectiveness), the ratio between total facility energy and the energy that reaches IT equipment. What matters for cooling is that cooling and air movement consume between 30 % and 45 % of a traditional datacenter’s energy: by a wide margin, the largest energy overhead on top of compute. The industry average has been stuck around 1.54 for years, according to the Uptime Institute, precisely because air cooling is inefficient and hard to improve on.

The second metric, which AI has pushed to the foreground, is WUE (Water Usage Effectiveness), the litres of water consumed per kWh of IT:

$$\text{WUE} = \frac{\text{litres of water}}{\text{kWh of IT}}$$

The average sits around 1.8-1.9 litres per kWh, although the most efficient facilities aim to go below 0.2. Water is the hidden face of cooling, and we will come back to it.

The third is the ASHRAE TC9.9 framework, the reference that defines the temperature at which equipment may operate. For air, its classes run from A1 (15-32 °C inlet) to A4 (up to 45 °C); for liquid, a recent update defines water classes W17 to W45, where the number is the maximum supply temperature in degrees. The underlying trend is clear and counterintuitive: raise operating temperatures. The hotter the coolant can come in, the less it has to be chilled artificially, and the more ambient air or water can be used without compressors. Operating hot is operating efficiently.

That idea crystallises in the concept of warm-water cooling and the ambition to operate chillerless. The reasoning is elegant: if the chip is perfectly happy with 40 or 45 °C inlet water, there is no point spending energy producing 15 °C chilled water with compressors; a dry cooler rejecting heat to outside air, which is almost always below 45 °C, is enough. NVIDIA promotes exactly that, 45 °C inlet water (class W45) for its Vera Rubin generation, with a wide thermal rise (45 °C in, 65 °C return) that reduces pumping flow and, above all, eliminates chillers in most climates. The term deserves a caveat: “chillerless” usually means “almost chillerless”, because on the hottest days some mechanical cooling backup is still needed. But the direction is unmistakable, and it is the one that saves energy, saves water and, as we will see, allows heat reuse.

One measurement caveat to close on: PUE does not capture the consumption of the fans that live inside the server, which in an AI node is not negligible. Hence TUE (Total Usage Effectiveness), which combines PUE with the equipment’s own internal efficiency (ITUE) to give a more honest picture of how much energy actually reaches the transistors. As liquid removes server fans, the distinction gains importance: in an air-cooled AI node, internal fans can consume a far from negligible percentage of the server’s own power, energy that liquid recovers by making them unnecessary. It is one of the gains of liquid that PUE, on its own, cannot even see.

Water: the new constraint

If energy is the AI datacenter’s input constraint, water is becoming an equally hard output constraint. Many efficient cooling techniques, cooling towers and evaporative cooling among them, trade water for energy by evaporating water, and at AI scale that bill has become enormous. Estimates of AI datacenter water consumption already run into hundreds of billions of litres per year, with projections doubling toward 2030. The per-query figures in circulation, from a few millilitres to half a litre per request, should be handled with care, because they vary enormously with methodology, but the aggregate order of magnitude is real and concerning.

It is worth understanding the underlying technical dilemma, because it is a trade-off rather than a defect. A datacenter’s heat rejection equipment moves inside a triangle of energy, water and temperature. A cooling tower evaporates water to cool better: little energy, a lot of water (good PUE, bad WUE). A dry cooler evaporates nothing, so it barely uses water, but it can only cool to near ambient temperature, which demands operating with hotter water and, in warm climates, more backup energy. A mechanical chiller cools to any temperature but at a high energy cost. Raising operating temperatures, as noted above, is what allows the move from tower to dry cooler without an energy penalty: warm water and the waterless datacenter are two sides of the same coin. AI, by concentrating so much heat, forces a conscious choice inside that triangle where the choice used to be made by inertia.

And it is not only a technical problem but a social one. Competition for water with communities and agriculture has begun to stall projects: in the United States, “Project Blue” near Tucson, which expected to consume hundreds of millions of gallons a year, was cancelled after local opposition. The industry’s answer is twofold: on one side, waterless designs using dry coolers in exchange for operating with hotter water (Microsoft reports reaching a WUE of 0.30 and deploying zero-evaporation cooling); on the other, direct liquid to chip itself, which by capturing heat at a higher temperature makes it easier to reject to ambient without evaporating. The paradox of 2026 is that liquid, well designed, can use less water than the evaporative air it replaces.

Waste heat: from problem to resource

There is an optimistic twist. All that heat that costs so much to remove is, after all, energy, and reusing it rather than dumping it into the atmosphere is increasingly on the table. Waste heat reuse for district heating has gone from curiosity to regulatory obligation: the European energy efficiency directive pushes in that direction, and national laws already set quotas, with Germany requiring 10 % reuse in 2026 and up to 30 % in 2028.

Here liquid again has a decisive advantage over air. Modern district heating networks need water at 50-70 °C, a temperature that a liquid-cooled datacenter with its wide thermal rise (45 °C inlet and 65 °C return, for example) can deliver almost directly, whereas the hot air of a traditional room is too cool and too diffuse to be useful. It is no accident that the large reuse cases, heating tens of thousands of homes in northern Europe with Microsoft, Google and Meta datacenters wired into district heating networks, rest on liquid cooling. For an operator, waste heat can go from cost to revenue: heat delivered to a district network can be worth considerably less than the gas it displaces, with payback periods under a decade when the datacenter sits one or two kilometres from the network. At continental scale the potential is large: European datacenter waste heat is estimated to be able to cover around 10 % of EU space heating by 2030. What was vented to the atmosphere for decades is starting to be seen for what it is: wasted high-quality energy.

What is holding the transition back

If liquid is thermodynamically inevitable, why has it not swept air aside already? Because changing cooling paradigm touches everything an operator fears touching. The first brake is fear of leaks: bringing water within centimetres of the most expensive electronics on the planet makes anyone nervous, and although modern solutions use dripless quick connectors, leak detection and drip trays, the perception of risk weighs. The second is lack of standardisation: for years every vendor brought its own fluid, its own connectors and its own temperatures, which locked customers into a supplier and complicated maintenance; the sector is converging, but slowly. The third is cost and construction: liquid demands piping, CDUs, manifolds and often a rebuilt room, an investment that only high densities justify. And the fourth is operational and cultural: a team used to managing air conditioning has to learn to operate a hydraulic system, with its water chemistry, its pumps and its gaskets.

The result is that air does not disappear: it is reserved for loads that need nothing more, while liquid enters where density forces it. The question for an architect is rarely “air or liquid?” in the abstract, but “which part of my room needs liquid, when, and how does it coexist with air everywhere else?”. Retrofit, adapting existing air rooms to host liquid islands, is one of the big businesses of 2025-2026, precisely because almost nobody starts from scratch.

Where the market is in 2026

How far along is the transition? Uptime Institute data from 2025 gives a nuanced picture: direct liquid to chip is already present in around 22 % of operators, but perimeter air remains the majority at 75 %, and 61 % of those not yet using liquid are considering it. The barriers they cite are revealing, being lack of standards, cost and fear of a leak failing over the electronics, the same ones that hold back any paradigm change. The liquid cooling market, however, is growing at high double-digit rates, with estimates multiplying it by five or six over a decade, and the large players are repositioning fast: Vertiv leads with a good share of the first GB200 racks and has launched solutions combining direct liquid to chip and air in a single module; Schneider Electric bought Motivair to enter in earnest; and alongside them live cold-plate specialists (CoolIT, Boyd, JetCool, Asetek) and immersion specialists (Submer, GRC, LiquidStack). The dominant pattern is hybrid: rooms combining liquid for the GPUs and air for everything else, as a bridge to a mostly liquid future. That vendors package both worlds into a single product says a great deal about where the transition is heading: not to a clean cut, but to a managed coexistence lasting years.

For an inference factory

What should someone building an inference factory take from this? Three ideas that run through the whole series. First, that cooling is not an add-on but a co-design: the density of AI racks forces the cooling method to be decided at the same time as the power method, because a 130 kW rack cannot simply be dropped into a room designed for air. Second, that the move to liquid is no longer optional above a certain density; the question is not whether, but which, direct to chip or immersion, and how to manage coexistence with air for years. And third, that the metrics that matter have widened: watching PUE is no longer enough, water (WUE) has to be watched too and, increasingly, what is done with waste heat, because operating cost, the social licence to build and, in some geographies, legal compliance are all decided there.

It is also worth keeping in mind that cooling is not only efficiency: it is availability. As the power series showed, a thermal failure can bring a factory down as fast as a power cut, in fact faster, because a 130 kW rack without cooling reaches dangerous temperatures in seconds. That is why critical cooling is designed with the same philosophy as power: redundancy in pumps and CDUs, early detection, and testing. A cooling system nobody has exercised at full load is, once again, an unverified promise.

This article has drawn the map of the thermal challenge. The next three go down into the detail of each rung of the spectrum: starting with air cooling and its limits, the point everyone is coming from, moving on to the direct liquid to chip that is feeding GB200 racks, and finishing with immersion and the horizon of the megawatt rack. Feeding the GPUs was half the battle; the other half is keeping them from melting.

Sources

Network World, Why AI rack densities make liquid cooling nonnegotiable — https://www.networkworld.com/article/4149069/why-ai-rack-densities-make-liquid-cooling-nonnegotiable.html
Introl, Liquid Cooling vs Air: The 50kW GPU Rack Guide (2025) — https://introl.com/blog/liquid-cooling-gpu-data-centers-50kw-thermal-limits-guide
ToneCooling, NVIDIA GB200 NVL72 Cooling Requirements — https://tonecooling.com/nvidia-gb200-nvl72-cooling-requirements/
TweakTown, NVIDIA full-spec Blackwell B200 uses 1200W — https://www.tweaktown.com/news/97059/nvidias-full-spec-blackwell-b200-ai-gpu-uses-1200w-of-power-up-from-700w-on-hopper-h100/index.html
DCD, Nvidia’s Rubin Ultra NVL576 rack expected to be 600kW — https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/
Computer Weekly, Huge grid and heat challenges ahead as Nvidia set for 1MW rack — https://www.computerweekly.com/news/366639658/Huge-grid-and-heat-challenges-ahead-as-Nvidia-set-for-1MW-rack
mgrid / Uptime Institute, Global Data Center PUE Stalls at 1.54 (2025) — https://mgrid.org/2025/10/01/uptime-institute-data-center-pue-stagnation-2025-liquid-cooling/
ClearComfort, 2026 PUE and WUE Checklist — https://clearcomfort.com/pue-wue-ai-data-centers/
Introl, Water Usage Effectiveness (WUE) AI Data Center Cooling Guide 2025 — https://introl.com/blog/water-usage-efficiency-wue-ai-data-center-cooling-guide-2025
CKY, ASHRAE TC 9.9 Thermal Guidelines (5th Ed.) — https://www.cky.com.tw/en/insights/ashrae-tc9-datacenter-thermal-guidelines
DCD, Hot water, cold water: the right temperature for water cooling — https://www.datacenterdynamics.com/en/analysis/hot-water-cold-water/
EESI, Data Centers and Water Consumption — https://www.eesi.org/articles/view/data-centers-and-water-consumption
Microsoft Datacenters, Measuring energy and water efficiency — https://datacenters.microsoft.com/sustainability/efficiency/
Energy Solutions Intelligence, Data Center Waste Heat Recovery 2026 — https://energy-solutions.co/articles/sub/data-center-waste-heat-district-heating
NVIDIA Blog, Hotter Than a Hot Tub: 45°C liquid cooling for AI factories — https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/
Uptime Institute, Global Data Center Survey 2025 — https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2025

Cooling the AI datacenter (3/4): direct liquid to chip

Sat, 27 Jun 2026 18:00:00 +0200

The previous article ended with air exhausting its headroom on geometry: there is no floor through which to push seven thousand cubic feet per minute into a cabinet with a two-tile footprint. The way out is to bring the fluid to the hot spot until it touches the chip package itself. That is direct-to-chip liquid cooling, DLC, and it is the technology feeding the GB200 and GB300 racks shipping today.

The figure that justifies the change is textbook: water carries on the order of 4,000 times more heat per unit volume than air and has around 30 times the thermal conductivity. What follows is the engineering between that statement and a room in production, which turns out to be considerably richer than “add a pipe”.

Inside the cold plate

The part that touches the silicon is the cold plate: a copper block with an internal cavity crossed by microchannels, resting on the processor package through a thermal interface material. The fluid enters cold, crosses the channels absorbing heat by forced convection, and leaves hot. Typical geometry for a plate serving a processor of around 2,000 W works with fins and channels 0.15 to 0.25 mm wide, channel heights of 1 to 2 mm, and a copper base of about 2.5 mm.

Its figure of merit is thermal resistance, defined by the OCP specification as the ratio between the package temperature rise over the fluid inlet and the applied power:

$$R = \frac{T_c - T_L}{Q}$$

OCP insists on a nuance with practical consequences: resistance must be characterised as a curve against flow rate, not as a loose catalogue number. Published bench tests with commercial loops give values on the order of 0.02 °C/W at 550 W, with pressure drops of about 6.5 psi in the best of the three loops tested. That same bench produced a result anyone designing a rack should keep in mind: most of the circuit’s pressure drop came not from the cold plate but from the quick disconnects.

The cold plate is also where material constraints concentrate. Fluid velocity inside the microchannels must stay below 1.5 m/s to avoid erosion. The electrochemical potential difference between any pair of metals in contact inside the plate must not exceed 0.15 V, a far stricter limit than for the rest of the circuit, because slow galvanic corrosion in a 0.2 mm channel blocks it. The mechanical qualification OCP demands includes hydrostatic testing at three times maximum operating pressure, a pressure-drop test with less than 0.5 % loss, salt spray per ASTM B117, and internal X-ray inspection to find voids and manufacturing debris. Fin pitch on commercial plates runs around 50 microns and tends to narrow, which directly sets the loop’s filtration requirement.

How much heat the plate captures

Here is the nuance that separates the brochure from the engineering. A cold plate cools the component it covers, and a server has many more components than the GPU. The fraction of heat the liquid removes depends on how many of them are covered.

Experimental tests on real racks give capture ratios of 94 % on a 53 kW rack and 93 % on a 128 kW multi-rack assembly. Vendor deployment guides are considerably more conservative: Vertiv talks of 70 to 75 % in cold plates and 25 to 30 % left to air. Reference designs sit in the middle band: OCP’s model for 125 kW racks assumes 80/20, the Vertiv 360AI design for a 1.2 MW pod splits 76/24, and Chilldyne’s design for an NVL72 declares 72/28, which leaves 320 kW of residual air load in the room. At the high end, Supermicro announces up to 98 % capture per rack on its DLC-2 platform, and achieves it the only way possible: by putting cold plates on memory, PCIe switches and voltage regulators too, not only on CPU and GPU.

That residual fraction has two design consequences. The first is that the air room still exists, with its containment and its perimeter equipment, sized for 5 or 25 % of the heat depending on how complete the cold-plate solution is. The second is economic: internal fans in a high-performance server consume 10 % to 20 % of system power, and that line item only disappears entirely when the need to move air through the chassis disappears. A DLC capturing 72 % leaves the fans spinning, slower but spinning. The difference between 72 % and 95 % capture is not a nuance of thermal efficiency: it is the difference between paying that bill and not paying it.

The two loops

A direct-to-chip liquid system always has two separate hydraulic circuits whose fluids never mix.

The TCS (Technology Cooling System) is the secondary loop: it runs from the CDU to the cold plates, through manifolds and quick disconnects. It is a closed circuit of known volume with a controlled fluid, and it is the one that touches the IT equipment. The FWS (Facility Water System) is the building’s primary loop: it carries water from the heat rejection equipment to the CDU. Between them sits a plate heat exchanger inside the CDU, transferring heat without transferring fluid.

That separation is what makes the system governable. The TCS can have its own chemistry, its own pressure and its own filtration, far finer than will ever be reasonable in a building’s pipework. And because the exchange is never perfect, the TCS is always hotter than the FWS, by a difference that is one of the numbers most conditioning the design of the whole facility.

The CDU: the thermal transformer

The coolant distribution unit (CDU) is the boundary between the server vendor’s world and the building operator’s. It holds the heat exchanger, the pumps, the filtration, the buffer tank and the control. There are two architectures and three locations.

By architecture, a liquid-to-liquid (L2L) CDU rejects heat to building water, and is what gets installed where there is pipework. A liquid-to-air (L2A) CDU rejects it to the room’s own air, with no need to connect anything to the hydraulic installation, in exchange for returning all that heat to the existing air system. The usual sizing rule places L2A in deployments of one to ten racks and recommends L2L above 500 kW. By location, they come in-rack, in-row and perimeter.

Commercial capacities today span three orders of magnitude. Vertiv’s range runs from 70 kW in L2A format to 2,300 kW in L2L. The Liebert XDU1350 declares 1,368 kW nominal at 4 °C approach and up to 2,912 kW at 8 °C, with a nominal flow of 1,200 l/min in N+1 two-pump mode and triple-redundant 50 micron filtration cleanable in line. The CoolIT CHx2000 reaches 2,000 kW at 5 °C approach and serves twelve GB300 NVL72 racks from a single unit. At the hyperscale end, Google’s Project Deschutes CDU, contributed to OCP in 2025, is a 2 MW unit with a 3 °C approach, 80 psi of available pressure and 0.2 micron side-stream filtration. Motivair, now inside Schneider Electric, announces architectures scaling to 10 MW.

Of all those numbers, the one that governs the whole facility is the approach temperature: how many degrees separate the building’s inlet water from the coolant leaving toward the equipment. The state of the art is between 2 and 4 °C, and some vendors offer 6. The consequence is direct and often overlooked at purchase time: the lower the CDU’s approach, the higher the building water setpoint can be, and the higher that setpoint, the more hours a year the dry cooler runs without compressor help. Two degrees of approach translate into weeks of free cooling.

Redundancy follows the logic already seen in the electrical chain. Pumps are configured N+1 with automatic changeover, with isolation valves and quick disconnects so they can be replaced without draining the circuit. And there is a failure-domain rule that deserves a place in any specification: size the CDUs so that each one serves less than 10-20 % of the liquid-cooled servers, so a loop failure does not take out the whole room. The CDU also keeps secondary supply above the room’s dew point, which guarantees purely sensible cooling and avoids condensation on the electronics.

Flow, temperature rise and the price of warm water

The flow specification almost every design works with is 1.5 l/min per kW, with the server side sitting between 1.2 and 1.5. The design temperature rise of the TCS loop is around 10 °C for GPUs, although full-load tests with 128 kW racks have measured rises of nearly 15 °C. A 1 MW pod split across twenty 50 kW racks asks for on the order of 1,100 l/min in total, a figure that gives a sense of pipe sizing and pump size.

On that basis comes the most interesting argument of the moment. NVIDIA has set as a platform target that all MGX racks operate with 45 °C inlet water, and describes the full free-cooling scenario: the facility delivers 41 °C to the CDUs and these supply 45 °C to the racks. The argument is sound, because at that temperature almost any climate allows heat rejection with dry coolers, without mechanical compression, and the electrical budget freed up translates into more racks per contracted megawatt.

The price of that decision is in the hydraulics, and platform data for the GB300 quantifies it:

Supply temperature	Flow per rack	Pressure drop
25 °C	59 l/min	2.3 psi
30 °C	71 l/min	3.2 psi
35 °C	89 l/min	4.9 psi
40 °C	119 l/min	8.5 psi
45 °C	177 l/min	18.4 psi

Going from 25 to 45 °C triples the flow and multiplies pressure drop by eight, paid for in pumping, manifold sizing and CDU size. Which is why OCP defends a different position in its roadmap: 30 °C coolant as a durable minimum, arguing that at that temperature free cooling or mixed mode already covers almost the whole year in many climate zones, that fifteen years of PUE improvement are preserved, and that heat reuse options stay open. The same document warns that dropping to 20 °C raises annual chiller power by around 20 %, and that even at 30 °C a chillerless facility cannot be promised in every scenario, because capacity has to be sustained at full load and under fault conditions.

The two positions contradict each other less than it seems. NVIDIA specifies the upper limit the equipment tolerates; OCP recommends the operating point that optimises the whole. A reasonable design in southern Europe sits in the 30 to 35 °C supply band, with hydraulic capacity held in reserve to go higher if climate or tariff advise it.

Manifolds and connectors

Between the CDU and the cold plates lies a two-level distribution network. Row manifolds join the CDU to the racks and mount overhead or under the raised floor. Rack manifolds distribute inside the cabinet: vertical ones mount at the rear, horizontal ones can go at the front. NVIDIA already integrates tray manifolds, rack manifolds with UQD08 connectors and liquid-cooled busbars up to 5,000 A into the Vera Rubin platform, a sign that the boundary between electrical and thermal distribution is blurring inside the cabinet.

The part that makes all this maintainable is the dripless quick disconnect. OCP maintains two separate specifications: UQD for manual connection and UQDB for blind mate, the one that lets a compute tray connect hydraulically as it slides into the rack with nobody screwing anything. The UQDB test methodology validates interchangeability between suppliers with a full millimetre of radial misalignment, and vendors declare full flow with gaps of one to two millimetres. Spillage on disconnecting a reference flat-face connector is below 0.015 cc, on the order of one drop. Typical working ranges reach 150 psi with burst pressure four times higher and lives of more than 10,000 cycles.

That an open connector specification exists matters more than it seems. For years every vendor brought its own, which tied the customer to one supplier for the room’s whole life. The connector is the piece of plumbing that decides whether another vendor’s rack fits in your row.

The chemistry of the loop

The TCS fluid is not plain water. The de facto standard is PG25, a mixture of roughly 25 % by volume propylene glycol in water with corrosion inhibitors and antifoaming agents. Its properties at 20 °C give a specific heat of about 3.9 kJ/kg·K, thermal conductivity of 0.49 W/mK, viscosity of 2.4 cP and a freezing point of -10 °C. Pure deionised water has more than 4 % more heat capacity and pumps better; glycol is paid for in thermal performance and pumping energy, and in exchange gives two things: freeze protection and biological control.

That biological control has a surprising detail that explains a fair number of failures. Glycols are biostatic from 25 % by volume, but at concentrations on the order of 10 % biomass tends to feed on the propylene glycol rather than be inhibited by it. A badly dosed loop, or one topped up with water after a purge, is worse off than if it carried no glycol at all. The resulting biofilm blocks 0.2 mm channels and ruins heat transfer with no warning beyond a pressure drop that creeps up slowly.

The fluid quality specification is short and deserves a place in the maintenance contract:

Parameter	PG25 (TCS)	Plant water (FWS)
pH	8.0-10.5	7.0-9.0
Total hardness	< 50 ppm	< 200 ppm
Copper	< 2 ppm	—
Iron	< 2 ppm	—
Chlorides	< 25 ppm	< 50 ppm
Sulfates	< 25 ppm	< 100 ppm

Chlorides and sulfates accelerate pitting and galvanic attack, especially in aluminium and copper alloys. The design recommendation is to keep wetted materials as monometallic as possible, copper and brass or stainless steel, with a close anodic index between parts, and to avoid uninhibited aluminium and galvanised steel.

Filtration changes completely between loops. The TCS works below 25 microns, and commercial CDUs fit redundant 50 micron filters; the FWS accepts the 300 to 500 micron range. At the extreme, Google’s 2 MW CDU filters to 0.2 microns in a side stream to extend coolant life. OCP recommends on-site fluid testing quarterly and laboratory analysis annually, with standardised methods for metal ions, pH, reserve alkalinity and chlorides. Common operating practice is more intense at the start: daily analysis during commissioning, weekly after the first week and monthly in steady state. Documented coolant degradations are almost always attributed to the same causes: a badly executed initial flush, insufficient maintenance, or air ingress into the circuit.

Leaks, or the fear that was not the problem

The objection that always turns up in the first meeting is water over the electronics. The system defends itself in three layers.

The first is detection. Direct sensors, leak detection cables and point probes at the low points of trays and manifolds are combined with indirect detection through buffer tank level and bubble sensors: if air comes in or liquid is missing, something is open. The GB300 NVL72 platform already integrates leak detection at tray and rack level, with the signal available to the BMS.

The second is hydraulic design. The most elegant solution is to operate the server side below atmospheric pressure: with supply at about 0.13 bar of vacuum and return at 0.61 bar, a rupture draws air in rather than pushing coolant out. Incoming airflow is limited with check valves and venturis, and the measured thermal effect during a leak is a rise of one to three degrees in CPU temperature. A DOE cluster operated on this scheme has accumulated several years without a single recorded leak.

The third is factory qualification: hydrostatic testing at three times operating pressure, shock and vibration testing with subsequent leak-tightness verification, and at some vendors PG25 leak testing on 100 % of units plus an acceptance test simulating megawatts of load.

The result in production is good. Google reports CDU fleet availability of around 99.999 % since 2020, with liquid cooling in production since 2018, close to a gigawatt of liquid-cooled capacity and roughly half its global footprint with liquid deployed or enabled.

And here is the data point that dismantles the usual narrative. Asked about the barriers holding back DLC adoption, operators surveyed by Uptime in 2025 cite lack of standardisation (39 %), cost (38 %) and reliability (35 %), while safety concerns, the fear of water over electronics, appear in barely 7 %. What holds liquid back is not the leak. It is that there is still no comfortable standard, that it is expensive, and that it introduces failure modes operations teams do not know. Uptime adds an honest caveat: the sector is not applying enterprise IT redundancy standards to liquid today, because most installed capacity is in HPC and AI, where batch workloads tolerate interruptions a transactional system would not.

Getting liquid into an air-cooled room

Almost nobody starts from scratch, which is why the factor weighing most in the decision to adopt DLC, according to the same survey, is ease of retrofit onto existing infrastructure (46 % of responses), ahead even of operating savings.

The canonical retrofit route is the liquid-to-air CDU in sidecar format: a cabinet placed next to the rack, closing its own circuit with the cold plates and rejecting heat to room air. Commercial units run from 200 kW, at 1,145 kg and 19 kW of own consumption, to 500 kW at 2,300 kg and 55 kW. Building pipework does not have to be touched, and that is the whole appeal: it turns a construction problem into a purchasing problem. The cost is that the room has to be able to absorb those kilowatts in air, which in practice limits the approach to small islands.

The two physical limits of retrofit are weight and clear height. A GB200 NVL72 rack weighs on the order of 1,360 kg empty and reaches 1,500 kg with coolant, which demands floor positions rated for that load and, frequently, a concrete slab rather than a raised floor. Adding a one or two tonne sidecar alongside does not help. The conclusion deployment guides keep repeating is uncomfortable but realistic: most existing rooms cannot take a rack of this class without structural reinforcement, electrical expansion and thermal redesign, and the typical timeline for a full conversion, across manufacturing, delivery, integration and commissioning, is around a year.

For an inference factory

Four decisions concentrate almost all the value.

The first is to specify the capture fraction, not just the power. A tender asking for “liquid cooling for 130 kW racks” without saying which components carry a cold plate ends up with a room that has to remove 30 kW of air per cabinet and with internal fans consuming what liquid was supposed to save. The right question to the vendor is what percentage of node heat the fluid captures and on which components.

The second is to treat CDU approach temperature as a purchasing parameter. Two degrees of difference between a 4 °C CDU and a 2 °C one translate into two more degrees of FWS setpoint, and those two degrees are compressor-off hours in every year of the facility’s life. It is the thermal equivalent of choosing a UPS efficiency well.

The third is chemistry and filtration, which is where facilities are lost at the two-year mark. PG25 at a measured 25 %, not at the 12 % left after a top-up with water; a documented initial flush; filtration matched to the plates’ fin pitch; and a maintenance contract with quarterly analysis and written thresholds. A dirty loop gives no warning: performance drops slowly until a GPU starts throttling.

The fourth is the failure domain. Spread the load across CDUs so none serves more than 10 to 20 % of the servers, fit N+1 pumps with automatic changeover, and feed the CDUs from the backed-up system. The reason was given in the first article: a 130 kW rack without flow reaches dangerous temperatures in seconds, and critical cooling in an inference factory is designed with the same discipline as its electrical chain. The comparison is exact: the CDU is to cooling what the UPS is to power, and nobody would buy a UPS without module redundancy.

Direct liquid to chip is today the default answer between 50 and 150 kW per rack, and the whole ecosystem, from NVIDIA reference architectures to OCP specifications, is built around the cold plate. What remains to be seen is the next rung, the one that submerges the entire electronics, and what happens when the rack approaches a megawatt. That is the last article in the series.

Sources

Open Compute Project, Cold Plate Development and Qualification — https://www.opencompute.org/documents/ocp-cold-plate-development-and-qualification-with-integrated-comments-pdf
Open Compute Project, 30 °C Coolant: A Durable Roadmap for the Future, Rev 1.0 — https://ashrae.org.vn/wp-content/uploads/2024/12/30%C2%B0C-Coolant-A-Durable-Roadmap-for-the-Future-REV1_0.pdf
Open Compute Project, Universal Quick Disconnect (UQD) Specification Rev 1.0 — https://www.opencompute.org/documents/ocp-universal-quick-disconnect-uqd-specification-rev-1-0-2-pdf
Open Compute Project, Universal Quick Disconnect Blind-Mate (UQDB) Specification Rev 1.0 — https://www.opencompute.org/documents/uqdb-spec-1-0-pdf
Open Compute Project, Open Rack V3 Blind Mate Manifold Specification Rev 1.0 — https://www.opencompute.org/documents/open-rack-v3-blind-mate-manifold-specification-rev-1-0-review-april05-2024-pdf
ASHRAE TC 9.9, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers — https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf
Dell Technologies / OCP, Liquid Coolants Guidance for Technology Cooling System and Facility Water System — https://www.delltechnologies.com/asset/en-us/products/servers/industry-market/liquid-coolants-guidance-for-technology-cooling-system-and-facility-water-system-whitepaper.pdf
Schneider Electric, Direct Liquid Cooling System Challenges in Data Centers (White Paper 210) — https://media.datacenterdynamics.com/media/documents/WP210_V1_EN.pdf
Schneider Electric Blog, What coolant for data center liquid cooling: PG25, dielectric fluids, ASHRAE recommendations — https://blog.se.com/datacenter/2026/08/14/what-coolant-for-data-center-liquid-cooling-pg25-dielectric-fluids-ashrae-recommendations/
Schneider Electric Blog, Liquid cooling failures in AI data centers: causes, risks, prevention — https://blog.se.com/datacenter/2026/07/15/liquid-cooling-failures-ai-data-centers-causes-risks-prevention/
NSF Public Access Repository, Experimental evaluation of direct-to-chip cold plate liquid cooling for high heat density data centers — https://par.nsf.gov/servlets/purl/10529800
NVIDIA Technical Blog, NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer — https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/
NVIDIA, NVL72 AI Factory Enterprise Reference Architecture: System Hardware and Components — https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/components.html
NVIDIA Blog, Chill Factor: Blackwell Platform Boosts Water Efficiency by Over 300x — https://blogs.nvidia.com/blog/blackwell-platform-water-efficiency-liquid-cooling-data-centers-ai-factories/
Uptime Intelligence, Data Center Cooling Systems Survey 2025 (UI Field Report 181) — https://intelligence.uptimeinstitute.com/sites/default/files/2025-07/UI%20Field%20181_Data%20center%20cooling.pdf
Uptime Institute Journal, Performance expectations of liquid cooling need a reality check — https://journal.uptimeinstitute.com/performance-expectations-of-liquid-cooling-need-a-reality-check/
Uptime Institute Journal, AI embraces liquid cooling, but enterprise IT is slow to follow — https://journal.uptimeinstitute.com/ai-embraces-liquid-cooling-but-enterprise-it-is-slow-to-follow/
Vertiv, Deploying Liquid Cooling in the Data Center: A Guide to High-Density Cooling — https://www.vertiv.com/4926c8/globalassets/documents/white-papers/liquid-cooling/deploying-liquid-cooling-in-the-data-center-a-guide-to-high-density-cooling-white-paper.pdf
Vertiv, Liebert XDU1350 Coolant Distribution Unit, datasheet — https://www.vertiv.com/49971a/globalassets/shared/liebert-xdu1350-coolant-distribution-unitcdu-ds-en-na-sl-70799-web.pdf
Vertiv, 360AI High Density Reference Design RD0019 — https://www.vertiv.com/493c85/globalassets/documents/ai-hub-reference-design/vertiv-360ai-reference-design---rd0019.pdf
CoolIT Systems, CHx2000 CDU — https://www.coolitsystems.com/cdu-product/chx2000/
Chilldyne, FAQ Guide to Data Center Liquid Cooling — https://chilldyne.com/wp-content/uploads/2025/03/Chilldyne-FAQ-X.4.pdf
Chilldyne, Reference Design for AI: NVIDIA NVL72 — https://chilldyne.com/wp-content/uploads/2025/03/Chilldyne-Reference-Design-for-AI-NVIDIA-NVL72-X.3.pdf
Businesswire / Boyd, Google’s Two Megawatt Project Deschutes Coolant Distribution Unit at OCP Summit 2025 — https://www.businesswire.com/news/home/20251010498843/en/Boyd-to-Showcase-Googles-Recently-Released-Two-Megawatt-Project-Deschutes-Coolant-Distribution-Unit-at-the-2025-Open-Compute-Project-OCP-Summit
StorageReview, Inside Google’s plan to deliver 1MW racks and cool them too — https://www.storagereview.com/news/inside-googles-plan-to-deliver-1mw-racks-and-cool-them-too
Supermicro, DLC-2 Next Generation Direct Liquid Cooling Solutions — https://www.supermicro.com/en/pressreleases/supermicros-dlc-2-next-generation-direct-liquid-cooling-solutions-aims-reduce-data
Supermicro, Liquid-to-Air Sidecar CDU datasheet — https://www.supermicro.com/datasheet/datasheet_Sidecar_CDU.pdf
Dell’Oro Group, Data Center Liquid Cooling Market to Approach 7 Billion USD by 2029 — https://www.delloro.com/news/data-center-liquid-cooling-market-to-approach-7-billion-by-2029-as-ai-deployments-accelerate/

Liquid-Cooling on lo0 — Tech Blog

Cooling the AI datacenter (1/4): the thermal challenge

The law that governs everything: power is heat

AI density breaks the model

The spectrum of solutions

The journey of heat, end to end

The metrics of heat

Water: the new constraint

Waste heat: from problem to resource

What is holding the transition back

Where the market is in 2026

For an inference factory

See also

Sources

Cooling the AI datacenter (3/4): direct liquid to chip

Inside the cold plate

How much heat the plate captures

The two loops

The CDU: the thermal transformer

Flow, temperature rise and the price of warm water

Manifolds and connectors

The chemistry of the loop

Leaks, or the fear that was not the problem

Getting liquid into an air-cooled room

For an inference factory

See also

Sources