Liquid Cooling Deployment Cases Sorted by Where Heat Goes

Direct liquid cooling (DLC), meaning coolant pumped through cold plates on the chips, always ends at a handoff where the rack’s loop gives its heat to something else. Microsoft’s heat exchanger units for older halls hand it to the chilled air the hall already supplies, Google’s in-row coolant distribution units (CDUs) hand it to a facility loop, and Microsoft’s Fairwater campus sends it outdoors to cooling fins fitted with 172 fans. This data center liquid cooling case study sorts six public disclosures from five operators by that handoff rather than by headline capacity, because the handoff is the part the host building has to supply.

Four of five operators name direct-to-chip, and each pairs it with a different handoff

The sources call the method direct-to-chip, D2C or DTC. This piece writes DLC for all three and orders the disclosures by date in Table 1.

Table 1. Six liquid cooling disclosures, as the operators framed them (Ordered by announcement date)

* The figures in the fourth column count different things (sites, rack rating, fleet, site feed) and are not comparable with one another.

Source: Equinix press release (Dec 11, 2023); Microsoft Tech Community post and Meta Engineering post (Oct 15, 2024); Google Cloud Blog (Apr 29, 2025); Microsoft Blog (Sep 18, 2025); Byline Network report of a KT Cloud announcement (Nov 6, 2025)

Equinix’s December 2023 release says DLC was commercially supported in more than 45 metros and that support would reach more than 100 of its IBX sites. The release gives no kilowatt figure. The count therefore measures where a customer may install DLC, not how much liquid-cooled load exists.

Meta’s October 2024 Catalina post is the one disclosure that does not name the cooling method. It says the full solution is liquid cooled and that the ORv3 high-power rack supports up to 140 kW.

Google’s April 2025 post describes in-row CDUs that keep the rack loop separate from the facility loop. Microsoft’s September 2025 Fairwater post describes liquid cooling built into the facility, with hot water piped to cooling fins on each side of the building. Its October 2024 post on heat exchanger units (HXUs) covers existing air-cooled halls, where a standalone unit rejects heat into the chilled air that the hall already supplies.

The handoff follows what the building already has. Where a facility loop exists, as in Fairwater and in Google’s CDU design, the rack loop hands its heat to liquid. Where only room air exists, as with Microsoft’s HXUs, it hands the heat to air. KT Cloud’s Gasan center, reported by Byline Network from the company’s opening announcement, is new construction using cold plates and cold water, which fits the first pattern, although the report gives no facility-loop detail.

A retrofit pod spends four of its six rack positions on cooling

Datacenter Dynamics (DCD), citing SemiAnalysis and WCCTech, reports that Meta showed a Catalina pod at Hot Chips 2025 for halls designed around 20 kW air-cooled racks.

The pod has two compute racks and four liquid-to-air cooling racks, two on each side. The side racks pull air from the hall floor through a heat exchanger that cools a local closed liquid loop, so the heat still ends in hall air.

DCD’s compute figures do not agree with each other. Its headline says 120 kW a rack, and its text says two 120 kW racks.

It also gives 144 GPUs per rack, double the 72 that Microsoft’s Fairwater post puts in a GB200 rack. Two readings follow, and each can be run through the pod.

Reading A takes the text literally: 2 × 120 kW = 240 kW over 6 rack positions, which is 40 kW per position, twice the 20 kW hall figure.

Reading B takes 120 kW as the total for the pair: 120 kW ÷ 6 positions = 20 kW per position, equal to the hall figure with nothing spare.

The page does not settle which reading holds. Under either, the side racks hand heat to hall air, so the limit the pod meets is the air-side capacity of the hall, not the temperature of the chip. For an owner weighing a retrofit that avoids rebuilding the data center, that moves the first question from the cold plate to the room.

DCD also relays SemiAnalysis’s remark that colocation rent follows kilowatts rather than floor area. On that basis the four cooling positions look cheaper than the hall kilowatts the pod uses, an inference of mine and not a figure from either source.

Rack ratings, site feeds and fleet totals cannot be ranked against one another

Byline Network’s report gives KT Cloud’s Gasan center 40 MW of receiving power and 26 MW of IT capacity. The ratio is 40 MW ÷ 26 MW = 1.54. It compares two design capacities, so it is not a PUE. At full IT load it leaves at most 14 MW for cooling, power conversion and reserve, and it is an upper bound because the receiving figure may include margin.

Set 26 MW against the two rack figures in this material. At Meta’s 140 kW rating, 26,000 kW ÷ 140 kW = 185.7, so 185 racks fit. At the 20 kW hall figure DCD cites, 26,000 kW ÷ 20 kW = 1,300 racks. Byline Network’s report gives neither a rack count nor a density, so both numbers illustrate the spread, and 140 kW is a design ceiling rather than an operating load.

The coolant side scales the same way, using Q = ṁ × cp × ΔT. Assume a ΔT of 10 K, a cp of 4.18 kJ/(kg·K) and 0.997 kg/L for water at 25 °C. None of these comes from a disclosure. For one 140 kW rack, ṁ = 140 kW ÷ (4.18 × 10) = 3.35 kg/s, or about 202 L/min. For 26 MW, ṁ = 26,000 kW ÷ 41.8 = 622 kg/s, or about 37,400 L/min. Both results assume all the heat goes to water. A rack that sheds part of its heat to air needs less, and a glycol mix needs more.

Google’s numbers need their clocks separated. Its post counts seven years of liquid cooling, starting with TPU v3 in 2018.

The 99.999% covers a shorter span, fleet-wide CDU availability since 2020. At that availability, 525,600 min × 0.00001 = 5.3 min a year is the downtime budget for whatever unit is measured, and the post does not say whether that unit is one CDU, all CDUs, or a pod that keeps running through a CDU fault.

DCD reports around 1 GW across some 2,000 TPU pods at the same event. Google says more than 2,000, so 1,000,000 kW ÷ 2,000 = 500 kW per pod is a rough upper estimate of the average, since “around 1 GW” is itself approximate.

Each number describes a different object. Equinix’s counts sites where service is offered, Meta’s is a rating, KT’s is a power feed, and Google’s is a fleet. Sorting them by size sorts units, not cases.

These disclosures support a placement claim, not an efficiency or reliability ranking

Every figure here comes from an operator’s own announcement or from a trade report of one, and none was measured independently. Equinix’s 2023 release states a plan, Meta’s 140 kW is a rating for a design that was then upcoming, and Google’s 99.999% is self-reported for CDUs rather than pods or sites. The KT Cloud facts come from Byline Network’s article because the company’s own press page did not load as text. None of the six reports a PUE, a water figure, a cost or a failure count, so they say nothing about which handoff is cheaper or more efficient to run. Qualified engineering review is required before any of this informs a design.

What they do support is a placement pattern. A facility loop makes the handoff liquid, and its absence makes it hall air, paid for in rack positions and hall capacity. Four checks sort the next announcement of this kind:

• Whether each figure is a rating, a plan or an operating value
• Where the rack loop’s heat ends: a facility loop, outdoor fans or hall air
• What a capacity or availability figure counts: one CDU, a fleet, a rack or a site feed
• Whether the hall’s per-rack limit changed or only moved to side racks

References

• AI infrastructure is hot. New power distribution and liquid cooling infrastructure can help / Google Cloud Blog (Madhusudan Iyengar, Amber Huffman) / 2025-04-29 / https://cloud.google.com/blog/topics/systems/enabling-1-mw-it-racks-and-liquid-cooling-at-ocp-emea-summit
• Inside the world’s most powerful AI datacenter / The Official Microsoft Blog / 2025-09-18 / https://blogs.microsoft.com/blog/2025/09/18/inside-the-worlds-most-powerful-ai-datacenter/
• Liquid Cooling in Air Cooled Data Centers on Microsoft Azure / Microsoft Tech Community, Azure Infrastructure Blog / 2024-10-15 / https://techcommunity.microsoft.com/blog/azureinfrastructureblog/liquid-cooling-in-air-cooled-data-centers-on-microsoft-azure/4268822
• Meta’s open AI hardware vision / Engineering at Meta / 2024-10-15 / https://engineering.fb.com/2024/10/15/data-infrastructure/metas-open-ai-hardware-vision/
• Equinix to Accelerate and Simplify Liquid Cooling Deployments to Power Enterprise AI Workloads / Equinix newsroom / 2023-12-11 / https://newsroom.equinix.com/2023-12-14-Equinix-to-Accelerate-and-Simplify-Liquid-Cooling-Deployments-to-Power-Enterprise-AI-Workloads
• KT클라우드, 액체냉각 채택 ‘가산 AI 데이터센터’ 개소 (KT Cloud opens Gasan AI data center with liquid cooling) / Byline Network, reporting a KT Cloud announcement / 2025-11-06 / https://byline.network/2025/11/6-283/
• How Meta achieves 120kW a rack in 20kW air-cooled data centers / Datacenter Dynamics / 2025 (day not shown on the page) / https://www.datacenterdynamics.com/en/news/how-meta-acheives-120kw-a-rack-in-20kw-air-cooled-data-centers/
• Hyperscalers prepare for 1MW racks at OCP EMEA; Google announces new CDU / Datacenter Dynamics / 2025 (day not shown on the page) / https://www.datacenterdynamics.com/en/news/hyperscalers-prepare-for-1mw-racks-at-ocp-emea-google-announces-new-cdu/