Shared vs Dedicated Heat Rejection in Hybrid Air-Liquid Data Centers

Every liquid-cooled AI hall is still an air-cooled hall. Sixty to ninety percent of a liquid-cooled server's heat leaves through the cold plates, but the remaining ten to forty percent — drives, NICs, DIMMs, VRMs, and every storage and switch rack in the building — still leaves as air [1]. So every hybrid facility answers one architectural question first: do the air side and the liquid side reject their heat through the same plant, or through two?

The industry's shorthand for the first option is a shared or common heat rejection source — in Chinese vendor literature, wind-liquid same source (“air-liquid shared source”); in ASHRAE's liquid-cooling architecture taxonomy, it is the “reject heat to existing heat rejection system” architecture [1][2]. The second option is a dedicated heat rejection system for the liquid loop, independent of the air side [1]. Both work. Both are deployed at scale. And their trade-offs are almost perfectly mirrored: the shared plant buys capital efficiency and pays in temperature compromise; the dedicated plant buys thermal freedom and pays in duplicated equipment.

This article defines both architectures precisely, works through the physics that decides between them — one number, the heat-rejection temperature ladder, dominates everything — quantifies the advantages and disadvantages of each, and ends with a decision framework by deployment scenario.

Two ways to reject one hall's heat

First, pin down the plumbing. A liquid-to-liquid (L-L) CDU isolates the clean IT-side technology cooling system (TCS) from the facility water. The question is what facility water feeds the CDU's heat exchanger, and whether the air-side CRAHs drink from the same loop [2][3].

Table 1 — The two architectures. Adapted from the ASHRAE-affiliated liquid cooling architecture guide and vendor deployment patterns [1][3][4].

A third pattern exists on the liquid side — a liquid-to-air CDU dumping its heat into the room for the CRAHs to pick up — but at AI-hall scale it merely borrows the air plant's capacity and airflow budget, and most facilities past a few megawatts move to L-L designs into facility water [1][2]. Note also why the CDU sits in the path at all: facility water flows through the microchannels inside a cold plate only at the cost of clogging risk, chemistry control inside IT hardware, and warranty exposure nobody wants [2]. The CDU is the isolation boundary between a plant you can treat loosely and a loop you must keep immaculately clean.

The temperature ladder decides everything

The entire shared-vs-dedicated decision reduces to one physics chain. A heat-rejection device — dry cooler or cooling tower — can only make water that is warmer than the outdoor air (dry cooler) or the wet-bulb temperature (tower), by the approach temperature it can hold. So the supply temperature your loop accepts sets the outdoor threshold below which you can run free cooling, refrigeration off [4][5].

The ladder from published design practice [1][4][5][6]. Legacy facilities ran chilled water at 6–10°C; modern liquid-cooled AI factories run 30–32°C, with 40°C and beyond under discussion [3].

The consequence is enormous. In the Delhi analysis that published this method, a conventional 6–12°C plant can free-cool only when outdoor temperature sits below roughly 1°C — about 150 hours a year — while a 35–40°C liquid loop free-cools below roughly 27°C — about 5,200 hours, or 59% of the year, in the same city [5]. The chiller that never idles in a legacy hall becomes a trim machine in a warm-water hall. In cold climates the gap widens to near-total: a Norway facility running 45°C supply on dry coolers reports 98% free-cooling utilization and PUE 1.08 [7], and Schneider's Green Mountain example achieves roughly 330 pre-cooling days per year [3].

ASHRAE's liquid-cooling supply classes formalize the ladder [4]:

Table 2 — ASHRAE liquid-cooling supply classes and what each implies for heat rejection [4]. Every class the loop climbs buys free-cooling hours; every degree a shared plant gives back sells them.

Shared source: the advantages

The shared architecture is the default for a reason. Its advantages are structural:

• Capital reuse. An existing chiller plant, heat-rejection equipment, pump house, and pipe infrastructure serve the new liquid racks; the CDU simply taps the facility water loop. The architecture guide's first-listed advantage is exactly this — reduced investment by using the existing heat rejection system [1]. One plant, one pump room, one pipe network instead of two.
• Dynamic capacity allocation. One plant sized for the whole hall can shift capacity between air and liquid branches as the mix evolves — the property Tencent and Shenling's joint design exploits by dynamically adjusting cooling output between the air and liquid sides against real load [6]. Under-built liquid share today, air-heavy tomorrow: the plant follows.
• Operational unity. One water-chemistry program, one controls/DCIM integration, one maintenance contract and spares pool. For an operations team, one plant is one thing to instrument, commission, and staff.
• Better part-load efficiency. A single large plant serving both sides aggregates the hall's whole load; machines run closer to their efficient range than two half-sized plants each loafing on their own partial load.
• A bridge for phased AI retrofits. This is the decisive advantage for the installed base. A hall can run air-cooled racks today with the plant it already owns, then add liquid pods as GPUs arrive — vendor programs are explicitly built around this progression (“air first, upgrade to liquid later” is the marketing line; the plant tap-in is the engineering substance) [8][9]. Huawei's wind-liquid fusion design — distributed air cooling plus CDU off shared infrastructure — frames the same bridge: one side handles general computing, the other absorbs AI racks, with the ratio adjustable [9].

Published results show the ceiling of a well-executed shared design. Tencent and Shenling's same-source architecture — a shared outdoor dry-wet integrated cooling tower, dry/wet modes switched intelligently, cooling output dynamically allocated between air and liquid branches — has run in a Tencent cloud data center for over a year at real-time PUE as low as 1.13 and annual PUE ≤1.15 [6]. Shenling's container-scale same-source deployment in Malaysia — liquid-led, air-assisted, in a hot tropical climate — holds PUE ≤1.15 year-round [10]. Alibaba Cloud's CUBE 5.0 makes mode-switchability the headline: the same cold-source module feeds air-cooled or liquid-cooled configurations at PUE ≤1.15 or ≤1.10 respectively [11].

Shared source: the compromises

The disadvantages are equally structural, and they all trace back to the ladder.

• The temperature compromise is quantitative, not rhetorical. If the shared plant is held at air-side legacy temperatures, the liquid side's free-cooling window collapses. Using the threshold arithmetic above with a coarse climate model: a Beijing-area hall with a dedicated 40°C liquid loop free-cools roughly 97% of the year; hold the shared plant at an 18°C warm-air-conditioning setpoint instead and the same hall's free-cooling window falls to roughly 50%; at legacy 7°C it approaches the conventional single digits. In a hot-climate hall (Dallas-like), the same compromise cuts the window from ~95% to ~22% [5].

Estimated free-cooling share of annual hours by climate type, dedicated 40°C loop vs a shared plant pinned to 18°C. Simplified normal-distribution climate model, 5 K approach — directional arithmetic, not a site study [5].

• The approach-temperature trap in retrofit design. Feeding a CDU from a shared plant at 18°C with a 3–4 K heat-exchanger approach means the TCS cannot deliver the inlet temperature a cold-plate vendor may have specified near plant supply. The documented procurement error is sizing the CDU for nameplate power and forgetting the approach — the fix is either lowering the facility setpoint (giving up chiller efficiency) or selecting cold plates rated for warm-water operation. Cheap to catch in design; a rework after commissioning [12].
• Compressor economics cut the wrong way. The chiller efficiency gain from warmer water is steep — raising supply from 7°C to 18°C lifts the theoretical COP from about 5.0 to about 7.8, a 56% reduction in compressor energy at equal lift assumptions. A shared plant pinned cold for the air side pays that penalty on every hour it mechanically cools the liquid side, hours the dedicated loop would have spent free cooling.
• Coupled failure domains. One plant serving both sides means a plant event — pump trip, controls fault, chemistry excursion — stresses both cooling paths at once. The plant is a chain of single systems that all have to hold, and a CDU's facility-water feed remains a single point of failure unless dual valved connections are provided; mitigation is real (redundant pumps and loops, dual feeds), but it is engineering the shared design must buy back, alongside control logic that coordinates two temperature zones and dynamic allocation from one source [12][14].
• Humidity handling on the air side. Warmer shared water shrinks the CRAH coil's margin below room dew point. A hall held at 18°C supply is already trading latent capacity; the moisture side then needs attention (dew-point control, dedicated dehumidification) that a 7°C coil handles incidentally. Our piece on dew-point calculation covers the arithmetic.

Dedicated source: the advantages

The dedicated architecture's case is the mirror image: each loop gets exactly the temperature it wants, from rejection equipment sized and selected for it [1].

• Thermal freedom on the liquid side. The ASHRAE-affiliated guide's dedicated example is the cleanest statement: a dry cooler with trim compressor supplying 40°C/104°F to the liquid loop while a free-cooling chiller supplies 20°C/68°F to the air side — “optimizes temperature and flow of the TCS and heat rejection loops in the most efficient way without the constraints imposed by a shared air-cooled heat rejection system” [1]. The liquid loop rides W4 economics — year-round dry-cooler free cooling in most climates — while the air side keeps its own economizer strategy.
• Chillerless is only reachable dedicated (mostly). Running an AI hall with no compressor plant at all — dry coolers year-round — requires the liquid loop to live at W4 temperatures in a climate that cooperates. That design does not survive a shared plant pinned to air-side temperatures; it is a dedicated-loop architecture by definition [3][7]. Hybrid chiller-plus-dry-cooling remains the most popular configuration precisely because pure chillerless is climate-gated [3].
• Failure isolation. Two plants are two failure domains. A fault on the air side does not touch the liquid loop's rejection capacity and vice versa — the fault-domain separation that large AI operators want when a single hall's availability is measured in GPU-hours [13].
• Phased build without retrofit economics. A new AI hall can energize the dedicated liquid rejection first (dry coolers land fast — see our container deployment timeline piece), with the air-side plant added when the general-compute halls are.
• Heat reuse becomes a design variable. A W4/W5 loop returns water at 45–55°C, warm enough for district heating or process heat — an option that only exists because the loop runs hot, and one that dies if the same water must also serve a 13°C CRAH coil [4][7].

Dedicated source: the costs

• Duplicated plant. Two rejection systems, two pump groups, two pipe networks, two controls integrations, two water-chemistry regimes — the liquid loop's facility side and the air side do not even share chemistry, since the TCS side is segregated at the CDU in any case but the two facility loops still carry separate treatment programs and monitoring [1][12].
• Floor space and civil works. Two plant spaces instead of one, two sets of penetrations and risers. In retrofit halls where space is already scarce, this is frequently the disqualifying constraint — which is why the shared architecture dominates the installed base [1].
• Part-load inefficiency. Two half-loaded plants each sit deeper on their inefficient end of the curve than one right-sized plant would. A hall whose liquid share grows to 80–90% leaves its dedicated air-side plant structurally underloaded — wasted capital idling at poor efficiency.
• Double operations. Two commissioning programs, two maintenance cycles, two spares pools, and a controls integration story twice as long. None of this is exotic; all of it compounds [12].

What the energy difference is worth

Published PUE bands put numbers on the stakes. Against the industry average of 1.56, flat for five consecutive years [13]:

Table 3 — Published PUE by architecture and climate. Moving 1 MW of IT from the industry average to a well-run 1.15 design is worth roughly 3.6 GWh a year — about $0.29M–0.54M annually at $0.08–0.15/kWh, whichever architecture delivers it [6][13].

Read the table honestly: architecture is not the only variable — climate is doing enormous work in the Norway row. What the shared and dedicated designs prove jointly is that the temperature ladder, climbed by either means, is where the energy lives. Tencent's shared tower climbs it by dry/wet mode switching and dynamic allocation; Norway's dedicated dry coolers climb it by running W4 supply. The designs that stay low on the ladder — whatever the topology — pay compressor hours all year.

A decision framework

Table 4 — Architecture selection by scenario. The pattern underneath: existing plant and uncertain mix pull toward shared; thermal freedom, climate luck, and isolation pull toward dedicated.

Frequently asked questions

What does “shared heat rejection” mean for air and liquid cooling?

Both the air side (CRAH coils) and the liquid side (CDU heat exchangers) reject heat into the same facility water plant — one chilled-water plant, one cooling tower or dry cooler array, one pump and pipe network. ASHRAE's taxonomy calls it the “reject heat to existing heat rejection system” architecture [1][2].

Is a shared plant always cheaper?

On capital, usually yes where a plant already exists — the reuse advantage is the architecture's headline benefit [1]. But the operating side can reverse it: a shared plant pinned to air-side temperatures sells back the liquid loop's free-cooling hours, and in warm climates that can be most of the year's compressor savings [5]. Total cost depends on whether the shared design climbs the temperature ladder (two-temperature gradient, dry/wet switching) or stays pinned cold.

Why do liquid loops want warmer water than air loops?

Cold plates transfer heat efficiently enough to tolerate warm inlet water, and warmer loop supply means the outdoor temperature threshold for free cooling rises — from roughly 2°C for a 7°C plant to roughly 35°C for a 40°C loop. Modern AI facilities run 30–32°C supply against legacy 6–10°C, with Meta publicly targeting around 29°C and 40°C under active discussion [3][4].

Can one plant serve both sides at different temperatures?

Yes — this is the state of the art of the shared design: a two-temperature gradient from one source, dynamic allocation between air and liquid branches, and dry/wet mode switching on the rejection side. Tencent and Shenling's deployed system runs real-time PUE as low as 1.13 this way; Huawei's fusion scheme runs primary 35°C/43°C and TCS 40°C/50°C [6][9]. The cost is control complexity — coordinating two temperature zones and dynamic allocation from one plant — which is the architecture's main engineering challenge [6].

Does a dedicated liquid loop mean no chiller at all?

Not necessarily — it means its own rejection system. The common dedicated design pairs dry coolers with a trim compressor for the hottest hours [1]. Fully chillerless operation is achievable when the climate cooperates with W4 supply temperatures — Norway reports 98% free-cooling utilization and 330 pre-cooling days per year at the best sites — but most regions still want mechanical trim [3][7].

What should I check first when retrofitting liquid racks into an air-cooled hall?

Three numbers: the plant's supply temperature and whether it can run warmer, the CDU heat-exchanger approach against your cold plates' specified inlet, and the pump curve's head at the CDU connection points — retrofit teams most often trip on the approach-temperature mismatch and on facility pumps sized for an even CRAH grid, not for concentrated, modulating CDU loads [12].

The trade in one sentence

Shared source converts capital you already own into cooling capacity and pays for it in degrees; dedicated source converts degrees into free-cooling hours and pays for it in duplicated steel. Every hall between those poles is a bet on which currency — capital or compressor-hours — is scarcer where you build, and the temperature ladder is the exchange rate between them.

Planning a hybrid hall's cooling architecture — shared plant tap-in, dedicated dry-cooler loop, or a two-temperature gradient in between? Request a quote and we will work the ladder arithmetic against your load curve and climate.

References

1. ASHRAE-affiliated guide, “Navigating Liquid Cooling Architectures for Data Centers with AI” (IS-2024-173): existing vs dedicated heat rejection architectures, 40°C/20°C dedicated example, 60–90% liquid heat capture, ashb.com (PDF)
2. Schneider Electric, “Optimising AI infrastructure: Liquid cooling” (EB3): CDU liquid-to-air vs liquid-to-liquid, why facility water cannot feed cold plates directly, download.schneider-electric.com (PDF)
3. Data Center Dynamics, “Q&A with Steven Carlini, Schneider Electric”: 6–10°C legacy vs 30–32°C AI factories, Meta 29°C target, Green Mountain 330 pre-cooling days, chiller + dry cooling as popular configuration, datacenterdynamics.com
4. Reynolds & Bauhm, “HPC & Supercomputing Cooling — Warm Water”: ASHRAE liquid classes W1–W5, free-cooling and approach definitions, heat-reuse return temperatures, reynoldsbauhm.co.uk
5. KVRM Engineering, “Data Centre Water Usage Effectiveness: Cooling Tower vs Adiabatic vs Dry Coolers”: free-cooling threshold method, Delhi ~150 h/yr at 1°C threshold vs ~5,200 h/yr (59%) at 27°C threshold, kvrm.in
6. ChinaIOol, “Shenling & Tencent joint same-source elastic cooling wins 2025 CDCC award”: shared dry-wet integrated tower, dry/wet mode switching, dynamic air-liquid allocation, real-time PUE 1.13 / annual ≤1.15, deployed >1 year, chinaiol.com (Chinese)
7. Boyi Cooling, “Maximize Efficiency with Dry Coolers as Natural Cooling Systems”: 40–45°C supply designs, 5 K approach, Norway 98% free-cooling utilization at PUE 1.08, boyicooling.com
8. Envicool, “High-elasticity air-liquid fusion cooling for high-density compute” (CDCC 2025 talk): air-liquid compatible design, “air first, upgrade later” phased path for existing halls, envicool.com (Chinese)
9. Digital Economy Observer, “Huawei wind-liquid fusion for AI compute centers”: distributed air + CDU shared architecture, primary 35°C/43°C, TCS 40°C/50°C, Gui'an PUE 1.1, szw.org.cn (Chinese)
10. Shenling Environment, data center solutions: Malaysia MY06 container project, liquid-led air-assisted same-source design, PUE ≤1.15 in tropical climate, shenling.com (Chinese)
11. Science Net (China Science News), “Alibaba Cloud's fully modular data centers: 100-day delivery”: CUBE 5.0 air/liquid switchable cold source, air PUE ≤1.15, liquid ≤1.10, portal.nstl.gov.cn (Chinese)
12. Mislier, “Chilled Water vs Liquid Cooling: How the Two Layers Work Together”: retrofit patterns, the approach-temperature procurement trap, CDU load profile vs CRAH pumping, mislier.com
13. Data Center Knowledge, “Data Center Industry Survey Highlights Cost, AI, and Sustainability Challenges” (Uptime Institute 2024 Global Survey: average PUE 1.56, fifth year flat), datacenterknowledge.com
14. Anvilfield, “Data center cooling system types and how to select one”: chilled-water plant as shared engine, warm-water loops on dry coolers with chiller trim, DX and small-site economics, anvilfield.com