Choose by the power each rack will draw, not by the server brochure. Most server rooms still run below 10 kW per rack, and well-managed air cooling handles that. Air with good containment can stretch to one 8-GPU server at around 14 kW. From roughly 20 kW per rack, rear-door heat exchangers or direct-to-chip liquid start to make sense, and both need water piped to the rack. At the top end, a GB300 NVL72 rack draws about 132–142 kW and is liquid-cooled by design, with a coolant distribution unit (CDU) and facility water. If your plan includes that tier, the cooling decision is really a facility decision, and in Singapore that usually means a liquid-ready colocation site.
Which cooling fits which rack load?
| Load per rack | Typical AI hardware | Cooling that fits | What the room needs |
|---|---|---|---|
| Under 10 kW | Storage, networking, one or two PCIe GPU servers | Room air (CRAC/CRAH), blanking panels, containment | Standard server room |
| 10–20 kW | One 8-GPU HGX B300 server (about 14 kW) | Air with hot/cold aisle containment; rear-door exchanger as headroom | Strong airflow, 3-phase power |
| 20–80 kW | Several GPU servers per rack (NVIDIA's B300 reference design: four per rack, about 56 kW) | Rear-door heat exchanger (vendor-rated up to 80 kW) or direct-to-chip liquid | Chilled or warm water to the rack |
| 132–142 kW | GB300 NVL72 rack | Direct-to-chip liquid with CDU; residual heat to air | Facility water, 400/480 V feeds, heavy-floor rack space |
Source: Uptime Institute Global Data Center Survey 2026; NVIDIA DGX SuperPOD B300 reference architecture; Vertiv (18 Mar 2025) and Dell (19 May 2025) rear-door ratings (vendor claims); GB300 NVL72 ratings from HPE, Supermicro, GIGABYTE, Vertiv and Schneider. Band boundaries are SourceIT's rule of thumb, not a standard.
Where are most server rooms today?
Lower than AI headlines suggest. In Uptime Institute's 2026 survey (fieldwork March to May 2026), 69% of operators said their most common rack density was under 10 kW, 20% said 10–19 kW, 8% said 20–49 kW and 3% said 50 kW or more. Uptime's July 2026 release noted "a growing number of operators now reporting peak rack densities of 30 kW or higher", but dense racks are still the exception.
That matters for SMEs. If your AI plan is one or two GPU servers, you are adding a hot rack to a mostly air-cooled room, not building a liquid-cooled hall. Our guide on why an AI data centre needs more than GPUs covers the power side of the same question.
How far can air cooling go?
Air-cooled 8-GPU servers are real products: Supermicro sells an air-cooled 8U HGX B300 system, and Dell offers both air and liquid versions. One such server draws about 14 kW, which a rack with good containment, blanking panels and sufficient room cooling can handle.
Vendors are clear that air runs out of road above that. Vertiv says racks are moving beyond 20 kW towards 50 kW and more, and "traditional air cooling can't cool these high-density racks efficiently and sustainably". Schneider Electric's white paper on AI cooling architectures says air cooling "becomes less practical when considering heat sink dimensions, server airflow and energy efficiency", though it sets no fixed kW threshold. Treat both as vendor positions, but the physics is not in dispute: more heat per rack needs a medium that carries more heat than air.
Basic airflow discipline still pays at any density. Blanking panels, sealed cable cut-outs and temperature sensors at the rack inlet cost little. For monitoring the power side, see the silent danger of overloaded server racks.
What does a rear-door heat exchanger add?
A rear-door heat exchanger (RDHx) replaces the back door of the rack with a water-cooled coil. Hot exhaust passes through it and leaves the rack close to room temperature. Vertiv's chilled-water RDHx is rated "up to 80kW per rack"; Dell's enclosed rear-door system claims "air cooling capacity up to 80 kW per rack" with facility water at 32–36 °C. Both are vendor ratings.
The appeal is that the servers stay air-cooled, so you can keep standard equipment and mixed rack densities in one room. The catch is that water still has to reach the rack, and the design must handle heat while a door is open for service. Note too that Singapore's new SS 726:2026 liquid-cooling standard covers direct liquid cooling and excludes rear-door, two-phase and hybrid systems.
When is direct-to-chip liquid cooling mandatory?
When the hardware is designed for it. A GB300 NVL72 rack is rated at about 132 kW by HPE and Supermicro and up to 142 kW by Vertiv and Schneider, with HPE advising 192 kW busway provisioning. It uses cold plates on the chips, a CDU and facility water; there is no air-cooled version. We explain what the rack contains in GB300 NVL72's 37TB of memory.
Direct-to-chip doesn't remove all heat. Vertiv says it removes about 70–75% of rack heat, with the rest going to air. Supermicro claims its newer design captures up to 98%. CDUs come in different forms, which Schneider groups into liquid-to-air units that dump heat into the room and liquid-to-liquid units that connect to facility water. Supermicro also sells a 200 kW liquid-to-air sidecar that needs "no facility water", though the room's air cooling must then absorb that heat.
Cool the rack you will actually have. Below about 20 kW, good airflow is usually enough. Above about 50 kW, plan for water at the rack. At NVL72 scale, liquid is part of the product.
What does liquid cooling cost to run and maintain?
The energy case rests on warm water. Direct liquid cooling can run with water at 40–45 °C, warm enough to reject heat without chillers or compressors in many climates, which cuts the facility overhead that PUE measures. Vendor savings claims are large but use their own baselines: Supermicro claims up to 40% data-centre savings against air, HPE claims a 90% cut in cooling power against traditional air, and Dell claims up to 60% lower cooling energy for its rear-door system at 36 °C versus 20 °C water. Ask any vendor what the comparison assumes.
Maintenance is the part buyers underestimate. Vertiv recommends leak detection and drip pans under all fittings, and rack integrators test manifolds, CDUs and leak detection before shipping. Coolant change intervals, filter service and pump life are not published consistently, so get the service schedule in writing and agree who owns it: the colocation provider, the CDU vendor or your team. Workstation liquid cooling is a different problem, covered in liquid-cooled workstation maintenance.
What can an SME server room in Singapore actually cool?
Start with power, because it caps everything else. SP Group's single-phase supply goes up to 23 kVA (100 A), and larger loads need 400 V three-phase. A GB300 NVL72 needs 400/480 V 60 A feeds per HPE, so it is not an office build. A deskside DGX Station is a different story; see putting a DGX Station under an office desk.
For anything above one or two GPU servers, colocation is usually the practical route. CBRE puts Singapore pricing at about US$330–475 per kW per month, averaging about US$403, with vacancy around 2%. As a CALC at the average rate, one 14 kW HGX server costs about US$5,600 a month to host, and a 132–142 kW NVL72 rack about US$53,000–57,000 a month. Singapore's Green Data Centre Roadmap targets PUE of 1.3 or better at full IT load, and SS 726:2026 (August 2026) sets out how to plan and run direct liquid cooling in tropical conditions, including fluid temperatures above 40 °C. Liquid-ready capacity exists, for example Nxera's direct-to-chip DC Tuas. For the path from one server to an AI factory, see from RTX PRO 6000 to GB300 NVL72.
Buying in Singapore
We don't list cooling equipment: rear-door heat exchangers, in-row coolers, CDUs and containment are sourced to order. Ask us to source through our partners; we will confirm availability, lead time and facility requirements rather than assume stock. For the air-cooled rack around your GPU servers, we list APC 1U blanking panels (pack of 10) at S$102.50, the APC AP9335T temperature sensor at S$153.75, the APC Network Management Card 3 at S$596.25, the APC NetShelter SX 42U rack at S$2,737.50 and the Eaton G4 managed three-phase 22 kW PDU at S$3,502.80 (listed 9 Nov 2026). The passively cooled RTX PRO 6000 Server Edition is S$32,999 and needs server airflow. GPU servers such as the Lenovo ThinkSystem SR680a V3 and HPE ProLiant DL380a Gen11 are price on request. Prices change; ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can an 8-GPU AI server be air-cooled? Yes. Air-cooled HGX B300 servers exist, and one draws about 14 kW, which a well-contained rack can handle. Several per rack push you towards rear-door or liquid cooling.
At what rack density do you need liquid cooling? There's no single official threshold. Vendors say air struggles beyond about 20 kW per rack, and GB300 NVL72 racks at 132–142 kW are liquid-cooled by design.
Is a rear-door heat exchanger liquid cooling? It uses water at the rack, but the servers stay air-cooled. Singapore's SS 726:2026 standard covers direct liquid cooling and excludes rear-door systems.
Does liquid cooling lower running costs? It can, mainly by allowing warm-water cooling without chillers. Vendor savings claims of 40–90% use their own baselines, so ask what each comparison assumes.
Can a Singapore office server room host a GB300 NVL72? Generally no. It needs about 132–142 kW, 400/480 V feeds and facility water, which points to a liquid-ready colocation site.
