Liquid vs Air Cooling for Australian Data Centres
| Dimension | Air cooling | Direct-to-chip liquid (cold plate) |
|---|---|---|
| Practical rack density | ~8–25 kW per rack | Well above 40 kW per rack |
| AI GPU support (GB200/H200) | Inadequate above ~40 kW/rack | Required — cold plate on the die |
| Cooling efficiency | Lower; large fan/CRAC energy | Higher; lower PUE, less fan energy |
| Retrofit | Limited for high-density AI | Cold-plate kits retrofit GPU/CPU trays |
| Acoustic / space | Noisy, space-hungry at density | Quieter, compact at the rack |
| AU deployment | Legacy halls, lower-density colo | AirTrunk MEL2, NEXTDC AI Factory & new AI halls |
Why AI workloads force liquid cooling
Air cooling is fine for traditional racks up to roughly 8–25 kW. AI GPU clusters built on platforms like the NVIDIA GB200 NVL72 run at rack densities far beyond what air can remove in the available space. Above about 40 kW per rack, direct-to-chip liquid cooling (a cold plate mounted on the GPU/CPU die) becomes the practical requirement, delivering higher efficiency and lower PUE.
Liquid cooling vs air cooling for high-density racks: where is the kW-per-rack line?
| Rack density | Typical load | Cooling approach that fits | ToneCooling scope |
|---|---|---|---|
| Up to ~8–25 kW | Enterprise, colocation, storage and network | Room air (CRAC / CRAH) with hot- or cold-aisle containment | — (air-cooled) |
| ~25–40 kW | Mixed GPU and CPU rows | Air plus containment, or a rear-door heat exchanger | Rear-door heat exchangers |
| ~40–80 kW | GPU training and inference racks | Direct-to-chip cold plates and a CDU; residual heat to air or RDHx | GPU and CPU cold plates, CDU |
| ~100 kW+ (AI GPU racks) | GB200 NVL72 and GB300-class racks | Direct-to-chip liquid cooling by design, with a facility water loop | GB300 4G+2C kit, GB200 kit, CDU |
Planning bands compiled from the ranges on this page (air ~8–25 kW, liquid above ~40 kW, AI factory halls at ~100 kW per rack, as cited above). Confirm the final design against the server vendor’s thermal guide and the ASHRAE TC 9.9 thermal guidelines.
How many more GPUs can a rack support with liquid cooling?
Rack power, not floor space, is usually the limit. Using the ranges above, a liquid-cooled AI rack at ~100 kW carries about four times the power of an air-cooled rack at up to 25 kW (calculated from the figures on this page, not measured), so the same row can host several times more GPUs once the CDU and facility loop are sized for that heat.
Immersion cooling vs direct-to-chip cold plates: which suits an Australian AI data centre?
| Factor | Direct-to-chip cold plates | Immersion cooling (single- or two-phase) |
|---|---|---|
| How heat is removed | Cold plate on each GPU / CPU; remaining components cooled by air or a rear-door heat exchanger | Whole server sits in dielectric fluid; server fans are removed |
| Server hardware | OEM liquid-cooled trays (GB200, GB300-class) in standard vertical racks | Immersion-rated servers; check the OEM warranty for fluid contact |
| Facility | Fits existing raised-floor or slab halls; CDU per row or rack | Horizontal tanks; check floor loading and fluid containment |
| Coolant | Water-glycol (for example PG25) in a closed secondary loop — see the coolant selection guide | Hydrocarbon or engineered dielectric fluids; some two-phase fluids are fluorinated, so check your PFAS policy |
| Servicing | Quick disconnects; trays swap much like air-cooled servers | Servers are lifted out of the fluid; drip and handling procedures needed |
| Australian heat and water | A warm secondary loop suits dry coolers, which use no evaporative water; size them for local design-day temperatures (water use guide) | Heat also leaves through a heat exchanger to the facility loop, so the same dry-cooler sizing question applies |
| ToneCooling scope | Cold plates from our own manufacturing facility in Huizhou; CDUs, RDHx and quick disconnects configured to order | Tanks and dielectric fluids are not supplied |
When to choose direct-to-chip cold plates
- You are deploying GB200, GB300-class or other OEM liquid-cooled trays.
- The hall keeps standard racks, aisles and the existing service model.
- Storage and network gear can stay on air while GPU racks go liquid.
- You are retrofitting an existing hall by adding a CDU and rack manifolds.
When to choose immersion cooling
- The servers are immersion-rated and the vendor supports fluid contact.
- The site is new, with floor loading and containment designed for tanks.
- Your operations team accepts dielectric fluid handling and servicing procedures.
- Fanless, low-noise operation is a priority.
What does a direct-to-chip load look like on a spec sheet?
Measured example from our GB300 4G+2C cold plate kit (4 GPU + 2 CPU), steady state on pure water:
| Plate | Flow | Heat load | Inlet / outlet | Pressure drop | Thermal resistance |
|---|---|---|---|---|---|
| GPU cold plate | 1.7 L/min | 1,066.94 W | 19.9 / 27.3 °C | 12.06 kPa | 0.0130–0.0133 °C/W |
| CPU cold plate | 0.7 L/min | 485.46 W | 23.6 / 31.5 °C | 10.39 kPa | 0.0220 °C/W |
Source: ToneCooling datasheet TC-BZP-GB300-01A, test dated 20 March 2026, thermal interface material TC-5888.
What does liquid cooling in data centres need beyond the cold plate?
- Coolant distribution units — in-rack and in-row; the Australian and New Zealand electrical-safety pathway is set per project at quotation.
- UQD quick disconnects and hose assemblies for tool-free tray service.
- Rear-door heat exchangers for the residual air-side heat.
- Closed-circuit cooling towers where the site allows evaporative heat rejection.
Cold plates come from our own manufacturing facility in Huizhou (around 1,000,000 cold plates per year, IATF 16949-aligned quality system). CDUs, rear-door heat exchangers, quick disconnects and cooling towers are supplied by ToneCooling from China and configured to order, with scope agreed per project. Cold plate samples typically take 4–6 weeks including design and simulation; CDU prototypes take 8–10 weeks; volume lead times are confirmed at quotation.
The Australian data-centre buildout is going liquid
Australia is in a multi-billion-dollar AI data-centre buildout — Microsoft (A$25bn), AWS (A$20bn) and others. Operators are designing for liquid at the core: AirTrunk's MEL2 (354 MW) integrates large-scale direct-to-chip liquid cooling, and NEXTDC's M4 'AI Factory' runs liquid cooling at ~100 kW per rack. New AU AI halls are being built liquid-ready.
ToneCooling Australia supplies direct-to-chip cold plates
ToneCooling Australia supplies direct-to-chip liquid cold plates for Australian data-centre and HPC programmes — GB200, GB300, H200, AMD EPYC SP5, Intel Birch Stream — designed, engineered and manufactured at our own IATF 16949-aligned manufacturing facility in Huizhou, China, with customer engineering support from Carlton, Victoria. See our data centre liquid cooling and GPU cold plate pages.
Frequently asked questions
When does a data centre need liquid cooling instead of air?
Air cooling is practical up to roughly 8–25 kW per rack. AI GPU racks (e.g. NVIDIA GB200 NVL72) run far above that range, so above about 40 kW per rack direct-to-chip liquid cooling with cold plates becomes necessary.
Is liquid cooling more efficient than air cooling?
Yes — direct-to-chip liquid cooling removes heat at the die with far less fan/CRAC energy, lowering PUE and enabling much higher rack density than air cooling.
Are Australian data centres switching to liquid cooling?
Yes. Major AU operators are building liquid-ready AI capacity — AirTrunk's MEL2 uses large-scale direct-to-chip liquid cooling and NEXTDC's M4 AI Factory runs ~100 kW per rack with liquid cooling.
Can existing air-cooled racks be retrofitted to liquid?
High-density AI deployments are best designed liquid-ready, but direct-to-chip cold-plate kits can retrofit GPU/CPU trays where the facility supports a liquid loop and CDU.
Does ToneCooling supply data centre cold plates in Australia?
Yes. ToneCooling Australia supplies direct-to-chip cold plates for GB200/GB300/H200, AMD SP5 and Intel Birch Stream, with customer engineering support and project documentation coordinated from Carlton, Victoria.
Is immersion cooling better than direct-to-chip liquid cooling?
Neither is better everywhere. Direct-to-chip cold plates suit OEM liquid-cooled GPU trays such as GB200 and GB300-class platforms in standard racks; immersion suits immersion-rated servers in halls designed for tanks. AirTrunk MEL2, referenced on this page, uses direct-to-chip liquid cooling.
How many more GPUs can a rack support with liquid cooling compared to air cooling?
Using the ranges on this page, a liquid-cooled rack carries about four times the power of an air-cooled rack (~100 kW against up to 25 kW; calculated, not measured), so a row can host several times more GPUs once the CDU and facility loop are sized for it.
Does ToneCooling supply immersion cooling tanks?
No. ToneCooling Australia supplies direct-to-chip cold plates, including GB300 4G+2C and GB200 kits, plus CDUs, rear-door heat exchangers and quick disconnects for Australian and New Zealand projects. An engineer reviews every RFQ and replies by email with next steps.
Discuss your data centre liquid cooling project
Send your rack heat load, server platform and site location to the Carlton customer engineering support team — an engineer reviews every RFQ and replies by email with next steps. Standard kits: GB300 4G+2C cold plate kit and GB200 cold plate kit. Custom samples are available from a single piece.
Request a QuoteSources and standards referenced: ASHRAE TC 9.9 Mission Critical Facilities · Open Compute Project – Cooling Environments · Bureau of Meteorology climate data · ToneCooling datasheet TC-BZP-GB300-01A (20 March 2026).

