Coolingthe cloud
A concept design proposal: a warm chilled water plant that rejects heat to the air and uses water only when the community can spare it
- Facility
- 20 MW IT, AI
- Location
- Western Sydney, NSW
- TCS supply
- 30 °C
- Stage
- Concept design
Concept design · 2026 · James Gianoutsos
01 — Summary
Reject heat to the air. Use water only when the community can spare it.
This proposal sets out a cooling plant for a 20 MW liquid-cooled AI data centre in Western Sydney, NSW. It supplies water at 30 °C from dry coolers, falls back to a high-temperature chiller, and treats adiabatic assist as a third mode that the local water supply can switch off.
83%
of the year on free cooling: dry coolers alone, no chiller and no water.
−94%
water against an evaporative cooling tower plant: about 15 ML a year instead of 254 ML.
0 ML
in a drought year. With restrictions in force, the plant runs entirely dry.
+1.2%
facility energy for that saving: annualised PUE 1.09 against 1.07 for the tower plant.
Indicative · concept-stage estimate. Hourly model of a synthetic Western Sydney year; method and assumptions under Indicative performance
Recommendation
Adopt the three-mode warm water plant, with the adiabatic mode permitted only while the community water system is unrestricted, and take it to an hourly design study on measured weather data.
A cooling tower plant is the cheapest to run, but it evaporates water every hour of the year, and most of all on the hottest days, when the local supply is under the most pressure. A plant with no water at all avoids that, but pays for it in chiller energy through every warm afternoon.
The proposed plant sits between them. It keeps most of the tower plant’s efficiency, uses a small fraction of its water, and hands the decision to use that water to the state of the supply it draws from.
02 — Design brief
A hot site, a constant load, a shared water supply.
The client and site are hypothetical, and chosen to be difficult: a warm, humid climate where dry cooling is least proven, and a load that never lets up.
Brief
- Facility
- AI training and inference data centre (hypothetical client)
- Site
- Western Sydney, NSW
- IT load
- 20 MW, near-constant. AI load can’t be shifted to another hour or another site.
- Cooling
- Direct-to-chip liquid cooling. supply 30 °C (ASHRAE S30).
- Redundancy
- N+1 on heat rejection and on chillers
- Design ambient
- 47 °C . Penrith recorded 48.9 °C in January 2020.
- Water
- Potable supply from the local utility, subject to drought restrictions
Design criteria
- 01Water. Minimise potable water use, and use none at all while the community supply is restricted.
- 02Reliability. Hold 30 °C supply on the design day with any one unit out of service.
- 03Energy. Stay within a few per cent of a cooling tower plant’s annual .
- 04Compliance. Meet the Commonwealth expectation of sustainable water use, and the NSW requirement to reduce water use under drought restrictions.
- 05Future-proofing. Keep working as the climate warms and as next-generation hardware asks for cooler water.
03 — Context · the water problem
Cooling towers drink, and they drink most when water is scarcest.
Cooling is the overwhelming share of a data centre’s on-site water use, and the conventional plant, the evaporative cooling tower, consumes water by thermodynamic necessity.
5–40
per day sought by individual data centre applications to Australian water utilities, up to about 20 times the draw of the largest existing customer.
Daily water demandML / day
Source · Water Services Association of Australia (2025); Han et al. (2026). The largest-customer figure is implied: 1/20 of the upper request.
1.4–1.5 L/kWh
Evaporated for every kilowatt-hour of heat a tower rejects, before drift and blowdown. The heat leaves as the latent heat of the vapour, so the only way to save the water is to change the mechanism.
Evaporated per day
696,000 L
Equivalent to
0.28
Olympic pools per day
MW × 1,000 × 24 h × 1.45 L/kWh. Starts at this brief’s 20 MW.
Source · ASHRAE (2020); Mytton (2021)
A year’s average can’t see a heatwave. Supply fails on coincidence, not on total draw.
The reporting metric, , averages over a whole year. A plant that draws heavily through a heatwave reports the same figure as one that draws steadily, but the impact lands on one local supply system, at the moment it is most stretched. This design is judged on when it uses water, not only on how much.
Facility A draws the same every month and never exceeds the local headroom.
Illustrative profiles to show the concept. Not measured data.
Source · Concept after Li et al. (2025); Han et al. (2026); Lei & Masanet (2022)
04 — Context · floor and ceiling
The climate raises the floor. The hardware lowers the ceiling.
ASHRAE’s warm-water classes (W27, S30) let IT equipment accept warmer supply water, which is what makes dry heat rejection possible. But the next generation of accelerators needs cooler fluid, not warmer.
Source · Shape after Vertiv & Open Compute Project (2026). Indicative, not to scale
30 °C
The coolant target silicon and system vendors have converged on, driven by accelerator . Running warmer isn’t free on the IT side either: server power rises about 8.5% as coolant goes from 30 to 50 °C.
Source · ASHRAE TC 9.9 (2021); Vertiv & Open Compute Project (2026); Stahlhut et al. (2025)
Meanwhile the hours in which heat and humidity limit have risen over forty-five years of observations, and are projected to keep rising. A dry cooler can only hold its setpoint while the air is cooler than the water.
The ceiling
Hardware sets the warmest supply water it will accept, and that ceiling is heading down, towards 30 °C.
The floor
Climate sets the lowest temperature a dry cooler can reach. With rising heat, humidity and local heat build-up, that floor is heading up.
The window
Between them is the free-cooling window. A dry cooler holds a 30 °C loop only while the dry-bulb sits several degrees below it.
The crossing
Once the floor rises above the ceiling, mechanical cooling is no longer an efficiency option. It becomes a requirement. The literature doesn’t resolve what happens where they cross.
Source · Conceptual diagram. Direction of trends: Karamperidou et al. (2026); Vertiv & Open Compute Project (2026); Han et al. (2026)
Design response
Fix the setpoint at 30 °C
The durable target vendors have converged on. Designing for warmer water would buy free-cooling hours the next hardware generation takes back.
Keep a chiller
A chiller-less plant only works where the air is reliably cooler than the supply. In Western Sydney it isn’t, and the gap is widening.
Size for the hot tail
Plant is sized for a 47 °C design day, close to the 48.9 °C recorded at Penrith, so a warming climate eats into margin before it eats into capacity.
05 — Options appraisal
Five ways to reject 20 MW. One keeps both the water and the energy.
Each option is scored against the brief. The towers win on energy and lose on water; a plant that never uses water wins on water and pays in chiller energy. The proposal takes the water only when it is cheap for the community to give.
| Option | Annual water | Water in drought | Energy | Holds 30 °C year-round | Compliance |
|---|---|---|---|---|---|
| A · Evaporative cooling towersThe baseline. 254 ML/yr, heaviest in heatwaves. PUE 1.07. | – | – | ✓ | ✓ | – |
| B · Towers + fleet workload shiftingMoves load to less stressed regions. AI training and inference are near-constant, so little can move. | ◐ | ◐ | ✓ | ✓ | ◐ |
| C · Chiller-less dry coolingNo water, lowest plant cost, but can’t hold 30 °C for 17% of the year. | ✓ | ✓ | ✓ | – | ◐ |
| D · Dry coolers + chiller, never wetNo water at all. Chiller runs 1,481 h/yr. PUE 1.11. | ✓ | ✓ | ◐ | ✓ | ✓ |
| E · Proposed: dry + chiller + gated adiabatic15 ML/yr in a normal year, none in a drought. PUE 1.09; becomes D under restrictions. | ◐ | ✓ | ✓ | ✓ | ✓ |
✓ meets the criterion · ◐ partly · – does not.
Indicative · concept-stage estimate. Water, chiller hours and PUE from the model under Indicative performance
Why not move the load?
Stress-aware scheduling cuts stress-adjusted water use by up to a quarter across a fleet, but it needs load that can wait or travel. A facility running AI training or inference can’t shift its heat. Its plant has to decide.
Source · Talukder et al. (2026)
Why not switch on the plant’s own needs?
Cooling-mode switching is established, but the switching signal has always come from inside the facility. Here it comes from the community’s water system, so the plant stops drawing water at the moment the supply is restricted.
Source · Gnibga et al. (2024)
Precedents
The pieces of the proposal have each been built and measured somewhere. None of the published work combines them at a 30 °C setpoint, in a warm, humid climate, with the switch at the plant.
| Work | Non-evaporative (dry) rejection | 30 °C TCS with elevated LCWT | Warm, humid climate like Sydney | Mode switching at the plant | Driven by community water state |
|---|---|---|---|---|---|
| Karimi et al., 2022Chiller-less warm-water plant, hot-arid climate | ✓ | ◐ | – | – | – |
| Jin et al., 2024Water-side economisation over 8,760 h, cooling towers at conventional CHW temperatures | – | – | – | ✓ | – |
| Gnibga et al., 2024Cooling-mode flexibility, internal switching criterion | ◐ | – | – | ✓ | – |
| Talukder et al., 2026Stress-aware workload scheduling at the fleet layer | ◐ | – | – | – | ✓ |
| This proposalThree-mode warm chilled water plant, Sydney, hourly | ✓ | ✓ | ✓ | ✓ | ✓ |
✓ addressed · ◐ partly · – not addressed.
06 — The proposed plant
Three modes. Water only when the community can spare it.
A warm chilled water skid supplying water at 30 °C. It rejects heat through dry coolers, engages mechanical cooling only when it has to, and treats adiabatic assist as a third mode that is triggered by the state of the local water system.
Community water system
Active mode
Free cooling
Dry coolers hold 30 °C unaided. The chiller is bypassed and no water is used.
Thresholds are concept-stage assumptions (5 K dry-cooler approach, adiabatic band to 32 °C), the same ones the performance estimate uses. The summer-day profile is illustrative.
Mode 1
Dry coolers + water-side free cooling
The chiller is bypassed whenever ambient conditions let the dry coolers hold the supply setpoint unaided.
Mode 2
High-temperature chiller
Mechanical cooling only when free cooling can’t hold the setpoint, e.g. Carrier 19MV4 with leaving condenser water to 55 °C.
Mode 3
Gated adiabatic pre-cooling
A third mode, gated on the state of the local water system rather than only on the plant’s needs.
Equipment schedule · indicative
| Item | Duty | Installed | Basis |
|---|---|---|---|
| Adiabatic dry coolers | 18 × 1.4 MW | 19 (N+1) | Sized for 24 MW: IT load plus compressor heat at the design day. Pre-cooling pads or mist, isolatable per unit. |
| High-temperature chillers | 5 × 4.5 MW | 6 (N+1) | Two-stage centrifugal, magnetic bearings, leaving condenser water up to 55 °C (e.g. Carrier AquaEdge 19MV4). |
| TCS heat exchangers and pumps | 20 MW | N+1 | Separates the 30 °C technology cooling loop from the facility water loop. |
| Adiabatic water supply | Break tank | 1 | Potable make-up behind a motorised isolation valve, closed by the water-state gate. |
Indicative · concept-stage sizing. Unit capacities are placeholders until equipment is selected on the design day.
The dry coolers are sized for the chiller’s full heat of rejection, 24 MW, so the same coils serve free cooling in winter and condenser duty in a heatwave. Nothing on the rejection side evaporates water unless the gate allows it.
07 — Heat rejection · modes 1 and 3
Dry by default. A little water, only when it counts.
Rejecting heat without evaporation is established practice, from IBM’s Aquasar to instrumented chiller-less plants. Adiabatic pre-cooling extends it into the hours a dry coil alone can’t reach.
Water
−66%
Measured of a chiller-less warm-water plant against conventional evaporative plant, in a hot, arid climate.
Energy
+13%
The price paid there: roughly 13% higher . The proposal keeps that price down by pre-cooling when water is available instead of running harder.
Source · Karimi et al. (2022); Zimmermann et al. (2012)
Dry coolers, not hybrid towers
A hybrid tower still wets its fill by default. A dry cooler is dry by default and wet only by exception, which is the order the brief asks for.
Pre-cooling on the inlet air
Water is sprayed or misted into the air entering the coil, never onto the coil, so the heat exchanger stays dry and scale-free, and each unit can be isolated.
A 5 K approach
Coils sized so the dry coolers hold 30 °C supply unaided up to 25 °C ambient. A closer approach buys free-cooling hours at the cost of coil area and fan power.
Example ambient: 35 °C, humid-hot afternoon
Entering air
35.0 °C
up to −9.4 K with spray
Condensing temp.
0.0 K
reduction, up to 7.2 K
Mist approach
0.5 K
to wet-bulb, at best
The reductions are the maxima reported, and depend on how far the wet-bulb sits below the dry-bulb. Humid air leaves less room.
Source · Yang et al. (2021)
assist still uses water, about 0.5 L per kWh of heat while it sprays, against 1.45 L for a tower every hour of the year. That is why it is the plant’s third mode, not its default, and why the decision to use it belongs to the water system rather than to the plant.
08 — Mechanical cooling · mode 2
When the air isn’t enough: the chiller.
Past the crossing, the plant needs mechanical cooling. The design uses a high-temperature, two-stage centrifugal chiller with magnetic bearings, able to deliver up to 55 °C.
Overview
A heat pump for the whole plant
When the dry coolers can’t hold 30 °C on their own, the chiller takes the heat out of the IT loop and lifts it to a temperature the dry coolers can reject. Follow the refrigerant round the loop as you scroll.
Why 55 °C matters
A conventional chiller rejects heat at temperatures that need a cooling tower. Leaving the condenser at up to 55 °C keeps the rejection side dry, so even Mode 2 uses no evaporative water.
Source · Carrier (n.d.), AquaEdge 19MV4
The price: energy
Every hour the chiller runs costs compressor power. The annual chiller energy penalty of a low TCS setpoint is severe in hot climates and negligible in cold ones, which is why the hours it must run in Sydney matter.
Source · Vertiv & Open Compute Project (2026)
The chiller is the plant’s fallback, not its default. In the modelled year it runs 359 hours while water is available, helped by pre-cooled air, and 1,481 hours when drought restrictions lock the adiabatic mode out. Those extra hours are the energy the plant spends to leave the community’s water alone.
09 — Control strategy
The community’s water system holds the switch.
The plant chooses its mode from the weather. Whether it may use water is decided by a separate gate that watches the local supply, and fails closed.
Sequence of operations · concept
| Mode | Enters when | Leaves when | Plant |
|---|---|---|---|
| 1 · Free cooling | Dry-bulb ≤ 25 °C | Dry-bulb > 25 °C for 15 min | Chillers bypassed. Dry cooler fans modulate on 30 °C supply. Sprays off. |
| 3 · Adiabatic assist | Dry-bulb > 25 °C and the water gate is open | Dry-bulb < 24 °C, or the gate closes | Sprays staged unit by unit on supply temperature. Chillers stay off. |
| 2 · Chiller | Dry-bulb > 32 °C, or > 25 °C with the gate closed | Dry-bulb < 31 °C (or < 24 °C, gate closed); 20 min minimum run | Chillers staged on load. Dry coolers reject condenser water at up to 55 °C. Sprays only if the gate is open. |
Setpoints, deadbands and timers are concept-stage and are tuned in the design study. is measured at the dry cooler inlets, shielded from their own exhaust.
Inputs · any one closes the gate
- Drought restrictions in forceUtility restriction level
- Break tank below reserveOn-site level transmitter
- Operator lockoutPlant control system
- Restriction signal lostWatchdog on the utility feed
Water gate
Open
Adiabatic assist is permitted. The plant may use water in modes 2 and 3 when the weather calls for it.
Fails closed
Lose the utility signal, the tank level or the controller, and the gate shuts. The safe state for the community is a dry plant.
Costs energy, never capacity
The chillers carry the full load without water, so closing the gate can raise PUE but can never let the supply temperature drift.
Separable and auditable
The gate is its own logic block with its own log of restriction state and spray hours, so compliance can be shown, not asserted.
Keeping the gate separate from the mode logic is the point of the design. Earlier work switched cooling modes on signals from inside the facility; here the signal comes from the supply the plant draws on, and nothing the plant needs can override it.
10 — Indicative performance
94% less water, for 1.2% more energy.
A first-pass hourly model: every hour of a synthetic Western Sydney year run through the plant’s control logic, once with the community supply normal all year and once with drought restrictions all year.
- Free cooling
- Adiabatic assist
- Chiller
Hours per year by modeof 8,760
- Free cooling
- 7,279 h83%
- Adiabatic assist
- 1,122 h13%
- Chiller
- 359 h4%
Indicative · concept-stage estimate. Hover or drag across the year to read an hour
Free cooling carries winter, spring, autumn and every night. The warm hours cluster on summer afternoons, exactly when a cooling tower would be drawing hardest. With water available, adiabatic assist absorbs most of them; under restrictions the chiller takes them all, and the plant uses no water.
Indicative · concept-stage estimate. Tower evaporation only; drift and blowdown would add to the baseline
Results
| Scenario | Water, ML/yr | Peak day, ML | WUE, L/kWh | PUE | Chiller h/yr |
|---|---|---|---|---|---|
| A · Cooling towers | 254 | 0.70 | 1.45 | 1.075 | — |
| E · Proposed, supply normal | 15 | 0.22 | 0.09 | 1.088 | 359 |
| E · Proposed, drought year (= D) | 0 | 0.00 | 0.00 | 1.114 | 1,481 |
Indicative · concept-stage estimate. and are annual; peak day is the highest single day of water use
Method and assumptions
Every figure on this page comes from one small model with the assumptions listed here, kept in the site’s source so they can be checked and changed. It is deliberately simple and conservative on the chiller. It is not a substitute for the hourly design study on measured weather, which is the next step.
- Weather
- Synthetic hourly year matched to Western Sydney monthly means; peak 46.1 °C
- IT load
- 20 MW, constant, all 8,760 hours
- Dry cooler approach
- 5 K: free cooling up to 25 °C dry-bulb
- Adiabatic band
- Pre-cooling keeps the chiller off up to 32 °C dry-bulb
- Adiabatic water
- 0.5 L per kWh of heat rejected while spraying
- Chiller COP
- 5 dry, 6 with pre-cooled air; carries the full load when running
- Fans and pumps
- 3% of IT (+0.5% while spraying)
- Other overhead
- 5% of IT: electrical losses, lighting, offices
- Tower baseline
- 1.45 L/kWh evaporated every hour; fans and pumps 2.5% of IT
11 — Compliance and risk
The gate is becoming a requirement, not a preference.
Commonwealth
Sustainable water use is, for the first time, an explicit expectation of new data centre and AI infrastructure developments.
Source · Department of Industry, Science and Resources (2026)
New South Wales
Facilities that use potable water must be designed to reduce water use under drought restrictions. The gate does exactly that, and its log shows it.
Source · Infrastructure NSW (2026), NSW data centre guidelines
Risk register · concept
| Risk | Consequence | Mitigation |
|---|---|---|
| Heatwave beyond the 47 °C design day, with water restricted | Chillers at their highest lift; dry cooler capacity derates. | N+1 chillers and dry coolers; design day close to the Penrith record; agreed emergency provisions with the utility, logged. |
| Next-generation hardware needs supply below 30 °C | Fewer free-cooling hours, more chiller energy. | Chillers sized for the full IT load; setpoint reset tested in the design study. |
| Humid heat weakens adiabatic pre-cooling | Chiller runs more often than the dry-bulb model suggests. | Design study models wet-bulb depression hour by hour; the chiller covers any shortfall. |
| Controls fault or lost utility signal | Gate state uncertain. | Gate fails closed; watchdog on the feed; manual overrides logged. |
| Water quality in the pre-coolers | Microbial growth in wetted media. | Units drained when idle; water treatment and a water risk management plan. |
| Refrigerant leakage | Direct emissions. | Low-GWP refrigerant, leak detection and a refrigerant management plan. |
12 — Validation and next steps
What the design study has to answer.
The design question
For a 30 °C setpoint in a warm, humid climate, how many hours a year can a dry cooler hold that setpoint unaided, and what is the energy penalty when a high-temperature chiller covers the rest?
01
Hourly design study
Replace the synthetic year with measured Western Sydney weather, dry-bulb and wet-bulb, and today’s and projected climates. Compare against the chiller-less and cooling-tower baselines on energy, water and water-scarcity cost.
02
Equipment selection
Select dry coolers and chillers at the design day; replace the assumed COPs and fan power with manufacturers’ part-load data.
03
Controls and the utility
Detail the sequence of operations, and agree the restriction signal, its source and its failure behaviour with the water utility.
04
Basis of design
Issue the basis of design report, plant schematics, sequence of operations and performance report for schematic design.
Reject heat to the air.
Use water only when the community can spare it.