The Desert Heat Engine

How overnight temperatures turn AI waste heat into recovered power — and why the cold sink is as valuable as the power supply in any honest accounting of data centre site economics.

Memo9 — Energy Series
AuthorBrett Murrell
Versionv1.1
Date14 May 2026
SeriesMMA Memos
Word count~5,800
An AI data centre does not consume electricity. It converts it — almost entirely into heat. A 1 GW campus delivers roughly 900 MW of waste heat to its cooling system continuously, 24 hours a day, every day of the year. In every current data centre deployment, that heat is a liability: something to be expelled, cooled away, and forgotten. This memo argues it is an asset — and that the size of that asset depends almost entirely on the overnight temperature at the site. Using the Organic Rankine Cycle (ORC), waste heat at 60°C can be converted back into electricity, with efficiency determined by the temperature differential between the hot source and the cold sink. At the MMC inland desert corridor sites, monthly average nighttime lows range from 3.9°C in July to 22.3°C in January. That range produces a 2× swing in recovered power — from 51 MW recovered per GW of compute in summer to 76 MW in winter — before any solar boosting of the hot side. With solar-thermal boosting to 90°C, winter nights recover 107 MW per GW of campus. These numbers are not yet in any data centre site selection model anywhere in the world. They should be. This memo does the calculation, explains the physics, and draws the site selection conclusion.
3.9°CJuly average overnight low at MMC inland desert sites — the coldest cold sink in the thermodynamic cycle
76 MWRecovered per GW of compute on winter nights at 60°C hot side — no solar boost
107 MWRecovered per GW with solar-thermal boost to 90°C hot side — winter nights
Swing in recovered energy between summer and winter nights — priced into no current site model

1. The Problem Nobody Has Solved

Every watt of electricity that enters a data centre eventually leaves as heat. The servers convert electrical power to computation, and computation generates heat. The cooling systems move that heat from the servers to the environment — via air handlers, chilled water loops, cooling towers, or direct liquid cooling. At the end of every pathway, the heat is rejected to the ambient environment and lost.

In a 1 GW AI campus, the continuous heat rejection rate is approximately 900 MW. That is not a rounding error. It is a thermal output comparable to a large industrial furnace, operating around the clock. In urban and coastal environments, it is a planning liability — contributing to urban heat islands, consuming scarce water in cooling towers, and generating community opposition. Nobody has treated it as a resource.

The reason is straightforward: heat is only useful if you can do something with the temperature differential between the hot source and a colder sink. In a data centre built next to a city, the hot source (the cooling loop water) is at 35–45°C and the ambient air is at 15–25°C. The differential is modest. The thermodynamic work you can extract is small. It has not been worth the engineering effort.

In a data centre built in an Australian inland desert, with a cooling loop deliberately designed to output water at 60°C, and winter nights that drop to 3.9°C, the differential is 56°C. That is a meaningfully different thermodynamic situation — and it produces meaningfully different results.

2. What is the Organic Rankine Cycle

The Organic Rankine Cycle is a heat engine — a machine that converts a temperature differential into mechanical work and then into electricity. It works on the same principle as a steam turbine power station, but uses an organic working fluid with a low boiling point instead of water. This allows it to operate efficiently at the low temperatures (60–120°C) typical of industrial waste heat streams, where water-based steam cycles are impractical.

The ORC cycle has four stages:

  1. The evaporator. Waste heat from the data centre cooling loop is passed through a heat exchanger. The organic working fluid absorbs this heat and vaporises.
  2. The expander. The high-pressure vapour expands through a turbine or scroll expander, doing mechanical work and generating electricity.
  3. The condenser. The expanded vapour is cooled back to liquid by the cold sink — ambient night air, or a water stream. This is the critical stage. The colder the condenser, the greater the pressure drop across the expander, and the more work extracted.
  4. The pump. The liquid working fluid is pumped back to high pressure and the cycle repeats.

The theoretical upper limit on ORC efficiency is the Carnot efficiency — a function purely of the hot and cold temperatures:

Carnot efficiency formula
η = 1 − (Tcold / Thot)where temperatures are in Kelvin (°C + 273.15)

Real ORC systems achieve approximately 40–60% of the Carnot theoretical maximum, depending on working fluid selection, heat exchanger design, and operating conditions. For the calculations in this memo, 50% of Carnot is used as the realistic ORC efficiency — consistent with the peer-reviewed literature on low-grade waste heat ORC systems.

The implication is direct: every degree of reduction in the cold sink temperature increases ORC efficiency and increases recovered power. The night temperature is not background context — it is a primary engineering variable.

3. The Cold Sink Is the Asset

Current data centre site selection evaluates power cost, power reliability, land cost, connectivity, and jurisdiction. It does not evaluate the cold sink. This is a significant omission.

The cold sink — the ambient temperature available to reject heat from the ORC condenser — determines how much of the waste heat stream can be converted back to useful electricity. A site with cheap power but warm nights recovers less energy from its waste heat than a site with slightly more expensive power but cold nights. If the recovered energy value is large enough, it changes the site economics entirely.

The global locations already recognised for cold-sink advantages are Iceland and Scandinavia — where data centres have been built specifically to exploit year-round cold ambient temperatures for free cooling. What has not been recognised is that the Australian inland desert offers cold-sink conditions competitive with Scandinavia for six to eight months of the year — combined with a solar resource that Scandinavia cannot match, and an aqueduct water supply that Iceland cannot offer.

The night temperature is not a comfort metric for data centre operators. It is a thermodynamic asset worth quantifying in dollars per megawatt-hour of recovered electricity. Nobody is doing this calculation. The MMC inland corridor sites perform well when it is done.

4. The Numbers — Monthly ORC Recovery at MMC Desert Sites

The following analysis uses the monthly average nighttime low temperatures at the MMC inland Pilbara corridor zone as the cold sink, a data centre cooling loop output of 60°C as the hot source, and 50% of Carnot efficiency as the realistic ORC conversion rate. Waste heat is assumed at 900 MW per 1 GW of installed compute — 90% of electrical input becoming recoverable thermal output.

Month Night low (°C) ΔT (°C) Carnot efficiency Real ORC efficiency Recovered per 1 GW campus
January22.337.711.3%5.7%51 MW
February21.138.911.7%5.8%53 MW
March17.942.112.6%6.3%57 MW
April12.947.114.1%7.1%64 MW
May8.052.015.6%7.8%70 MW
June4.755.316.6%8.3%75 MW
July3.956.116.8%8.4%76 MW
August5.854.216.3%8.1%73 MW
September11.248.814.6%7.3%66 MW
October14.945.113.5%6.8%61 MW
November18.341.712.5%6.3%56 MW
December20.739.311.8%5.9%53 MW
Annual average13.546.514.0%7.0%63 MW

The winter months of June, July, and August — when actual overnight lows frequently dip below 0°C with ground frost — represent the peak ORC recovery window. At 3.9°C average, July delivers 76 MW recovered per GW of compute. January at 22.3°C delivers only 51 MW. That 25 MW difference per gigawatt of campus is the value of the cold sink — unpriced in any current data centre economics model.

On an annual average basis, 63 MW of electricity is recovered per GW of installed compute. At a power cost of 5¢/kWh, that is approximately $27.6M per year in recovered electricity value from a single 1 GW campus — before the solar-thermal boost described in the next section.

5. Solar-Thermal Boosting — Raising the Hot Side

The ORC hot side in the base case is limited by the data centre cooling loop temperature: 60°C. This is achievable with liquid immersion cooling — a deliberate design choice discussed in section 6 — but it is not the ceiling.

The MMC corridor AI campus sites receive some of the highest solar irradiance on Earth. Rooftop and perimeter solar thermal collectors — flat plate or evacuated tube — can raise a heat transfer fluid to 80–120°C during daylight hours. That heated fluid can be blended into the ORC hot-side feed, raising the effective source temperature from 60°C to 90°C or higher. This is the solar-boosted ORC configuration demonstrated by Rice University in 2025, which showed 60–80% more electricity recovery from the same waste heat stream.

The desert context makes this more effective than any previously tested environment:

Cold sink conditionHot side (°C)ΔT (°C)Real ORC efficiencyRecovered per 1 GW
July night avg (3.9°C) — no boost6056.18.4%76 MW
July night avg (3.9°C) — solar boost9086.111.9%107 MW
Annual avg night (13.5°C) — solar boost9076.510.5%94 MW
January night avg (22.3°C) — solar boost9067.79.4%84 MW

Solar-boosted ORC on a winter night recovers 107 MW per GW of compute — more than 10% of the campus electrical input returned as recovered power. Across a 5 GW campus, that is 535 MW of recovered electricity during the winter operating window — enough to power a city of 200,000 people.

Even on the worst summer nights, solar-boosted ORC recovers 84 MW per GW. The floor is high. The ceiling — on cold winter nights with solar thermal boost — is genuinely significant.

6. The Immersion Cooling Requirement

The 60°C hot-side temperature is not a given. It is a design choice — and it requires liquid immersion cooling rather than conventional air cooling.

In a conventional air-cooled data centre, servers sit in racks in cooled rooms. The cooling loop operates at 15–25°C supply, rejecting heat at 30–40°C. The temperature is too low for effective ORC. The system is designed for server comfort, not thermodynamic efficiency.

In a liquid immersion cooling system, servers are submerged in a dielectric fluid — a non-conductive liquid that draws heat directly from the chips. The fluid operates at higher temperatures because the chips can tolerate higher fluid temperatures when liquid-cooled. Modern immersion cooling systems operate with fluid outlet temperatures of 50–70°C — and next-generation two-phase immersion systems push higher still.

The largest AI campuses being designed today are specifying immersion cooling precisely because it reduces cooling energy consumption by 30–50% compared to air cooling, and enables higher server density. The ORC heat recovery benefit is additive — a further reason to choose immersion cooling that has not yet been costed into site selection.

The design requirement for the MMC inland AI campus is therefore: immersion cooling as the standard, with cooling loop output temperature maintained at 60°C or above, feeding directly into the ORC heat exchangers. This is not an experimental configuration. It is an engineering decision that must be locked in at campus design stage — which is why it belongs in the planning framework for Phase 1–3 corridor new towns, before the campuses are built.

7. The Warmed Water Loop

After the ORC condenser rejects heat into the cooling water stream, that water is warmer than when it arrived. Depending on flow rate and the heat load, condenser outlet water can reach 35–50°C. This is not waste — it is a warm water supply at no additional energy cost.

The productive uses of warm water at corridor new town scale are a subject for further study. The possibilities include controlled-environment agriculture, aquaculture in indoor facilities, and biological cultivation systems that thrive in warm water conditions. None of these requires the full quantification here. What matters for the planning framework is that the warm water output of the ORC condenser is a design feature to be captured, not a discharge to be managed. The campus plan should route condenser outlet water to productive use as a first principle, with the specific applications determined by the economic conditions at each corridor town site.

8. ORC vs Stirling — Choosing the Right Engine

ORC is the right technology for the base waste heat recovery case — steady 40–150°C source, commercial and modular, 3–8% thermal efficiency. But the desert adds a second option at higher temperatures: the Stirling engine.

A Stirling engine is an external combustion engine that runs on any temperature differential. It does not require combustion — it runs on heat from any source. At source temperatures above 200°C, Stirling engines achieve 10–25% thermal efficiency, significantly outperforming ORC. The challenge is reaching those temperatures. Concentrating solar thermal (CST) — parabolic trough or dish collectors — can raise a heat transfer fluid to 250–400°C during desert daylight hours, well above the Stirling threshold.

ParameterORCStirling (CST-boosted)
Optimal source temperature40–150°C200–400°C+
Thermal efficiency3–8% (low grade); up to 15% solar-boosted10–25% at high temperature
Heat sourceData centre cooling loop directlyCST collectors + thermal storage
Nighttime operationYes — runs on stored hot-side fluidYes — thermal storage essential
Cold sink benefitHigh — every °C drop improves outputHigh — same thermodynamic principle
MaturityCommercial — plug-and-play modules availableCommercial but fewer large-scale deployments
Best application at MMC campusPrimary waste heat recovery — always onPeak despatchable generation — daytime CST stored for night

The optimal MMC desert campus runs both in a cascade. ORC handles the continuous waste heat stream from immersion cooling loops — 24/7, baseload recovery. CST-boosted Stirling handles peak despatchable generation, charged during the day and discharged through the night. Thermal storage in molten salt or pressurised water tanks bridges the gap, enabling true 24/7 operation with reduced battery reliance. Studies show thermal storage combined with solar boosting recovers 60–81% more annual electricity than ORC alone on waste heat.

9. The Low Humidity Advantage

Desert air is dry. This is not incidental — it is a direct engineering advantage for the condenser stage of both ORC and Stirling cycles, and for data centre cooling generally.

In humid environments, air-cooled condensers struggle. Moisture in the air reduces the temperature differential between the condenser surface and the ambient air, degrades heat transfer efficiency, increases corrosion rates, and requires higher-maintenance equipment to manage biological fouling and mineral scaling. Cooling tower blowdown — the water wasted to control mineral concentration — is a significant water cost in humid climates.

In the Australian inland desert, humidity is consistently low — often below 20% relative humidity at night. This means:

The humidity advantage does not appear in any current data centre energy model. Combined with the cold-sink temperature advantage, it makes the desert condenser stage materially more efficient than any temperate or coastal equivalent — at no additional cost.

10. Site Scoring — The Desert Wins on Every Axis That Matters

The integrated analysis — ORC/Stirling recovery, cold sink, humidity, solar resource, land, and water — produces a clear site scoring result. The following ratings assess five location types specifically for the hybrid ORC/Stirling + solar thermal storage model at AI campus scale, across ten criteria: diurnal swing, humidity, solar resource, land availability, water security, power access, regulatory environment, climate resilience, connectivity, and biocircular upside.

RankLocation typeScoreKey strengthsKey weaknesses
1 Remote arid desert — Australian NT/WA/SA outback, MMC corridor zones 92/100 Extreme diurnal swing, low humidity, unlimited cheap land, world-class solar, water-positive via AWG/condensate, MMC power and water spine Longer fibre runs; remote labour logistics (solved by MMC new towns)
2 Inland temperate plains — inland NSW/VIC 74/100 Moderate temperature swings, good land, decent solar, easier grid tie-in Smaller diurnal ΔT, higher humidity nights, more water stress
3 Coastal temperate — near Perth, Adelaide, SE Queensland 67/100 Excellent fibre and latency, strong grid infrastructure, some renewable access Higher humidity, smaller diurnal swing, land constraints and cost
4 Urban/metropolitan fringe — Sydney, Melbourne, Brisbane 52/100 Best latency and talent pool, fast permitting in some zones High land cost, community opposition, no pond space, poor cooling efficiency
5 Tropical/humid — Far North QLD, Southeast Asia 41/100 Abundant water, some renewable options High humidity kills ORC efficiency, minimal diurnal swing, cyclone/flood risk

The 51-point gap between remote arid desert (92) and tropical/humid locations (41) is the thermodynamic argument made visible. Singapore and Jakarta — where most current hyperscale AI deployment in the Indo-Pacific is concentrated — score in the tropical/humid band. They are built there because the demand is there, not because the physics favours it. As AI inference becomes a commodity delivered via subsea cable from large inland campuses, the campus can be where the physics is best. The physics is best in the Australian desert.

11. The Broader Site Selection Implication

The desert ORC analysis generalises beyond the MMC corridor. Its conclusion applies to any data centre site selection globally: the cold sink temperature is an unpriced economic variable, and cold environments are systematically undervalued as AI infrastructure real estate.

The locations that score highest on this variable are, in rough order:

The insight is that tropical coastal deployments — Singapore, Jakarta, Mumbai — are thermodynamically the worst locations for waste heat recovery. Every watt of waste heat is rejected into already-warm air at minimal thermodynamic benefit. They are built there because the demand is there, not because the physics favours it.

As AI inference becomes a commodity exported from large campuses to end users via subsea cable — as described in Memo 8 — the geographic constraint of proximity to demand dissolves. The campus can be where the physics is best. And where the physics is best, combined with cheap power, water, and space, is the Australian inland desert on the MMC corridor.

A data centre in Singapore operates at a permanent thermodynamic disadvantage. Its waste heat is nearly worthless because the ambient temperature is too high to create a useful differential. A data centre on the MMC desert corridor in July, with immersion cooling at 60°C and night air at 3.9°C, recovers 76 MW per GW of compute — and 107 MW with solar boosting. The difference is not engineering. It is geography. And geography is fixed.

12. First-Mover Position

No large-scale AI campus has been designed with ORC waste heat recovery as a primary system component. The concept is well-understood in the academic literature — Rice University’s 2025 demonstration of solar-boosted ORC showed 60–80% more electricity recovery — but it has not been implemented at hyperscale. The integration of immersion cooling, solar-thermal boosting, ORC recovery, and desert cold-sink exploitation as a unified campus design is genuinely novel.

The MMC programme is the only entity currently positioned to develop this at meaningful scale, for three reasons:

The ORC heat recovery memo series continues. The next subject is the productive use of condenser water — a fuller treatment of what warm water at corridor scale can produce in the context of new town food, fuel, and agricultural systems.

13. What Needs to Happen

Lock in immersion cooling as the standard. The campus design specification for Phase 1–3 MMC AI precincts must mandate liquid immersion cooling with a minimum cooling loop output temperature of 60°C. This decision cannot be retrofitted. It must be in the planning framework before campuses are tendered.

Commission an ORC pilot at Phase 0. Before the Phase 1 campuses are built, a pilot ORC installation at a Phase 0 data facility — even at modest scale, 10–50 MW thermal input — validates the real-world performance of the desert ORC system against the calculated values in this memo. The pilot generates the data for the full Phase 1 business case and establishes Australia as the first nation to demonstrate integrated AI campus heat recovery at scale.

Price the cold sink into site selection. The national AI infrastructure planning framework should include a standardised methodology for valuing ORC recovery potential at candidate sites, using local overnight temperature profiles, proposed cooling loop temperatures, and solar thermal availability. Sites that score well on cold-sink performance should receive preferential treatment in the campus zoning and power allocation framework.

Engage CSIRO. The Commonwealth Scientific and Industrial Research Organisation has directly relevant expertise in ORC systems, desert thermodynamics, solar thermal, and biological systems that can use warm water productively. A formal research partnership with CSIRO on the integrated MMC desert campus energy system is the natural next step — and positions the work for international publication and IP protection.