Compute Waste Heat as a Low-Grade Thermal Offset for Lunar ISRU

1. The idea, and its hard limit

Space compute (lunar-surface data centers, autonomy/processing hardware) generates large quantities of waste heat that must be rejected. Lunar ISRUIn-Situ Resource UtilizationProducing usable materials (oxygen, water, propellant) from local resources instead of shipping them from Earth. is thermally limited. The tempting pitch is "use the compute waste heat to power ISRUIn-Situ Resource UtilizationProducing usable materials (oxygen, water, propellant) from local resources instead of shipping them from Earth.." Stated that broadly, it is wrong, and a reviewer will reject it on the Second LawSecond Law of thermodynamicsHeat flows only to lower temperature, so ~330 K compute waste heat can serve sublimation (~273 K) but never high-grade reduction heat (1073-1900 K).: heat flows only to lower temperature. Compute waste heat rejects at roughly 315-350 K (GPUGraphics Processing Unit (GPU)A massively parallel AI accelerator; here the funded product whose waste heat anchors the ISRU cascade and whose revenue pays for the heat source. junctions ~360- 375 K; coolant/radiator interface lower). The reduction and electrolysisElectrolysisUsing electric current to split a compound (here, water into H2 and O2, or oxide melts into metal and O2); an irreducibly electrical load. routes need high-gradeHigh-grade heatHeat at high temperature (1073-1900 K) needed to drive reduction and electrolysis routes. heat, hydrogen reductionHydrogen reductionReact ilmenite with H2 at ~1073-1373 K to make water, which is then electrolyzed; the one route with a clean published energy figure. 1073-1373 K, carbothermalCarbothermal reductionHigh-temperature (~1900 K) reduction of regolith using carbon or methane to liberate oxygen. ~1900 K, molten regolithRegolithThe loose, broken-rock surface layer of the Moon; the feedstock for every non-water route. electrolysisElectrolysisUsing electric current to split a compound (here, water into H2 and O2, or oxide melts into metal and O2); an irreducibly electrical load. ~1900 K, molten-salt ~1200 K. You cannot drive any of them with 350 K heat. Compute waste heat therefore cannot replace the dominant ISRUIn-Situ Resource UtilizationProducing usable materials (oxygen, water, propellant) from local resources instead of shipping them from Earth. heat demand.

There is exactly one exception, and it is the strategically important one.

2. Where grade matches: PSR water-ice sublimation

Thermal miningPolar water-ice miningSublimate water ice out of permanently shadowed regolith at ~273 K and cold-trap the vapor; the only low-temperature route. of polar water ice raises icy regolithRegolithThe loose, broken-rock surface layer of the Moon; the feedstock for every non-water route. from the PSRPermanently Shadowed RegionA crater floor near the poles that never sees sunlight, sitting at ~40-110 K: both the water resource and an ideal cold sink. floor (~40-110 K) to the sublimationSublimationA solid passing directly to vapor without melting; water ice sublimes at ~273 K under vacuum, the basis of the low-temperature water route. point (~273 K) and supplies the heat of sublimationSublimationA solid passing directly to vapor without melting; water ice sublimes at ~273 K under vacuum, the basis of the low-temperature water route. at ~273 K. A heat source at 315-350 K is comfortably above that, so it is thermodynamically valid to drive the entire water-extraction thermal chain with compute waste heat. The same is true, marginally, of the cold slice of feedstock preheating for any route (from feed temperature up to ~350 K), but that is a tiny fraction of the heat a reduction route needs.

3. Method

We extend the lpem common-basis energy model (which already resolves each route's heating demand against temperature) with a grade filter: given a reject temperatureReject temperatureThe temperature at which compute waste heat is shed (~315-350 K); it sets how much low-grade ISRU heat that stream can offset. T_reject, the usable source temperature is first derated by a heat-exchanger approach delta-T (nominal 15 K, range 5-30) and delivered heat is scaled by an exchanger effectiveness (nominal 0.85, range 0.70-0.95). The offsettable energy is then the electrical-equivalentElectrical-equivalent energyThermal demand converted to the electricity it would take to supply it (via an explicit heating-efficiency knob), so all routes compare on one basis. heating that derated source could displace: the sensible slice from feed temperature up to min(T_react, T_src) for regolithRegolithThe loose, broken-rock surface layer of the Moon; the feedstock for every non-water route. routes, and the full thermal-miningPolar water-ice miningSublimate water ice out of permanently shadowed regolith at ~273 K and cold-trap the vapor; the only low-temperature route. term for the water route when T_src ≥ 273 K. High-gradeHigh-grade heatHeat at high temperature (1073-1900 K) needed to drive reduction and electrolysis routes. reaction heat, fusion, and all electrical loads (electrolysisElectrolysisUsing electric current to split a compound (here, water into H2 and O2, or oxide melts into metal and O2); an irreducibly electrical load. work, liquefactionLiquefactionCooling a gas into a cryogenic liquid (e.g. O2 to LOX); a leading cost of the water route is LH2 liquefaction., compression) are excluded by construction. See src/lpem/waste_heat.py; reproduce with python -m lpem --waste-heat.

4. Results

Offsettable low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. thermal demand at T_reject = 350 K (per kg O2Molecular oxygenThe primary product; energy is reported per kg of O2 delivered to cryogenic storage., electrical- equivalent displaced after pinch and effectiveness; route totals from the current lpem model, v0.13):

The water route is the clear beneficiary: its entire thermal-miningPolar water-ice miningSublimate water ice out of permanently shadowed regolith at ~273 K and cold-trap the vapor; the only low-temperature route. demand is low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. and so fully offsettable (scaled only by the exchanger effectiveness), ~11% of its total energy. The high-temperature routes gain almost nothing (the share is even smaller than in earlier model versions, because their totals rose once continuous reactor heat loss was charged; the low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. preheat slice is a sliver of a larger denominator, and the energy that matters, the 1073-1900 K reactor heat, is untouchable by a 350 K source).

Magnitude. The constraint is not heat quantity; it is that the low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. demand is small. A 50 t O2Molecular oxygenThe primary product; energy is reported per kg of O2 delivered to cryogenic storage./yr water plant needs only ~10 kW of continuous low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. heat, so a ~12 kW compute load covers it entirely, displacing that fission heating and ~2.3 t of landed reactor mass (at ~225 kg/kWeKilowatts, electricalElectrical power output (distinct from thermal power or installed nameplate).). Conversely, on energy grounds alone a 100 kW compute facility's rejected heat could serve the low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. demand of roughly 500 t O2Molecular oxygenThe primary product; energy is reported per kg of O2 delivered to cryogenic storage./yr of water miningPolar water-ice miningSublimate water ice out of permanently shadowed regolith at ~273 K and cold-trap the vapor; the only low-temperature route., far more than any near-term plant. Two caveats keep this honest. First, these supportable-throughput figures are energy-balance upper bounds, not delivery designs: moving the heat into a granular icy bed (effective conductivity ~0.001-0.01 W/m/K in vacuum, ~50-77 K driving delta-T) is conduction-rate-limited, and an illustrative contact-area estimate for conducting even 12 kW into the bed is on the order of 5,000 m^2 (k_eff = 0.003 W/m/K, 0.1 m path). Delivery engineering, not energy supply, is the binding constraint. Second, within the energy boundary, co-located compute waste heat does not merely help; it saturates the low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. ISRUIn-Situ Resource UtilizationProducing usable materials (oxygen, water, propellant) from local resources instead of shipping them from Earth. heat demand.

5. What this is, and is not

• It is: a free (otherwise-dumped) heat stream that covers the low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. thermal demand of the PSRPermanently Shadowed RegionA crater floor near the poles that never sees sunlight, sitting at ~40-110 K: both the water resource and an ideal cold sink. water route, eliminating ~11% of its production energy and the corresponding reactor mass, and giving a co-located compute facility a productive heat sink instead of a pure radiator load.
• It is not: a replacement for ISRUIn-Situ Resource UtilizationProducing usable materials (oxygen, water, propellant) from local resources instead of shipping them from Earth.'s high-gradeHigh-grade heatHeat at high temperature (1073-1900 K) needed to drive reduction and electrolysis routes. reduction/electrolysisElectrolysisUsing electric current to split a compound (here, water into H2 and O2, or oxide melts into metal and O2); an irreducibly electrical load. heat (Second Law), and it is not a large energy win in absolute terms (~1-3 kWhKilowatt-hourA unit of energy (one kilowatt sustained for one hour). The paper reports kWh per kg of oxygen./kg O2Molecular oxygenThe primary product; energy is reported per kg of O2 delivered to cryogenic storage.). The value is that it is free, and that it lands specifically on the water route, the only route that yields complete propellant (LOXLiquid OxygenCryogenic liquid O2, stored at ~80-90 K; the oxidizer half of the propellant. + LH2Liquid HydrogenCryogenic liquid H2 (boils ~20 K); the fuel half of a complete LOX+LH2 propellant.).

6. Limitations

• Co-location is assumed, not shown. Orbital data-center waste heat cannot reach the surface; this requires a surface (ideally PSR-adjacent) compute facility, and heat transport (heat pipes / pumped loops) from racks to the regolithRegolithThe loose, broken-rock surface layer of the Moon; the feedstock for every non-water route., with losses not modeled here. The co-location architecture is treated in the separate PSR-co-location study.
• Electrical loads dominate the water route (electrolysisElectrolysisUsing electric current to split a compound (here, water into H2 and O2, or oxide melts into metal and O2); an irreducibly electrical load., LH2Liquid HydrogenCryogenic liquid H2 (boils ~20 K); the fuel half of a complete LOX+LH2 propellant. liquefactionLiquefactionCooling a gas into a cryogenic liquid (e.g. O2 to LOX); a leading cost of the water route is LH2 liquefaction.); waste heat does nothing for those, so the offset ceiling is ~11% regardless of how much compute heat is available.
• Delivery is conduction-limited. The model bounds the energy a waste stream can displace; it does not design the regolith-side heat exchanger. Conducting heat into a granular icy bed at cryogenic temperature is severely rate-limited (see Section 4), so the supportable-throughput numbers are upper bounds.
• Solar-thermal concentrators are the obvious competing low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. (and high-gradeHigh-grade heatHeat at high temperature (1073-1900 K) needed to drive reduction and electrolysis routes.) heat source; the case for compute waste heat rests on the compute existing anyway for its own reasons, not on heat being scarce.
• Reject temperatureReject temperatureThe temperature at which compute waste heat is shed (~315-350 K); it sets how much low-grade ISRU heat that stream can offset.. We take the compute coolant-loop reject as ~330 K nominal (a 315-360 K band). The GPUGraphics Processing Unit (GPU)A massively parallel AI accelerator; here the funded product whose waste heat anchors the ISRU cascade and whose revenue pays for the heat source. junction is hotter (~360-375 K, i.e. ~90 C, the figure the program literature quotes), but the loop that actually reaches the regolithRegolithThe loose, broken-rock surface layer of the Moon; the feedstock for every non-water route. is cooler; this paper's reject temperatureReject temperatureThe temperature at which compute waste heat is shed (~315-350 K); it sets how much low-grade ISRU heat that stream can offset. is the loop/radiator-interface value, not the junction. The water-route offset is insensitive to the exact value within this band: any post-pinch source temperature above the ~273 K sublimationSublimationA solid passing directly to vapor without melting; water ice sublimes at ~273 K under vacuum, the basis of the low-temperature water route. target serves the entire low-gradeLow-grade heatHeat available only at modest temperature; useful for ice sublimation but useless for high-temperature reduction. thermal-miningPolar water-ice miningSublimate water ice out of permanently shadowed regolith at ~273 K and cold-trap the vapor; the only low-temperature route. slice, so the headline offset is the same at 330 K as at the 350 K used for the table (at 330 K the derated source is ~315 K, still comfortably above 273 K). The pinch and exchanger effectiveness are now modeled explicitly (approach delta-T nominal 15 K, range 5-30; effectiveness nominal 0.85, range 0.70-0.95); the sublimationSublimationA solid passing directly to vapor without melting; water ice sublimes at ~273 K under vacuum, the basis of the low-temperature water route. credit vanishes if the post-pinch source falls below 273 K.

7. Reproduce

pip install -e .
python -m lpem --waste-heat           # offset table + a 50 t/yr water heat balance
pytest tests/test_waste_heat.py

Sources

• lpem energy model (this repo), route thermal demand split by temperature.
• Leger et al. 2025 (PNASProceedings of the National Academy of SciencesThe journal where Leger et al. 2025, the paper’s main validation anchor, appeared.); NASANational Aeronautics and Space Administration CFM/CryoFILL, water-route thermal miningPolar water-ice miningSublimate water ice out of permanently shadowed regolith at ~273 K and cold-trap the vapor; the only low-temperature route. and liquefactionLiquefactionCooling a gas into a cryogenic liquid (e.g. O2 to LOX); a leading cost of the water route is LH2 liquefaction..
• Colaprete et al. 2010 (LCROSS), PSRPermanently Shadowed RegionA crater floor near the poles that never sees sunlight, sitting at ~40-110 K: both the water resource and an ideal cold sink. ice grade and sublimationSublimationA solid passing directly to vapor without melting; water ice sublimes at ~273 K under vacuum, the basis of the low-temperature water route. context.