Lunar Surface: Where Waste Heat Becomes Feedstock

Every lunar-base architecture hits the same fundamental constraint: ISRU needs a heat source whose cost doesn't break the program.

Everywhere else, waste heat is a liability you pay to throw away. On the lunar polar surface the logic inverts. Every kilowatt of GPU waste heat at ~90 °C, instead of being rejected to space at near-zero benefit, can drive water-ice extraction from polar regolith. Heat that is pure cost elsewhere becomes the first useful step of a resource cascade.

This page documents Centradiant's lunar program: a data center as the head of an ISRU cascade. It is an independent effort, developed and funded separately from the company's orbital work, and is presented here on its own terms.

---

Why Lunar Now?

In February 2026, SpaceX publicly reordered its roadmap: Moon before Mars, a self-growing city in ~10 years, "data centers manufactured on the Moon," an AI satellite factory, first Starship V3 cargo to the lunar surface targeted for 2028. Every credible lunar-base program in the literature, NASA Artemis FSH, ESA Argonaut, CNSA/Roscosmos ILRS, Blue Moon, treats compute as a future luxury and ISRU as an agency-funded science exercise. The Feb 2026 SpaceX plan calls for data centers on the Moon but does not connect them to ISRU heat. This is the missing connection. Centradiant can fill it in either of two postures: operate its own compute at the ridge, or supply the heat-cascade integration to a third party's lunar data center. The second posture makes the ISRU layer a value-add on compute that someone else funds, at the cost of depending on that program's schedule. Which posture is primary is an open program decision, not a settled claim.

---

The Heat Problem on the Surface

The lunar polar surface has both halves of the thermal cascade available in close proximity:

• Peak-of-Near-Eternal-Light ridges at ~85–93% annual illumination (Connecting Ridge near Shackleton crater: 89.44°S, 222.69°E) support continuous solar PV.
• Permanently Shadowed Regions at 40–100 K, immediately adjacent, function as a passive cryogenic sink, for radiator overflow, for water-vapor cold-trap condensation, and for passive LOX storage at ~80 K with near-zero boil-off.

The cascade closes one direction:

1. Solar PV → GPU compute at ~90 °C waste heat
2. Waste heat → thermal bus (pumped fluid loop, Lunar Ice Miner heritage, ICES-2022-130)
3. Thermal bus → ORC turbine (~13% efficiency at a 90 °C source against a PSR-facing radiator sink) → offsets array draw
4. Thermal bus → ISRU water plant (~100 kW thermal → ~1.2 t/day H₂O via auger-dryer + cold-trap, at an assumed ~5 wt% accessible ice; the yield scales with the local grade, which is not yet ground-truthed)
5. Thermal bus → regolith TES + survival heating
6. Remainder → radiator → PSR cold sky

Compute is the funded product. ISRU process heat is the near-zero-marginal-cost byproduct.

---

Why Previous Lunar Architectures Did Not Scale

Existing waste-heat cascades for lunar ISRU all anchor on heat sources whose only product is heat:

• RTG (Lunar Ice Miner, ICES-2022-130; ACT + Honeybee Robotics), radioisotope cost and supply limit scale.
• Fission Surface Power (NASA FSP, 40 kWe target 2030), not flight-ready before ~2030, nuclear launch approvals add a separate regulatory track.
• Solar concentrator (Sowers & Dreyer, New Space 2019), large optical aperture; mechanically complex.
• Generic crew-base equipment, agency-funded, no revenue mechanism.

A heat source funded by independent compute revenue does not appear in the published literature. No credible lunar-base program carries a self-funded surface revenue mechanism. That is the gap this branch addresses.

---

The Inversion

Put a revenue-generating GPU compute facility, 500 kW installed, ~360 commercial-GPU-class accelerators, NVL72-class racks, at the head of the cascade. Compute is the primary, independently financed product. ISRU process heat falls out as a byproduct of heat that would otherwise cost a radiator to discard.

Two consequences follow:

1. Compute revenue funds the mission, rather than waiting for an agency appropriation. Mission-1 standalone does not pay back at the seed scale (~$2–4 B CapEx vs ~$5–10 M/yr Year-1 revenue), but break-even shifts into reach at multi-mission scale: base case ~18–20 missions and ~25 years to cumulative positive ROI; optimistic ~8–10 missions / ~10 years.
2. Waste heat becomes the means of transforming lunar matter. The near-term, self-funded product is water and oxygen: the ~90 °C heat drives thermal extraction directly. Turning that water into chemical propellant adds a separate electrical load (electrolysis runs ~5–6 kWh per kg of water, of order ~250–300 kWe for 1.2 t/day) that competes with the compute load, so propellant follows as installed power grows rather than coming free with the heat. The ISRU dividend still bends the cost-per-MW curve down with each subsequent mission.

Beyond the operating business there is a much larger notional resource figure, useful for scale but not a market and not a booked asset: at a $3,000/kg Starship-delivered displacement value, one km³ of polar regolith holds on the order of $45 T in oxygen + metals and the adjacent Shackleton PSR water-ice inventory on the order of $1,800 T. After a 99% execution-risk haircut, a reachable monetizable subset over 25 years is ~$200–300 B, roughly 15× cumulative program CapEx. The legal basis is narrower than a property claim: the Artemis Accords affirm that resource extraction and use are permissible and provide for deconfliction "safety zones," but they grant no mineral title, and ownership of in-place resources remains unsettled under Outer Space Treaty Article II (non-appropriation). First operational presence confers an operational and safety-zone advantage, not enforceable mineral rights.

---

Built on Proven Heritage

Every rung of the cascade has published precedent. The contribution is the integration around a compute anchor, not invention of the rungs.

Proven physics and individually ground-validated components (TRL ~4–6); none has yet been demonstrated on the lunar surface, and the contribution is the integration, not the rungs.

---

What Must Be Proven

This is open research, and three questions gate it. They are stated here rather than buried:

• The power balance. A published allocation of the 500 kW across compute, electrolysis, and thermal loads, settling how much water is stored versus electrolyzed to propellant at a given installed power.
• The cascade thermal model. Whether ~90 °C heat sustains the claimed extraction rate, validated against a distribution of plausible ice grades rather than a single point, since a 90 °C source cannot heat regolith above 90 °C and the budget tightens sharply below ~5 wt%.
• Destructive single-event latch-up. Soft errors are recoverable in software; latch-up can permanently kill an unscreened commercial part and is not stopped by ECC, scrubbing, replicas, or a regolith berm. It needs part-level beam characterization, addressed separately from the soft-error model below.

---

The De-Risking Artifact

The whole revenue thesis rests on commercial GPUs surviving the lunar polar radiation environment via ECC + lockstep + watchdog + a 50 cm regolith berm, not rad-hard silicon, which runs 10–100× slower than commercial parts and lags state-of-the-art by ~10 years. If that survival assumption fails, the compute thesis collapses.

cosmo-regulus is the Apache-2.0 open-source library that makes the assumption falsifiable. It anchors its fault model on measured Chang'E-4 LND lunar-surface dose data (13.2 ± 1 µGy(Si)/hr → ~116 mGy(Si)/yr unshielded) rather than CREME96 extrapolation, and produces an economic Pareto curve linking shielding mass × replica count × scrubbing rate → $/M-tokens at iso-quality. A first-cut version of that curve and its underlying numbers is at /lunar/cosmo-regulus.

---

Engagement

This branch is open research. The artifacts to evaluate at whatever depth is useful:

• Lunar whitepapers: the common-basis propellant-energy model plus two companion analyses (waste-heat offset, PSR co-location), in PDF.
• Full 15-section program plan (700+ lines): linked from the Contact page on request.
• cosmo-regulus open-source repository: github.com/dubthree/cosmo-regulus (Apache-2.0).
• Phase-0 deliverables and engineering analyses: covered in the program plan §12.

Centradiant welcomes partner conversations with satellite operators, launch providers, agency programs, and commercial-compute customers whose interest is in surface-based compute and resource infrastructure.