Frequently Asked Questions

Rad-hard silicon costs two to three orders of magnitude more per FLOP than commercial parts and is produced in volumes incompatible with 500 kW of compute. The economics collapse before the first cascade stage.

The polar radiation environment, while harsher than Earth orbit, is manageable for soft errors with commercial mitigation: ECC memory, lockstep CPU execution, hardware watchdogs, and a 50 cm regolith berm providing approximately 20 g/cm² shielding. The key question is whether the soft-error rate under that shielding keeps GPU availability above the threshold needed for the revenue model. Cosmo Regulus exists to answer that question with data anchored to Chang'E-4 LND measurements rather than orbital proxies. One risk it does not resolve is single-event latch-up: a heavy-ion-induced latch-up can permanently destroy an unscreened commercial part, is not caught by ECC, scrubbing, or replicas, and is not stopped by a 20 g/cm² berm. That requires part-level beam testing and is tracked as a separate open item.

GPU waste heat exits cold plates at 90 °C. That heat goes to three downstream uses:

• Water extraction: thermal energy sublimates ice in the regolith of the Permanently Shadowed Region. Vapor condenses in a cold trap as water; where electrical power allows, it is electrolyzed to hydrogen and oxygen.
• Organic Rankine cycle: a low-temperature ORC converts waste heat to electrical power during eclipse periods when solar input drops.
• Regolith thermal energy storage: excess heat is stored in regolith TES beds, buffering the cascade through thermal transients.

The cascade is not designed to be efficient in the thermodynamic sense. It is designed to be self-funding. Compute revenue covers capital and operations. The near-term product is water and oxygen: the ~90 °C waste heat does the thermal extraction directly. Turning that water into chemical propellant is 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 scales with installed power rather than coming free with the heat.

The Artemis Accords establish a framework for commercial lunar surface operations. As of 2026, the U.S. has no equivalent to FCC spectrum licensing or FAA launch licensing for permanent surface installations, but the regulatory trajectory is consistent with commercial activity at the Connecting Ridge on a 2030s timeline.

The lunar program is not on a regulatory critical path at this stage. The near-term work is science and engineering: characterizing the radiation environment, validating the cascade thermal model, and establishing the compute revenue baseline. Regulatory engagement follows demonstrated technical feasibility.

Cosmo Regulus is an Apache-2.0 library for modeling galactic cosmic ray and solar energetic particle flux at the lunar pole, anchored to Chang'E-4 Lunar Neutron and Radiation Detector (LND) data. It generates synthetic soft-error rate timeseries for commercial GPU architectures under configurable shielding configurations.

It is open-source because the underlying claim, that commercial GPUs survive the polar environment with standard mitigation, needs to be independently reproducible. A proprietary model that says "trust us" is not a scientific contribution. A public model that accepts peer critique is.

The secondary benefit: researchers working on lunar surface power, ISRU, or commercial lunar infrastructure can use Cosmo Regulus without reinventing the radiation model from scratch.

Only loosely. Centradiant also runs a separate orbital thermal program, but the two are independent on every dimension that matters here: funding, timeline, regulatory path, and customer. This lunar program does not depend on the orbital one, cites none of its performance figures, and stands on its own. It is documented separately and that work evolves on its own track.