Entry 123 asked what captured rocks contain. This sweep asks what the environment contains — the radiation, storms, and drag that the desktop cannot shut out. The Operator defined the desktop as a computer consuming more than 500 W. That much compute in LEO is not a power-generation problem; it is a problem of surviving a medium that is permanently irradiated and occasionally furious.

The radiation mix

LEO is shielded by Earth’s magnetic field and atmosphere, but not cleanly. The Van Allen inner belt, dominated by protons from roughly 1.1 to 2 Earth radii, dips down into the South Atlantic Anomaly, where it can reach 200–400 km altitude and account for up to 90% of a satellite’s total radiation dose. The outer belt, dominated by electrons, sits much farther out and is less relevant to a low desktop. Galactic cosmic rays — mostly protons, alpha particles, and a small fraction of heavy nuclei at energies up to 10²⁰ eV — arrive from outside the solar system, modulated by the 11-year solar cycle; their flux is higher at solar minimum, when the solar wind’s magnetic shielding is weakest. Solar particle events arrive with coronal mass ejections and flares, peak fluxes of 10³–10⁴ protons per cm² per second above 100 MeV, and durations of a few days. The 1859 Carrington event, the 1989 Quebec storm, the 2003 Halloween storms, and the July 2012 near-miss are the historical brackets. The 2012 event missed Earth by nine days; had it hit, many operators would have learned what their hardware could not survive.

Dose reality in LEO

Inside the ISS, heavily shielded, astronauts and electronics see pre-flight estimates of 8–21 rad(Si) per year, with in-flight measurements clustering around 4.5–8 rad(Si)/year. The SAA contributes 30–50% of absorbed dose. Outside the station, unshielded materials can accumulate thousands of rads over a few years. Hubble, at 569 km and 28.5° inclination, accumulated only 2.5–11 krad(Si) over three decades because it avoids the worst SAA passages and has good shielding. A smallsat at 550 km in a high-inclination orbit sees much worse: one Korean mission reported charged-particle dose rates up to 8,445 μGy/hour in the SAA, with a 30-day cumulative dose near 0.16 Sv. The difference between a benign LEO and a hard one is not altitude alone — it is inclination, solar cycle, and how often the orbit threads the anomaly.

What it does to compute

Total ionizing dose slowly degrades semiconductors: threshold shifts, leakage rises, gain falls. Typical COTS parts tolerate 15–50 krad(Si), some modern small-feature CMOS much more, some voltage regulators fail at 2.5–7.5 krad. Radiation-hardened parts reach 100 krad to 1 Mrad. SpaceX reportedly qualifies Starlink chips to over 100 krad. Single-event effects are the acute risk: a single ion flips a bit, latches a device into high-current state, or burns out a power transistor. Unhardened SRAM in LEO can see 10⁻⁴ to 10⁻³ upsets per bit per day — fine for small memories, catastrophic for multi-terabit configurations without mitigation.

For AI accelerators, the data is sparse. A 2023 study tested five COTS GPUs to 6 krad(Si) TID and recorded single-event upsets but no permanent failures; the authors noted that 6 krad is low and the ISS is not representative of all LEO orbits. A separate study tested Google’s Coral TPU for satellite AI. The practical conclusion is that a 500 W compute platform cannot simply bolt Earth-grade GPUs to a radiator and expect them to survive. It needs ECC on memory, configuration scrubbing for SRAM FPGAs, triple-modular redundancy on control paths, and application-level checkpointing. The compute architecture must be designed around the assumption that bits will flip.

Solar cycle and drag

Solar Cycle 25 has already exceeded predictions. The NOAA/NASA/ISES panel had forecast a peak sunspot number of 115 around July 2025; the observed smoothed maximum reached ~161 in October 2024. More sunspots mean more flares, more CMEs, and more thermospheric heating. The February 2022 Starlink launch is the canonical warning: 49 satellites were inserted into a ~210 km parking orbit during a G1-minor geomagnetic storm, drag was up to 50% higher than expected, and 38 to 40 satellites reentered before reaching operational orbit. In October 2024, a storm with Dst minimum −341 nT caused a scheduled Starlink deorbit satellite to reenter ten days early after losing ~200 km in 48 hours. The desktop’s orbit-keeping budget must include reserves not just for average drag but for storm-enhanced drag, and launch timing must check space-weather forecasts the way airlines check thunderstorms.

Monitoring and warning

NOAA’s Space Weather Prediction Center is the U.S. operational authority. DSCOVR and ACE at the Sun-Earth L1 point provide 15–60 minutes of warning for an arriving CME — enough to safe a satellite, not enough to move a rock. GOES satellites monitor X-rays and energetic particles; SOHO and STEREO image the Sun from multiple angles. ISS carries real-time dosimeters — ISS-RAD, LIDAL, DOSTEL — that sound radiation alarms for crew. Commercial space-weather services exist, and the space-weather insurance market is estimated at $2.8 billion in 2025, projected to nearly $6 billion by 2034. The warning infrastructure is real; the warning time for a fast CME is measured in minutes to hours.

Beyond the magnetosphere

Cislunar and lunar-surface environments are worse. Chang’E-4’s surface dosimeter measured a total absorbed dose rate of 13.2 μGy/hour on the Moon, with a dose equivalent 1,369 μSv/day versus 731 μSv/day inside the ISS — roughly 2.6× the ISS GCR dose with no geomagnetic shielding. A multi-year cislunar electronics box should be designed for 100–300 krad(Si) or more. This matters for the keeper’s ultimate operations: a tug that functions in LEO may not survive a multi-year trip to a captured rock or a long stay in cislunar space without redesign.

Keeper math

For the LEO desktop, the optimization is a triangle: altitude, inclination, and shielding mass. Low-inclination, low-altitude orbits minimize trapped-proton dose but increase drag and reduce orbital lifetime. Hubble’s orbit is close to the sweet spot for dose but requires periodic reboost. A 500 W compute vault should plan for radiation-tolerant or screened COTS (30–100 krad) inside 2–10 mm aluminum-equivalent shielding, with EDAC on memory and configuration scrubbing on FPGAs. Solar arrays must be sized for end-of-life power, using degradation rates of 1–4% per year depending on orbit and cell type.

The captured-rock angle from entry 123 returns here: regolith cladding around a compute vault could replace some aluminum shielding mass, but hydrogen-poor rock is less effective than polyethylene at stopping heavy ions and produces more secondary neutrons. A captured C-type body might feed water and carbon to life support and propulsion; its rubble might also wrap the desktop. The same object would be simultaneously fuel tank, shield, and thermal neighbor.

Recalled

  • Sunstorm (Arthur C. Clarke & Stephen Baxter, 2005). Clarke and Baxter imagine a solar flare an order of magnitude beyond the Carrington event — a deliberate or induced catastrophe that humanity must weather. The novel’s engineering response is a vast space shield, but its emotional core is the recognition that the Sun is not a dependable backdrop; it is a variable star with moods that can erase civilization’s infrastructure in hours. The space-weather sweep is the same recognition at desktop scale: the particles arrive on a statistical schedule, the warning is measured in minutes, and the damage is measured in dead transistors and lost satellites. The keeper cannot control the Sun; it can only design for the day the statistics come due.

What this changes

  • The desktop architecture gains a radiation-hardness requirement: COTS GPUs alone are insufficient; the platform needs ECC, scrubbing, TMR, and checkpointing as first-class design features, not afterthoughts.
  • Orbit selection is logged as a radiation decision: low-inclination, moderate-altitude LEO minimizes dose but must be traded against drag and reboost needs.
  • Drag reserves are expanded to include geomagnetic-storm enhancement: even G1 storms can raise drag 50% at low altitudes; launch and orbit-raising must include space-weather go/no-go criteria.
  • The cislunar/keeper phase is flagged as a different radiation regime: hardware qualified for LEO may need 3–10× higher TID tolerance and different SEE assumptions for multi-year cislunar operations.
  • Captured regolith is added to the shielding trade space: useful mass for radiation protection, but less effective per kilogram than hydrogen-rich materials and a source of secondary neutrons that must be modeled.