Radiation-hardened AI processors for spacecraft
Flight-qualified processors that run modern neural networks onboard spacecraft for years in deep-space radiation without upsets.
State of the art
NASA HPSC (Microchip PIC64-HPSC) in qualification testing (2026); commercial GPUs fly in LEO only with shielding and short mission lives.
- Rad-hard processes lag commercial nodes by several generations
- Single-event upsets in large memories corrupt model weights
- Power and heat rejection limits on spacecraft
Who’s working on it
- Microchip Technology ↗Company
Holds the NASA HPSC contract (awarded 2022) and shipped PIC64-HPSC early-access samples in 2025, targeting 100x current spaceflight compute
- NASA High Performance Spaceflight Computing (HPSC) programProgram
JPL-led programme that began radiation, thermal and shock testing in 2026; NASA's FY26 shortfall assessment ranked onboard computing its third-highest civil space technology gap
- NASA Jet Propulsion Laboratory ↗Government lab
Defined the HPSC requirements from future mission needs and runs the qualification campaign that decides whether the part is flight-usable
- SiFiveCompany
Supplies the RISC-V cores inside PIC64-HPSC, which is why the first modern rad-hard AI processor is RISC-V rather than PowerPC
- Ubotica Technologies ↗Startup
Irish company whose hardware ran the first in-orbit AI inference on ESA's PhiSat-1 (2020) using a commercial Myriad 2, proving the COTS-in-orbit alternative
- BAE SystemsCompany
Builds the RAD5545 class of flight-proven rad-hard processors that HPSC must displace, and the reference point for how slow qualified parts currently are
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| engineering | Rad-hard accelerators with error-corrected HBM sized for billion-parameter models | $$$ $100M – $1B |
Depends on
A root node: no prerequisites modeled.