3.5 KiB
| type | title | author | url | date | domain | secondary_domains | format | status | processed_by | processed_date | priority | tags | extraction_model | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| source | Starcloud Trains First AI Model in Space — NVIDIA H100 GPU in LEO, December 2025 | CNBC (@CNBC) | https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html | 2025-12-10 | space-development | article | processed | astra | 2026-04-14 | high |
|
anthropic/claude-sonnet-4.5 |
Content
Starcloud launched Starcloud-1 in November 2025, carrying the first NVIDIA H100 GPU into space. In December 2025, the company announced that the satellite had successfully:
- Trained NanoGPT (Andrej Karpathy's LLM) using the complete works of Shakespeare
- Run inference on a version of Google Gemini from orbit
- Fine-tuned an AI model in orbit
Technical specs of Starcloud-1:
- 60 kg satellite
- Based on Astro Digital's Corvus-Micro bus
- 325 km circular orbit
- Expected mission lifetime: 11 months (de-orbits and burns up)
- The H100 GPU is 100x more powerful than any GPU previously operated in orbit
Four industry firsts claimed: first H100 in space, first AI model trained in orbit, first orbital Gemini inference, first orbital model fine-tuning.
NVIDIA co-invested in Starcloud. Mission objective: determine whether data-center-grade GPUs can operate reliably in space radiation environment, vacuum exposure, and thermal cycling.
Agent Notes
Why this matters: This is the most concrete TRL validation for the ODC sector's central claim — that commercial-grade GPUs (not radiation-hardened military chips) can operate in LEO. The H100 demo at 325km altitude establishes TRL 7 for the LEO radiation environment at that altitude.
What surprised me: The 11-month expected mission lifetime. This is very short for any commercial system. At 325km, the orbital lifetime is naturally limited by atmospheric drag — de-orbit is natural and expected. But it also means we don't know what the long-term radiation degradation curve looks like for H100-class chips.
What I expected but didn't find: Any data on radiation-induced errors (single event upsets, bit flips) during operation. NVIDIA and Starcloud report "successful operation" but haven't disclosed error rates or performance degradation vs. terrestrial baselines.
KB connections: Validates the hardware feasibility component of ODC claims. But 325km is a much more benign radiation environment than the 500-1800km altitudes proposed by SpaceX and Blue Origin (well inside Earth's magnetic shielding, below the Van Allen belts' intense zone).
Extraction hints:
- Claim candidate: Starcloud-1's successful H100 operation in November-December 2025 establishes commercial GPU viability at 325km LEO but does NOT validate the 500-1800km radiation environment proposed for large-scale ODC constellations.
- Key scope condition: this demonstration is altitude-specific and duration-limited (11 months is not long-term reliability).
Curator Notes
PRIMARY CONNECTION: Starship achieving routine operations at sub-100 dollars per kg — the ODC cost case depends directly on Starship pricing, and this demo is the proof of concept that makes the case real. WHY ARCHIVED: The seminal ODC hardware proof-of-concept. Sets the TRL baseline for commercial GPU in space. EXTRACTION HINT: Focus on the altitude-environment gap (325km vs. 500-1800km) as the key caveat that limits what this demonstration proves.