spacetech
View All Tags
接著 上一篇 Claw 堆疊筆記 之後，今天把 NVIDIA Agent Challenge 2026 的參賽計畫「Spacesharks Mission Desk」做了一次重要的重整。
題目沒換，stack 沒換，但我把這個 agent 要證明什麼 重寫了一次。
原本的版本是「把現有 Jamia / Spacesharks GPT 移植到 Nemotron stack」。今天的版本是：
一個低成本、多模型、能拿出證據、可以連跑 24 小時的衛星營運 copilot — 它的可信任性不是來自最大的模型，而是來自一個被刻意設計過的 trust stack。
這篇筆記紀錄這次重整的原因、四層信任架構，以及我為什麼把 scope 鎖到只有四個 ingestion source。
Taiwan is nearly a world-class leader in the upstream of the LEO satellite supply chain (RF PA, filters, high-frequency PCB), and has decent participation in the downstream (ground terminals, antennas) — but a structural void exists in the hottest new theme of 2025–2026: mid-stream C, Orbital Data Center (ODC) hardware integration.
This is not an accident. It is a structural opportunity window.
In March 2026, Jensen Huang said something on the GTC stage:
"Space computing, the ultimate frontier, has arrived."
This was not a metaphor. In November 2025, US startup Starcloud launched an NVIDIA H100 into orbit (Starcloud-1) and completed the first large language model training session in space. Two months later, on January 11, 2026, Axiom Space launched two orbital data center nodes (ODC Node 1 & 2), connected them to Kepler Communications' optical relay network, and began offering cloud compute services to external customers.
Earth orbit has officially become a new location class for data centers.
The traditional space industry has a simple answer to radiation: test everything, apply twice the design margin, and only use military-spec parts.
That approach made sense when launching a billion-dollar GEO satellite. It makes much less sense when you're a New Space startup trying to ship a CubeSat in nine months on a shoestring budget.
This post captures my current engineering thesis across AI systems, space infrastructure, blockchain, and RF validation.
General-purpose LLMs are powerful, but when it comes to deep analysis in specialized domains, they have two hard limitations: knowledge that lags behind the cutting edge, and factual dilution for niche fields. Take the Polkadot ecosystem as an example — key 2024–2026 developments like Agile Coretime, RegionX, and JAM are either unanswerable by general models, or return outdated information.
The solution is not a bigger model — it's giving the model a structured, trustworthy external knowledge base, then wrapping it with a persona and task prompt. That's the motivation behind my myGPTs.
This page updates my radiation test blog into a practical engineering workflow and includes simple simulation tools.
These two calculators are for quick mission trade studies, not qualification signoff.
This page updates my radiation test blog into a practical engineering workflow and includes simple simulation tools.
These two calculators are for quick mission trade studies, not qualification signoff.