When an astronaut on a Mars-bound journey feels a sharp abdominal pain, they can’t call Houston. Communication delays and blackouts render traditional telehealth useless. NASA is tackling this challenge with the Crew Medical Officer Digital Assistant (CMO-DA), an AI system running entirely onboard the spacecraft, leveraging Large Language Models for local inference.

No-cloud architecture: AI in the silence of space

The project began as a cloud-connected proof of concept but was quickly moved to a fully disconnected edge deployment. The CMO-DA runs on HPE hardware, tested on the terrestrial twin of the Spaceborne Computer currently operating aboard the International Space Station. The goal: enable a crew member to diagnose and treat symptoms by querying spaceflight medical literature through Retrieval-Augmented Generation (RAG), without any real-time link to Earth.

RamaLama: AI models as portable, verifiable artifacts

The inference stack relies on llama.cpp via RamaLama, an open-source command-line tool backed by Red Hat. RamaLama wraps multiple inference engines—llama.cpp, MLX, vLLM—and lets users pull and run models much like container images, with automatic GPU detection and cryptographic verification. Models become portable, reproducible artifacts, a critical feature when deploying to unreachable hardware that demands accountability with every update.

Why the on-premise choice echoes on Earth

NASA’s decision isn’t just a space-constrained necessity; it points to a broader shift for enterprise deployments. Local-first architectures that treat models as immutable, cryptographically verifiable components address similar needs in finance, defense, and healthcare, where air-gapped or regulated environments require complete data control. The reproducibility of the deployment and the absence of cloud dependencies cut operational risk and ease compliance.

What space teaches us about edge AI

CMO-DA proves that LLM inference can work under severe constraints, on modest hardware, without external connectivity. For organizations evaluating self-hosted solutions, this case highlights the trade-offs between autonomy and management complexity—and underscores the strategic value of an AI infrastructure that keeps both data and critical decisions within physical boundaries. NASA’s experiment is an extreme test of technological sovereignty, one that speaks directly to anyone planning to bring AI inside their own secure walls.