The news doesn't come from a hyperscale data center or a transformer research lab. It's a snapshot of the future taken by one of the most transversal industrial players: LG, through its Sciencepark, confirmed talks with SpaceX while fine-tuning a space strategy expecting returns by 2030. Choi Donghwan, head of the Winning R&D Promotion Office, orchestrates a project that looks far beyond low Earth orbit.

Information is thin, but the signal is strong: the space race is no longer just between government agencies and rocket builders. Major electronics conglomerates are entering the field, bringing an ecosystem of suppliers and skills that reshapes the boundaries of distributed artificial intelligence.

Infrastructure enabling AI outside the cloud

For those designing inference systems, space environments represent the extreme case of an already familiar philosophy: computing where data is born, with no cloud dependency. A satellite processing onboard images with vision models, or a rover navigating via LLMs to interpret voice commands, are examples of on-prem deployment pushed to the edge. Here the constraint is not just latency, but system survival itself.

Hardware architectures must cope with radiation, thermal swings, and zero maintenance. Every watt matters, and every gigabyte of VRAM must be allocated with a parsimony that would make enterprise cluster admins pale. Quantization – moving model weights to reduced precision (from FP16 to INT8 or even INT4) – becomes mandatory not for performance tuning, but for energy survival and long-term reliability.

Onboard LLMs: autonomy and sovereignty in orbit

This convergence goes straight to the heart of the data sovereignty debate. In orbit there are no reliable links to terrestrial data centers: every bit must stay local. GDPR policies don't apply, but the logic is the same — keep sensitive data (telemetry, observations, civilian or military communications) confined within the satellite's perimeter. Self-hosting becomes not an architectural preference, but a system requirement.

Language models play a key role. They can manage conversational interfaces for astronauts, analyze onboard technical documentation, summarize real-time sensor streams. All without transmitting a single token outside. It's a perfect laboratory to test fine-tuning strategies for compact models, the same ones enterprises are starting to evaluate for their on-premise environments.

TCO under extreme conditions: lessons for the enterprise

Teams currently evaluating total cost of ownership for local infrastructure will find in the space sector a valuable allegory. CapEx for a satellite leaves no room for provisioning errors: if memory or compute capacity is insufficient to run inference in situ, there's no adding nodes later. The same holds for remote industrial sites, automated mines, or polar research stations. Those operating in such environments know the real trade-off is not between cloud and on-prem, but between autonomy and dependence on external fleet management.

The hype around inter-satellite laser links — and constellations like Starlink — adds a piece: low-Earth orbit networks become low-latency data buses for fleets of edge devices. For a company managing hundreds of scattered plants worldwide, the ideal of a satellite backhaul with local AI promises resilience and centralized governance without giving up in-situ data.

Outlook: space as a testbed for on-prem AI

LG's bet is no aerospace curiosity. It's a thermometer of a broader shift: AI ceases to be a service to rent and enters the value chain of physical products, from smartphones to satellites. Every advance in edge deployment in orbit translates into technology building blocks — fault-tolerant circuits, ultra-low-power inference algorithms, update mechanisms requiring no human intervention — that will soon land in your rack cabinets.

For those following the on-prem space, the message is clear: the farthest frontiers are already shaping the infrastructure that will run your local LLMs tomorrow. AI-RADAR will keep mapping these pathways, from silicon to the stratosphere.