The First LLM on the Moon: HPE's Shoebox Supercomputer

By Daniel Rao | October 21, 2026

NASA's Lunar Reconnaissance Orbiter mosaic of the Moon's south pole region, near where Astrolab's FLIP rover is due to land. Image: NASA/GSFC/Arizona State University (public domain).

As early as next month, ChatGPT-class AI will be running on the Moon. Not in a simulation, not on a server rack in Houston — on the actual lunar surface, inside a 1.2-kilogram computer the size of a small shoebox, rolling across the most hostile real estate in the inner solar system. Hewlett Packard Enterprise announced the mission on October 20: its Spaceborne Computer-4 will ride Astrolab's FLIP rover to the Moon's south pole, where it will run large language models, analyse LiDAR maps, and submit what would be the first Top500-style supercomputing benchmarks ever performed on another celestial body.

The stakes, in a sense, were set more than half a century ago. "The time delay, for example, on Mars can be up to 20 minutes... you are looking at stale data. In fact, the Soviets lost two rovers in the past because of operator errors caused by the time delay," said Mark Fernandez, HPE's chief scientist and principal investigator, during a press conference introducing the project. His point was blunt: every robot that must wait for permission from Earth is a robot driving half-blind. HPE wants to put the brain where the wheels are.

The mission: FLIP, Griffin-1, and a ride on a Falcon Heavy

FLIP — the FLEX Lunar Innovation Platform — is Astrolab's first rover under a NASA contract, the precursor to the larger buggies that will eventually carry Artemis astronauts. At 480 kilograms, with deformable wheels built for regolith and a design rated to survive at least 150 hours of the frigid lunar night, it is a serious machine. It will carry roughly 20 scientific payloads, including a LiDAR 3D mapper and a sensor for measuring concentrations of helium-3 in lunar soil.

The rover travels inside Voyager Technologies' Griffin-1 lander, flying under NASA's Commercial Lunar Payload Services programme, launching on a SpaceX Falcon Heavy from Launch Complex 39A. The destination: Mons Mouton, near the western edge of Nobile Crater, about 100 miles from the lunar south pole — the same general neighbourhood where Intuitive Machines' two Nova-C landers both tipped over. The intended timeline is tight: the mission is meant to launch late 2026, "as early as November," with the rover expected to operate on the surface for about ten days.

That timeline deserves honest handling. Griffin-1 has been delayed repeatedly — originally a November 2023 launch, then November 2024, then 2025, then July 2026, and now the back end of 2026 — after NASA cancelled its original VIPER rover payload in 2024 and Voyager agreed to acquire Astrobotic mid-2026. The hardware just cleared environmental testing at NASA's Jet Propulsion Laboratory, announced October 13, and is now back in Pittsburgh for final assembly before shipping to Cape Canaveral. That is real progress. But veteran space watchers are already betting on another slip, and this mission's "as early as November" should be read as exactly that: the earliest possible date, not the plan.

The computer: off-the-shelf ambition in a fanless box

Spaceborne Computer-4 is, physically, almost comically modest for the job it is being asked to do. It is a fanless module measuring 249 by 132 by 26 millimetres, weighing 1.2 kilograms, pairing a low-power 64-bit CPU with a GPU for parallel workloads, running a Linux environment with KIOXIA flash storage in its solid-state drives. It plugs into the rover over standard Ethernet — the same kind of port you'd find on an office switch — takes 28-volt DC power, and sheds heat through its aluminium chassis into the rover's own radiative cooling system.

What it is scheduled to run is the news. The manifest calls for large-language-model inference with retrieval-augmented generation, vision inference on camera data, LiDAR processing for mapping and navigation, and — the headline for the supercomputing world — the first High-Performance LINPACK and High-Performance Conjugate Gradient benchmarks on the lunar surface. HPL measures raw floating-point muscle; HPCG punishes memory access and communication patterns. Running both on the Moon would put the first off-Earth data point into the benchmarking world that decides which machines count as supercomputers.

HPE has not disclosed the CPU or GPU models, memory, storage, or power consumption — so until the benchmarks actually run, "high performance" remains an unanswered question. The mission itself is the experiment that answers it.

Why this matters: cutting the Earth–Moon command loop

Light takes about 1.3 seconds to reach the Moon, so commands and telemetry make a roughly 2.6-second round trip — manageable, but the data is always stale, and bandwidth is a trickle. On Mars, the round trip stretches to up to 40 minutes. Fernandez's Soviet-rover example is the cautionary tale: Lunokhod 1 was crippled in 1971 after driving into a crater, and Lunokhod 2 died in 1973 when regolith dust baked onto its radiators — both missions that depended on humans steering with delayed video. The principle hasn't changed in fifty years. A robot that can think locally doesn't wait.

"The next chapter of lunar exploration will depend on the ability to turn data into insights at the point of discovery," said Scott Wiest, HPE's vice president, fellow and global sales chief technologist. In practice that means FLIP's computer will analyse camera and LiDAR data onboard to aid navigation in terrain it has never seen, distil what it learns, and transmit back insights instead of raw gigabytes. On a rover operating ten days in a place where every watt and every bit is rationed, that is the difference between a technology demonstration and an actual science mission.

NASA is leaning into the same logic. Its FLUID experiment — Federated Learning for Unbounded and Intelligent Decentralization — will run an AI neural-network model coordinated across three nodes: Earth, the Moon, and the International Space Station. NASA will also test a tuned language model for searching and interpreting documents aboard the rover. Deloitte, meanwhile, is flying its Silent Shield system to detect cyber intrusions and provide near-real-time anomaly detection — because if AI is going to live on the Moon, somebody has to secure it.

The engineering bet: software instead of armour

The most radical thing about SBC4 is not what it does but what it refuses to be. The standard approach to space electronics is radiation hardening: specialised chips, years of qualification, eye-watering cost, and performance stuck in the late 1990s — the computers on NASA's Mars rovers are routinely compared to laptops from the Clinton era. HPE went the other way. It used commercial parts and built software that watches for upsets caused by charged cosmic particles, and when radiation spikes, the software throttles the computer's power and speed to protect the electronics.

"In space, we don't have the redundancy, we don't have the repairman, we don't have the spare parts," Fernandez said — a line that, delivered flat at a press conference, tells you everything about the design philosophy. The approach already has a track record: three earlier Spaceborne computers spent more than two years each aboard the ISS, and the Space Llama initiative ran Meta's Llama 3.2 on the station in 2025 to help astronauts with research and decision-making.

The International Space Station, photographed in 2009. Three of HPE's earlier Spaceborne computers operated aboard the station, including the Space Llama tests that ran Meta's Llama 3.2 in orbit. Photo: NASA (public domain). Image used to illustrate the ISS, not the computers themselves.

The ISS, though, is a sheltered environment at a balmy 20°C with Earth's magnetic field overhead. The Moon offers swings from roughly 80°C in sunlight to minus 200°C in shadow, and radiation levels nearly three times as intense as inside the station.

This is where the economics get interesting, and it is the part of the story the press releases don't say out loud. If a $1,200-gram commercial-part computer with clever software survives the lunar night and the radiation, the entire cost structure of space computing flips. Google is already testing AI chips in orbit with its Project Suncatcher TPU satellites, and Lunar Outpost is flying Nvidia Jetson modules to the lunar surface on a separate mission — but those are edge-inference plays. HPE is attempting something brasher: full HPC-class workloads and LLMs, validated by the industry's own benchmarks. If the benchmarks check out, the days of paying the space-hardened premium for every transistor could be numbered.

The scepticism is earned, though. LLMs hallucinate on Earth; on the Moon, as one AI and space-law researcher at the University of Valencia noted to MIT Technology Review, even a small error can have serious consequences. And the AI software here will be "specialized custom-built LLMs," not the open weights everyone knows — which means independent scrutiny of what the models actually do will have to wait for the mission's results.

What this proves — if it works

Strip away the superlatives and this is a plumbing story, and plumbing is what matters. A permanent lunar base cannot be operated like a remote-control car; the decision cycle is too fast and the bandwidth too thin. The same architecture that lets FLIP navigate a crater rim on its own is the architecture that will run a habitat's life-support diagnostics, process regolith-mining sensor data, and triage satellite imagery before anything gets downlinked. Mars, where Fernandez's 20-minute delay makes Earth control physically impossible for anything urgent, is the real destination. The Moon is the dress rehearsal.

The commercial logic is already moving money. Space startups banked a record $20.8 billion in a single quarter this year, and NASA's Moon Base programme is Phase One of an infrastructure build-out, with Griffin-1 as its next act after a string of private landing failures. SpaceX's $40 billion bet on space-bound chips points at the same conclusion the HPE mission does: computing is moving off the planet, and the companies that define its architecture will set the terms of the lunar economy.

The honest verdict on October 21, 2026: nothing has flown yet. The computer is a payload, the launch date is an aspiration with a troubled history, and the benchmarks are a promise. But the promise is a serious one — the first LLM on the Moon, the first Top500-style benchmark off Earth, and a design philosophy that could make every future lunar computer cheaper, faster, and autonomous. The Soviets learned the hard way that you can't steer a rover from a quarter-million miles away. Half a century later, HPE is finally building one you don't have to.

Sources: HPE press release and technical backgrounder (Oct 20, 2026); MIT Technology Review (Oct 20, 2026); Converge Digest (Oct 21, 2026); Tech Times and Eastern Herald on Griffin-1 JPL testing (Oct 13–14, 2026); NASA on FLIP payloads; we checked HPE's release against MIT Technology Review's reporting and the HPE technical backgrounder for spec consistency. Quote sources attributed inline.

---

About the author: Daniel Rao is TechPulse USA's science and space correspondent, covering AI, computing and commercial spaceflight. He previously reported on semiconductor supply chains and NASA's Commercial Lunar Payload Services programme.