Why Is a Powerful AI Computer Being Sent to the Moon?
Why Is an AI Supercomputer Going to the Moon?
JUST IN: Hewlett Packard Enterprise $HPE is preparing to send a supercomputer to the Moon for the first time, bringing high-performance computing and AI directly to the lunar surface.HPE's Spaceborne Computer-4 will fly aboard Astrolab's FLIP robotic rover, scheduled to land near the Moon's south pole aboard Astrobotic's Griffin-1 mission in late 2026.Equipped with a CPU, GPU and solid-state storage, the compact system is designed to run AI applications and high-performance computing benchmarks in the Moon's extreme environment.The experiments will explore how lunar vehicles can process imagery, map terrain and support autonomous navigation without relying on Earth-based computers for every calculation.The mission will also test whether advanced computing hardware can survive the temperature extremes and operating conditions of the lunar surface.If successful, it could help establish the computing infrastructure needed for future lunar bases, robotic exploration and commercial operations on the Moon.
— Outer Space Today (@outerspacetoday) October 20, 2026
Meet Spaceborne Computer-4
NO WONDER ELON MUSK IS PUSHING FOR SCALING ENERGY AND AI COMPUTE IN SPACE STARTING WITH NVIDIA ORBITAL VERA RUBINSNVIDIA Vera Rubin NVL72One orbital Vera Rubin NVL72 is a whole AI factory: 72 Rubin GPUs and 36 Vera CPUs in one NVLink domain, about 3,600 PFLOPS of NVFP4 inference, ~20.7 TB of HBM4, and ~120–150 kW. No building. No grid. No water loop.That is the unit cell. Heat goes straight into vacuum through radiators. Power comes from solar that never sees night in sun-synchronous orbit. The same NVIDIA stack runs on the ground and on Starmind, so software does not fork when the rack leaves Earth.Scale is just launch cadence. At ~150 kW each, roughly 7,000 satellites are a gigawatt. Tens of thousands are multi-gigawatt. Starship flying dozens per flight turns that into a production problem, not a permitting problem. Land, substations, and cooling towers stop being the ceiling.Petawatt-year is the end state of that loop: continuous orbital power, rack after rack, year after year, outside Earth’s energy and real-estate limits. Vera Rubin is how SpaceX makes each satellite a coherent supercomputer instead of a box of chips, so the constellation can actually add up.NVIDIA Vera Rubin NVL72 — space-optimized AI computerA rack-scale system that behaves like one giant GPU. SpaceX and NVIDIA redesigned it for orbit: radiation-hardened, vibration-tolerant, radiator-cooled, and built to run unattended.🚩Configuration• 72 Rubin GPUs + 36 Vera CPUs in a single coherent NVLink domain• 18 compute trays plus NVLink switch trays• ConnectX-9 SuperNICs and BlueField-4 DPUs• Ground version: cable-free MGX trays. Orbital version: denser, lighter, simpler than a standard rack🚩Compute• 3,600 PFLOPS NVFP4 inference• 2,520 PFLOPS NVFP4 training• 1,260 PFLOPS FP8/FP6• 288 PFLOPS FP16/BF16• 9,360 TFLOPS FP32 / 2,400 TFLOPS FP64• Fits trillion-parameter and long-context agentic models inside one rack instead of a cluster• Cuts GPUs needed for giant MoE training by about 4x versus the prior NVL72 generation🚩Memory• 20.7 TB HBM4 (288 GB per GPU)• ~1,580 TB/s GPU memory bandwidth (~22 TB/s per GPU)• Up to 54 TB LPDDR5X on the CPUs• Large models stay in one fast, unified memory pool🚩Interconnect• NVLink 6: 260 TB/s all-to-all inside the rack• 3.6 TB/s per GPU• 65 TB/s NVLink-C2C between CPUs and GPUs• Scale-out over Spectrum-X Ethernet and InfiniBand• Non-blocking fabric — program the rack as one accelerator🚩CPU• 3,168 custom Olympus cores (88 per Vera CPU)• Built for agent orchestration, RL, data prep, code, and simulation• Up to 1.8x faster than x86 on those workloads, so GPUs stay on the model🚩Power, cooling, space• Terrestrial rack: fully liquid-cooled, roughly 180–230 kW• Orbital version rejects heat into vacuum through deployable radiators• Typical satellite compute envelope: ~120 kW average, ~150 kW peak• No facility water loops. No on-orbit servicing.🚩What it can do• Train and serve frontier models off Earth’s grid, land, and cooling limits• Run Grok-class agentic workloads in low Earth orbit• Keep the same NVIDIA software stack from ground data centers to Starmind satellites• Scale from one optimized payload to a constellation of orbital AI nodes https://t.co/KXsDoAvFXS
— 🌋🌋 Deep₿lueCrypto 🌋🌋 (@DeepBlueCrypto) October 25, 2026
The FLIP Rover Will Carry the Technology
AI Could Help Rovers Navigate Without GPS
Why Processing Data on the Moon Matters
The Moon's Extreme Conditions Are a Major Test
Could This Change Future Space Exploration?
Unlock insightful tips and inspiration on personal growth, productivity, and well-being. Stay motivated and updated with the latest at My Life XP