Welcome to Issue 23, where Hewlett Packard Enterprise plans AI compute for the Moon
AI moved into more parts of space work this week: a compute module booked for the Moon, an AI anomaly test run on a CubeSat in orbit, and satellite embeddings released for the whole planet.
On the Space4AI side, Hewlett Packard Enterprise plans to send Spaceborne Computer-4, its new AI compute module, to the Moon on Astrolab's lunar rover. HPE says the mission is intended to launch in late 2026. MIT Technology Review put the flight as early as November 2026. HPE has not named a day.
On the AI4Space side, Ai2 released version 1.3 of its OlmoEarth satellite embedding model, with precomputed annual embeddings for the whole planet and a browser map tool that needs no sign-in. The release follows the OlmoEarth platform in Issue 11 and the Studio embedding exports in Issue 13. Korea Aerospace Industries completed an AI-based anomaly diagnosis test on a CubeSat that reached orbit on South Korea's Nuri rocket on October 7. Colorado-based WarpWare closed a $1.2 million pre-seed round for an AI engineering platform for space missions.
Specifics below.
Did you know?
Bits & Orbits Insights
Public dashboards on in-space compute, space autonomy, and the SpaceX lock-up
calendar:
Space4AI
Hewlett Packard Enterprise announces plan to send an AI compute module to the Moon
Astrolab's FLIP rover, which will carry Spaceborne Computer-4, driving off a Griffin lander (artist's rendering). Credit: Astrolab
On October 20, 2026, Hewlett Packard Enterprise announced that Spaceborne Computer-4, its new AI compute module, will fly on Astrolab's FLIP rover (FLEX Lunar Innovation Platform). HPE says the mission is intended to launch in late 2026 under NASA's Commercial Lunar Payload Services (CLPS) initiative. Voyager Technologies' Griffin-1 lander is to touch down near the lunar south pole and deliver the rover, with the module aboard.
MIT Technology Review reported the same day that the flight is slated for as early as November 2026. HPE's October 20 release does not name a launch day.
On the lunar surface, Spaceborne Computer-4 will test large language models, visual and LiDAR data processing, and autonomous monitoring. NASA and Deloitte will also run AI workloads on the module. NASA will run a neural-net model trained on Earth, Moon, and ISS data to show how several separate AI models can work together. Deloitte will deploy its Silent Shield system to detect cyber intrusions, provide near-real-time analytics, and flag anomalies.
HPE flew three earlier Spaceborne Computers to the ISS in 2017, 2021, and 2024. Those missions tested whether commercial off-the-shelf parts could work in space, then added compute power, storage, and onboard AI. Computer-4 is the first in the series to operate in the vacuum of space. It carries a low-power CPU and GPU and stores data on KIOXIA solid-state flash drives. It has no fans and relies on a space radiator to stay cool. The system also monitors radiation and lowers its compute power during events that could damage the hardware.
HPE will also run its software at Astrolab's mission control to manage the automated workloads and speed up delivery of results. The stated goal is to prepare future lunar missions to use edge computing, including mapping the surface from in-situ data and showing that AI applications can be deployed reliably at scale. Mark Fernandez, HPE's chief scientist of space technologies and solutions, said the module lets the rover interpret the data it collects on site. He said this gives Astrolab, NASA, and Deloitte a way to test whether onboard computing can deliver useful results sooner and make better use of communications with Earth.
AI4Space
Ai2 releases OlmoEarth v1.3 satellite embedding model and global embeddings
OlmoEarth v1.3, released October 22 with precomputed embeddings for the whole planet. Credit: Ai2
On October 22, 2026, Ai2, the Allen Institute for Artificial Intelligence, released the OlmoEarth v1.3 satellite embedding model. It also released precomputed annual embeddings for the whole planet at several resolutions. A browser-based map-making tool called OlmoEarth Embeddings Explorer launched alongside them and requires no sign-in.
The model takes a year of Sentinel-1 radar data and Sentinel-2 and Landsat visible and infrared imagery. It compresses those observations into a numerical fingerprint for each 10 m (33 feet) by 10 m (33 feet) patch of ground at its finest resolution, and two coarser resolutions are also available. Other software can use these fingerprints to tell crops, forests, mines and other land cover apart. Ai2 says v1.3 is its cheapest OlmoEarth model to train: it used 53% of the compute time of v1.2 and 18% of v1.
We covered the July 28 launch of the OlmoEarth Platform in Issue 11. That release's roadmap included a dedicated embedding model with precomputed global embeddings. Issue 13 then covered the August 12 custom embedding exports from OlmoEarth Studio, which let a user compute embeddings for a chosen region and export them as GeoTIFFs.
Ai2 says a fine-tuning run in OlmoEarth Studio takes hours. A first pass in the new explorer takes seconds. Analyzing a region from a handful of labels needs no GPU. The embeddings can show whether a task deserves a fine-tuning run and where the model falls short.
In one example, Ai2 compared embeddings across years in the La Pampa mining belt in Peru's Madre de Dios region, with no model training. That area lost more than 1,200 hectares (2,965 acres) of forest to illegal gold mining in 2017 and 2018, until Peru's Operation Mercurio shut most of it down in February 2019. The comparison shows 34 km² (13 square miles) of significant change from 2016 to 2020, falling to 9 km² (3.5 square miles) from 2020 to 2024.
KAI demonstrates onboard AI anomaly analysis on BEE-1012 CubeSat in orbit
On October 21, Korea Aerospace Industries (KAI) announced that it had completed an on-orbit pilot mission on October 20 to test AI-based anomaly diagnosis. The test ran on BEE-1012, a CubeSat that carries an AI module KAI developed in-house. The satellite reached orbit on the fifth launch of South Korea's Nuri rocket on October 7.
The test had three stages. A ground station sent a simulated fault signal. The onboard AI module analyzed the cause and scope of the anomaly. It then sent a technical report with proposed countermeasures back to the ground station. KAI built the module by adapting the fault diagnosis, life prediction and predictive maintenance technology it developed for aircraft. The module was trained on tens of thousands of fault scenarios. Defense AI company Funzin worked with KAI on the space software, and South Korean startup Mobilint supplied the module's neural processing unit (NPU). The project grew out of a memorandum of understanding that KAI and Space Lintech signed in March to run AI demonstrations on CubeSats. Yonsei University also took part.
KAI plans to start the main autonomous operations demonstration in December to further test the technology's reliability. The company also plans to make the technology a core standard specification for future satellite systems it develops. Seo Hyun-seok, executive vice president and head of KAI's Space Business Division, said the pilot confirmed that the technology works in the space environment. He said KAI will build AI into its space systems and apply it across its satellite business.
WarpWare closes $1.2M pre-seed for AI agentic flight software engineering platform
On October 19, 2026, WarpWare closed a $1.2M pre-seed round. The company is based in Colorado. It will use the money to expand what it calls an AI-enabled engineering platform for space missions. The round was raised through SAFEs and came mostly from angel investors. Backers included the Rockies Venture Club, Cowtown Angels, the Aggie Angel Network, individual angels, and two space-focused early-stage funds, Vesper Alpha and Marlboro Lane Ventures.
The founders started WarpWare while they were working at Sierra Space. They had asked how much of that company's flight software could be reused for a new vehicle. CEO and cofounder Alex Reynolds and COO and cofounder Mickey Mathew said the answer was almost none. Reynolds said roughly 80% of the space software stack is expensive and slow to build, and it is not proprietary. WarpWare's platform covers flight software, guidance, navigation and control (GN&C), simulation, and testing. A space-specific AI harness connects to these tools and automatically runs multilayered testing on what the AI produces. Engineers decide where AI is used, which work stays fully human-led, and when an output is ready to move forward.
WarpWare says more than 20 organizations use the platform. Their missions include in-space power beaming, NASA science missions, cislunar missions, and launch vehicles. LEAP flew the full platform on its Bullfrog rocket. WarpWare says this cut LEAP's timeline to launch by 65%. The company is also supporting Argotec on NASA's MiniCOR space-weather mission. WarpWare officials said the new funds will bring its agentic-engineering tools to more commercial and government space markets. The money will also build out sales and marketing.
Till next time,
Meta-beat Column of this week
Read also about the AI Pipeline that sits at the core, producing this Newsletter, including its ups and downs of this week: