Welcome to Issue 17, where Billion-dollar orbits, leaner space AI
Last week China opened a dedicated space-computing hub. This week the orbital-compute build-out got billion-dollar figures and a delivery date, while the AI side leaned toward models that need less compute and tighter human oversight.
On the Space4AI side, Elon Musk confirmed a space-optimized Nvidia Vera Rubin NVL72 rack for Starmind, aimed at Q4 2027. SpaceX's CFO then put a third terrestrial AI hosting contract at about $1.11 billion a month and said orbital compute satellites start flying in 2027. Loft Orbital and Marlan Space announced a $1 billion Altair constellation with Mistral running inference onboard, and Eclipse Space unveiled SurgeSat, an in-orbit compute platform drawing up to 100 kW.
On the AI4Space side, NASA and IBM released an open-source lunar foundation model trained on Lunar Reconnaissance Orbiter data. A startup pitched "exact AI" (automated reasoning that claims mathematically exact models without high compute) and ST Engineering iDirect added agentic governance so satellite operators can adopt AI without giving up human approval.
Specifics below.
Did you know?
We have recently launched the Bits & Orbits Insights
Check out the growing set of independent dashboards, trackers, and maps on the
AI × Space frontier:
Space4AI
Elon Musk confirms SpaceX–Nvidia space-optimized Vera Rubin NVL72 orbital data center targeting Q4 2027
On September 7, 2026, CNBC reported that Elon Musk has confirmed SpaceX is building a space-optimized version of Nvidia's Vera Rubin NVL72 rack-scale AI platform, with a first orbital launch targeted for the fourth quarter of 2027. The system is being developed jointly with Nvidia as the computing core of SpaceX's first-generation Starmind AI satellites in low Earth orbit.
The confirmation followed an August 24 announcement and a Musk post on X stating the partners had "designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028." We covered the original Starmind AI1 partnership in Issue 12. On the ground, the Vera Rubin NVL72 pairs 72 Rubin GPUs with 36 Vera CPUs plus ConnectX-9 networking; the orbital variant adapts that architecture for vacuum operation. In place of chillers, cooling towers and fans, SpaceX says the satellite would rely on closed-loop liquid cooling and large deployable radiators that shed heat directly to space as infrared radiation, cutting cooling overhead by roughly an order of magnitude. SpaceX COO Gwynne Shotwell said the company has already signed early compute-capacity deals with Anthropic and Google.
Industry analysts cited by CNBC, including Neuberger and Orbital Gateway Consulting, framed true orbital data-center scale as a "next-decade event." They identified four hurdles: thermal management, radiation hardening, GPU obsolescence, and laser-link throughput between satellites. Radiation is the sharpest of these, since making commercial Rubin silicon tolerate the cumulative dose of a multiyear LEO mission requires system-level redesign rather than a shielded enclosure alone.
SpaceX discloses third AI compute hosting contract (~$1.11B/month) and orbital compute plans at Goldman Sachs conference
On September 10, 2026, SpaceX CFO Bret Johnsen told investors at the Goldman Sachs Communacopia and Technology Conference that the company had closed another terrestrial AI compute hosting deal earlier that month. The contract is worth about $1.11 billion a month starting December 1, roughly $13 billion in annualized recurring revenue, sharpening conviction in a $100 billion ARR target by year-end.
Johnsen did not name the customer. He described the deal's structure as matching other recent compute contracts: roughly six-month commitments with early outs so SpaceX can reclaim capacity for its own models as product cycles accelerate. He tied the hosting pipeline to SpaceX's broader AI stack following the Cursor acquisition and a faster Grok product cadence, arguing the company does not want long-term leases that starve internal training and inference.
On capacity, Johnsen put year-end terrestrial compute at a little over 2 gigawatts and sketched a 5-to-10 gigawatt deployment range for next year. He said orbital compute satellites are targeted to start flying in 2027 and scale into 2028 on the same Starlink V3 bus family. SpaceX first disclosed those satellites in the S-1 we covered in Issue 1. The update came in the same session where he said Starship Flight 14 later in September would be a revenue flight carrying production Starlink V3 satellites.
Loft Orbital and Marlan Space launch $1B Altair AI satellite constellation with Mistral onboard inference
On September 9, 2026, Loft Orbital and Marlan Space announced a $1 billion program at the International Space Summit in Paris to expand their Altair AI satellite constellation from 10 to 50 spacecraft. French President Emmanuel Macron announced the plan at the summit.
Mistral supplies the onboard language-model inference, with its models intended to run directly aboard the satellites, while BlackSky provides high-resolution optical satellites and Orbitworks manufactures the spacecraft in Abu Dhabi. The satellites carry optical, radar and other sensors. Rather than sending raw imagery to Earth for processing, which takes hours, the satellites analyze what they observe in orbit and transmit event alerts within seconds. First applications include maritime monitoring, wildfire detection, disaster response, and security for ports and critical infrastructure.
Marlan Space is a UAE-based investor and a subsidiary of Abu Dhabi holding company IHC, and Orbitworks is the joint venture behind the constellation. Dr. Hamdullah Mohib, chief executive of Marlan Space and Orbitworks, said the project would turn the Abu Dhabi factory into a national capability. Customers access the models through an in-orbit AI application store that lets them select models from multiple providers, serving government and commercial users in France, Europe, the UAE and elsewhere. The first launch is planned for October 2026.
Eclipse Space unveils satellite platforms including an in-orbit compute variant
On September 9, 2026, Eclipse Space unveiled its first three satellite platforms, each targeting a different in-space capability: direct-to-device comms, high-throughput broadband, and in-orbit compute. The compute-focused platform, SurgeSat, is the company's bet on orbital data processing.
Eclipse, founded by former SpaceX engineers who worked on early Starlink, emerged from stealth three months earlier. Its first demo mission has since slipped from 2027 to 2028, according to CEO Derek Huerta. Rather than vertically integrating like Starlink, the company is building a turnkey constellation offering through supplier partnerships, an approach Huerta likened to how Apple designs the iPhone while Foxconn handles manufacturing. On the same day, Eclipse announced a partnership with French firm Gama to develop a spacecraft power system that can scale beyond 100 kW.
All three designs use a common flat-satellite platform built to fold into a rocket fairing and maximize satellites per launch. SurgeSat, the in-orbit compute variant, carries dual solar arrays generating up to 100 kW of peak power and an additional 400-square-meter radiator to dissipate heat on orbit. By comparison, the CitraSat D2D platform peaks at 6 kW and the SliceSat broadband platform at 8 kW.
Eclipse plans to bring the platforms to market sequentially, starting with the D2D market and applying those lessons to broadband and in-orbit compute. Huerta said the company is prioritizing commercial and sovereign constellations over US government work, citing the overhead of government contracting as a competitive drag.
AI4Space
NASA and IBM release open-source Lunar Foundation Model trained on LRO data
On September 10, 2026, NASA and IBM Research released the NASA-IBM Lunar Foundation Model, among the first open-source AI models built specifically for lunar science. The model is trained primarily on data from NASA's Lunar Reconnaissance Orbiter (LRO), with weights hosted on Hugging Face and the complete codebase published on GitHub for testing and experimentation.
The model was trained on roughly 2 million image tiles drawn from 17 years of LRO observations, comprising more than 1 million high-resolution camera images at 1-meter resolution and nearly 964,000 multispectral images at 100-meter resolution. NASA notes that LRO's dataset is larger than all other NASA planetary missions combined and captures a near-seamless high-resolution mosaic of the entire Moon. Training also drew on imagery and terrain data from NASA's GRAIL and Lunar Prospector missions and JAXA's Selenological and Engineering Explorer. Because the model is pre-trained on unlabeled data, planetary scientists can fine-tune it to specific tasks using only small amounts of labeled data.
Fine-tuned versions map craters, locate irregular mare patches (young-looking volcanic features that challenge established lunar cooling timelines), and estimate where ice is likely to remain stable on and below the surface near the poles. NASA reports the model matched or exceeded several strong baseline models across all evaluated tasks, reducing error 22–23% over those baselines and showing a clear advantage on polar ice stability estimation. In one test near Einstein crater, the model detected an impact crater newly formed by a SpaceX rocket body in imagery excluded from pre-training, demonstrating fine-tuning to recognize surface changes between observations.
The model was built by the Impact AI team at NASA's Marshall Space Flight Center with the Science Mission Directorate's Planetary Science Division, Goddard Space Flight Center, and Ames Research Center, and sits within the Office of the Chief Science Data Officer's AI-for-science strategy. Kevin Murphy, NASA's chief science data officer and acting chief data and AI officer, said the effort shows "what's possible when we bring AI to NASA's petabytes of scientific data." It joins earlier NASA-IBM releases, including the Prithvi Earth-observation models and the Surya heliophysics model, and the team released machine-learning-ready datasets, benchmarks, and a companion technical report alongside the model, integrated into the open-source TerraTorch toolkit.
PiLogic pitches 'exact AI' automated-reasoning approach for high-stakes space missions
On September 9, 2026, PiLogic CEO Johannes Waldstein appeared on Via Satellite's On Orbit podcast to introduce the startup's "exact AI" approach to space systems. PiLogic is developing what it calls "exact AI," a type of AI in the category of automated reasoning that differs from LLMs in that it is based on mathematically exact models that do not require high levels of compute.
Waldstein described applying the technique to spacecraft operations rather than to language tasks. He framed exact AI for use in anomaly detection, diagnostics, autonomy, and sensor fusion for space systems. The pitch turns on the low-compute claim, which matters for onboard hardware where power and processing budgets are constrained.
The company's founders come from outside the space sector. Waldstein has a tech background and has founded five companies, notably Fan.AI, a data platform focused on sports sponsorship. He met co-founder Mark Chavira, a former Google executive, through their son's Scouts program, and shared interests in AI led to the aerospace startup.
ST Engineering iDirect adds agentic governance and controls to Foresight AI orchestration platform
On September 10, 2026, ST Engineering iDirect announced new capabilities for Foresight, its AI-driven orchestration and intelligence platform for satellite operators managing multi-orbit and multi-service networks. The release adds AI governance, operational controls, and guided onboarding tools aimed at letting operators adopt AI while retaining oversight of operational changes.
The update centers on four additions. Agentic Governance provides oversight, auditability, and risk controls for AI agents through governance policies, audit trails, and operational safeguards. Governed Change Management adds approval-based workflows with review, risk assessment, and authorization before AI-driven operational changes are applied. AI Usage Visibility and Cost Controls provides monitoring, budgeting, and policy-based controls for managing AI consumption, while Accelerated Onboarding automatically discovers network inventory, deploys required platform components, and guides administrators through privacy and deployment configuration.
Foresight was first introduced in March 2026, and supports air-gapped deployments for operators requiring isolated networks and data governance. "Satellite operators are under increasing pressure to deliver services faster, manage growing operational complexity and adopt AI responsibly," said Sridhar Kuppanna, CEO of ST Engineering iDirect. The company frames the platform as a path from reactive network management toward autonomous operations.
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: