Welcome to Issue 14, where two $250M checks land on in-space compute — and one honest case against it
Last week nothing notable landed on the Space4AI side, so we doubled down on AI4Space. This week the capital showed up.
On Space4AI, Starcloud and Muon each closed a $250 million round for in-space compute — one building orbital data centers, the other a constellation foundry with on-orbit AI — and Space Capital published the candid case that the physics works while the economics do not yet.
On AI4Space, spacecraft autonomy kept leaving the lab. EraDrive, using its Era-Core flew optical navigation that does not depend on any GNSS system, and Slingshot's operations platform grew 400 percent as it added AI collision avoidance.
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
Starcloud and Muon each raise $250M for in-space compute
The same number hit twice this week, on two different bets.
On August 21, Starcloud closed a $250 million Series A extension at a $2.3 billion post-money valuation, taking total capital to $450 million. Manhattan West led, with NVIDIA and Cisco Investments joining as new investors. The money funds manufacturing, NVIDIA engineering, and launch allocation: Starcloud-3 production is underway at a new 100,000-square-foot plant in Woodinville, toward a stated 88,000-satellite, 20-gigawatt orbital-compute constellation. The NVIDIA work started when Starcloud-1 flew the first H100 to orbit in November 2025; the companies are now building NVIDIA's Space-1 Vera Rubin Module. Since launch, Starcloud says it trained the first AI model in space, ran Gemini in orbit, and demonstrated inference and fine-tuning on flight hardware.
A day earlier, Muon Space closed a $250 million Series C led by Eclipse Capital, with Google and Salesforce Ventures among the participants, taking total equity past $386 million. Muon is a constellation foundry, not a second orbital data-center vehicle. Proceeds go to production and dual-use platforms, and — the on-thesis line — on-orbit AI compute plus Starlink connectivity. Issue 3 covered Muon's Condor-Ultra Starship-class bus and planned NVIDIA Space-1 integration; this round finances scaling that platform. Seven satellites in the first half of 2026 brought the total to 11. FireSat (Earth Fire Alliance / Google.org) and Vindlér 2.0 for SNC come into operation this year; a San Jose factory is sized for 500 satellites a year by 2027. Muon declined to disclose a valuation.
Space Capital publishes 'Orbital Data Centers: The Honest Case' analysis
On August 20, Space Capital published "Orbital Data Centers: The Honest Case," an edited conversation between Muon Space CEO Jonny Dyer and Space Capital partner Tom Ingersoll recorded at the 2026 Space Capital Summit at Nasdaq MarketSite. Dyer's framing: the physics of orbital compute works, the economics do not yet, and nearly everything in between is an engineering problem rather than a physics one.
Dyer credits launch for making the conversation possible at all, and he traces the cost collapse from experience. He recalled buying converted Russian ICBMs to launch Skybox satellites fifteen years ago, versus today buying SpaceX slots "almost like a plane ticket," with Starship expected to reset the economics again. The second shift is manufacturing: he noted that Starlink satellites now resemble a data-center rack more than a traditional spacecraft, and that cheaper launch and commodity-style production feed each other.
On the upside, he pointed to orbits where the sun is baseload power and never sets, removing the storage a ground array needs for night and weather. Space offers a near-absolute-zero heat sink with no water, real estate, or permitting fights, and orbit carries no county councils, making any regulatory approach global. The catch, he said, is that capturing those advantages is "not physics, it's engineering."
The honest part is supply. Dyer put a large orbital data center at roughly eight to ten times the cost of its terrestrial equivalent at current launch prices, requiring launch cost to fall about five times and cadence to rise about ten. He noted Falcon 9 flies roughly 100 to 200 times a year while the math wants 500 to 1,000, the same scale needed to put gigawatts of compute in orbit. Demand, by contrast, he called "potentially unbounded" as AI grows and the U.S. grid falls behind.
He worked through the standard objections and downgraded them. Most AI traffic stays machine-to-machine, and Starlink's inter-satellite links already run about 100 gigabits per second, an order of magnitude short of a data center's terabit fabric but "a gap, not a wall." On radiation, he argued terrestrial chips are already hardened against cosmic-ray bit flips at scale, citing that Google's fleet effectively became "the world's largest cosmic ray telescope." What he would not wave away is that space remains a hostile environment of autonomous robots that cannot be physically repaired, with integration among the hardest problems in technology.
Dyer framed the shift as a progression rather than a switch, with national-security workloads and data best kept in orbit pulling the first systems up. He noted Google alone is spending about $100 billion on data-center capex this year, of which a small fraction goes far above the atmosphere, and that every hyperscaler will have to run the experiment. As an example of the convergence, he pointed to Muon's FireSat constellation — the wildfire-monitoring system the Series C above is now bringing into operation — which he described as built with Cal Fire and backed by Google and the Environmental Defense Fund, pairing orbital sensing with on-board AI and SpaceX optical links to give crews hourly fire situational awareness day and night.
AI4Space
EraDrive’s Era-Core flying on NASA’s Starling demonstrates FALCON GPS-free optical navigation
On August 17, NASA announced a successful flight demonstration of the FALCON optical navigation system aboard its Starling CubeSat swarm, a technology that fixes a satellite's position by referencing other objects in space rather than a navigation network.
The FALCON — Fast Autonomous Lost-in-space Catalog-based Optical Navigation — payload is a joint flight experiment by NASA and EraDrive, a startup spun out from Stanford University.
The experiment combines EraDrive's Era-Core flight software and embedded algorithms with Starling's cameras and an onboard catalog of known satellites to support GPS-independent navigation and space situational awareness.
Traditional satellite navigation depends on GPS signals, which can be unreliable or unavailable in lunar or deep-space environments.
FALCON made creative use of Starling's onboard star-tracker cameras — standard instruments that identify bright objects to inform orientation and position — matching observed spacecraft and orbital debris against a catalog of known objects maintained and made publicly available by the U.S. Department of War, then using those verified objects as reference points to determine Starling's orbit.
The mission team loaded the full catalog of approximately 20,000 space objects and their predicted orbits onto the spacecraft, and FALCON correlated that data with its own observations to pin down both Starling's location and the objects' locations with greater precision than the existing catalog. During a three-day period, FALCON improved the known orbits of more than 200 objects without intervention from ground operators.
NASA describes the self-orbit determination as a first for spacecraft using optical cameras to navigate by their relative position to other objects, and the catalog-update experiments produced better object position predictions onboard than those provided by ground stations.
FALCON began as a University SmallSat Technology Partnerships project before evolving into EraDrive, which is now commercializing its Era-Core software and related hardware.
Later this year, Starling — which launched in 2023 — will extend the experiment using Era-Core, enabling its four-spacecraft swarm to share tracking data and refine their positions collectively.
Slingshot Aerospace expands Portal with AI-powered collision avoidance as user base grows 400%
On August 18, Slingshot Aerospace announced that its Portal space-operations platform has surpassed 400% user growth since its April launch, and previewed new AI-driven capabilities including collision avoidance, high-fidelity object tracking, and on-demand processing. In four months, Portal has grown to more than 630 users representing 330 commercial, civil, and national security organizations across 19 countries.
The expansion targets an increasingly crowded environment. Slingshot's Seradata database counts more than 16,600 spacecraft now active in orbit, over 90% of which entered service since 2020, with annual spacecraft deployments up nearly eightfold since 2018. The headline addition is AI-driven collision avoidance powered by TALOS, the Thinking Agent for Logical Operations and Strategy, Slingshot's AI reasoning system. TALOS lets operators move from identifying a conjunction risk to evaluating maneuver options inside Portal, providing what the company calls explainable maneuver intelligence to compare tradeoffs before choosing a course of action. Slingshot notes TALOS already supports U.S. Space Force mission training and rehearsal, bringing an operationally proven reasoning capability into Portal's workflows.
Two further features round out the release. Next-generation object tracking will let operators request high-fidelity tracking of specific objects and check whether the Slingshot Global Sensor Network already holds the data or needs to collect it, while on-demand processing via the company's MFAST data fusion engine runs through Portal Marketplace. "A lot of space operators understand, there's such an exponential increase in the amount of space objects and the amount of data that's being collected...that, without AI, we're never going to be able to process and deconflict and make decisions fast enough for the real world," said Robin Dickey, Slingshot's director of policy and government affairs, on the sidelines of the Space and Missile Defense Symposium in Huntsville, Alabama. Chief Technology Officer Erik Ekwurzel framed the features as advancing Slingshot's Sense, Fuse, Decide, Act operations cycle, accessible through Portal or via APIs and MCP into existing mission systems.
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: