Published: June 03, 2026, 15:40 UTC
Muon Space Unveils Condor-Ultra — But Can Orbital Data Centers Actually Work?
MOUNTAIN VIEW, Calif. — Muon Space today unveiled the Condor-Ultra, a Starship-class satellite platform purpose-built for what CEO Jonny Dyer calls “the most compelling space infrastructure missions of the next decade.” At the top of that list: orbital data centers, a concept generating enormous hype — and enormous skepticism.
The numbers are arresting. Condor-Ultra delivers 20 kW baseline power, scalable to 100 kW, with over 18 square meters of nadir-facing payload area. It integrates Starlink Mini Lasers for 25 Gbps inter-satellite networking and NVIDIA’s Space-1 Vera Rubin Module — a radiation-hardened AI inferencing unit offering 25× the performance of an H100 GPU. The platform is designed for native stackability inside Starship’s massive fairing, enabling deployment of “thousands of satellites” in a single launch campaign.
Founded in 2021, Muon Space has moved fast. It achieved 95% in-house vertical integration — including propulsion, after acquiring Starlight Engines — and recently opened a San Jose facility with 10× expansion capacity capable of producing 500 satellites per year. The company’s first XL-class satellite launches in 2027 for Hubble Network, with a Condor-Ultra pathfinder slated for 2028. Configurations already exist for Falcon 9, Rocket Lab’s Neutron, and SpaceX’s Starship.
President Greg Smirin drew a sharp line between the company’s existing XL platform and the new Ultra: “Condor-Ultra is a different magnitude entirely.”
The Orbital Compute Gold Rush
Condor-Ultra arrives as the orbital data center narrative accelerates dramatically. AI compute demand is doubling every few months. Terrestrial data center construction faces mounting opposition — a World Economic Forum report published yesterday found that 70% of Americans oppose local AI data center builds, citing power consumption, water usage, and noise. Space offers an elegant escape hatch: unlimited solar energy, no zoning boards, no grid constraints, no water cooling bills.
The market signals are unmistakable. SpaceX has filed for a constellation of up to one million satellites for orbital data center operations. Starcloud, a separate venture, plans 88,000 satellites. Google and SpaceX are in preliminary partnership discussions. US data center spending jumped 70% between May 2023 and May 2024, and the Lawrence Berkeley National Laboratory projects energy consumption could double or triple by 2028.
Ramon.Space CEO Avi Shabtai calls orbital compute “the opportunity of the century,” though he readily admits space is “a hostile environment” plagued by radiation, thermal extremes, and punishing transport logistics.
The Physics Problem No Amount of GPUs Can Solve
Amid all this fervor, a fundamental question goes largely unasked: can orbital data centers actually work at meaningful scale?
The WEF’s June 2 report points to the central obstacle — cooling. In the vacuum of space, waste heat has exactly one escape route: infrared radiation. There is no convection, no conduction, no water loop to carry heat away. A two-sided radiator operating at a relatively warm 20°C emits just 633 watts per square meter. Compare that to a terrestrial data center’s water-cooled system, which rejects heat over 1,000 times faster. The math is brutal: a 1 MW orbital data center — roughly 1,000 times smaller than a hyperscale terrestrial facility — would require 1,600 square meters of radiator surface. That’s the size of an Olympic hockey rink, devoted entirely to dumping heat.
The geometry compounds the problem. Solar arrays must face the Sun to generate power. Radiators must face cold space to reject heat. These requirements directly conflict. Every square meter of radiator is a square meter that cannot be solar panel. Every watt of compute generates waste heat that must be radiated away from a surface area that competes with the very power generation it depends on.
NASA’s experience is instructive. The International Space Station’s External Active Thermal Control System rejects roughly 70 kW through radiator wings weighing around seven metric tons. Scaling that architecture to 1 MW would require approximately 100 tons of radiators alone — before a single server is launched.
The Economics of Launch
The Thales Alenia Space-led Ascend study, funded by the European Union, proposes a 10 MW orbital data center comprising 13 satellites in a 200-by-83-meter formation, housing roughly 5,000 servers. To reach just 200 MW — still an order of magnitude below a mid-tier terrestrial facility — the study concludes it would require 200 large-scale orbital infrastructures and 200 dedicated launches. The environmental calculus is no kinder: the rockets needed would have to be ten times less emissive than current models just to break even on carbon footprint versus terrestrial alternatives.
Dr. Domenico Vicinanza of Anglia Ruskin University puts it bluntly: “Launching hardware remains extremely expensive. Each kilogram costs thousands of dollars.”
SignalFeed’s analysis of the Google-SpaceX discussions echoes the caution: “The economic equation heavily favors terrestrial data centers.” Beyond launch costs, the obstacles include prohibitive maintenance and repair costs for hardware in orbit, specialized power systems with no terrestrial supply chain, and a regulatory framework that barely exists.
What Actually Makes Sense
None of this means Condor-Ultra is a solution in search of a problem. Muon Space’s platform is genuinely impressive engineering — the power budget, the AI compute density, the inter-satellite networking, the vertical integration, the production scalability. For Earth observation, communications infrastructure, and edge computing applications that genuinely benefit from being in orbit, Condor-Ultra represents a meaningful leap forward.
Where the skepticism is warranted is the vision of hyperscale AI training in space. Training a frontier model like GPT-5 or Gemini Ultra requires gigawatt-hours of energy and generates corresponding waste heat measured in megawatts. The radiator area, launch mass, and cost equations simply do not close at that scale. Not with Starship. Not with any rocket currently on any drawing board.
What does close is the edge inference use case. Latency-sensitive AI inference — where a model needs to run on-orbit and return results to Earth in real time — is a genuine, underserved market. Remote sensing data processing, satellite constellation management, secure government compute, and disaster response all benefit from having NVIDIA-class AI inferencing in space without beaming terabytes of raw data to the ground. Condor-Ultra’s 25× H100-class capability, paired with 25 Gbps laser links, makes this viable today.
The balanced verdict: Muon Space has built an exceptional satellite platform that will power real missions. But the vision of orbital data centers displacing terrestrial hyperscale facilities runs headlong into immutable physics. Cooling in a vacuum is fundamentally, exhaustively harder than cooling in a data center with a water pipe. Launch costs, though falling, remain punishing at scale. And the regulatory and maintenance frameworks for orbital server farms barely exist.
Orbital AI inference at the edge: viable, valuable, coming soon. Orbital AI training at hyperscale: a vision that will require breakthroughs in thermal management, launch economics, and on-orbit servicing that are nowhere on the near horizon.
Condor-Ultra is real hardware. But physics, as always, gets the final vote.