Orbital AI Compute Goes Operational — SpaceX Starweave and the Race for Space-Based Data Centers

SpaceX says orbital compute will be cheaper than terrestrial data centers within two years. The math says otherwise — but the deals are already signed.

In June 2026, SpaceX launched the first 120 satellites of Project Starweave, the company's AI-optimized Starlink constellation. Each satellite carries edge computing hardware capable of running inference for AI models in orbit, reducing latency from hundreds of milliseconds to single digits. A month earlier, the company had unveiled the AI1 satellite design, a 150 kW orbital compute platform described as "not a difficult engineering challenge" by Elon Musk himself.

The world's largest IPO — SpaceX filed for a $1.75 trillion listing in April 2026 — depends on an infrastructure category that does not yet exist. As we wrote in July, the race for space-based data centers is already reshaping how institutional capital thinks about compute infrastructure. The difference now is that the customers have signed contracts.

The Starweave Constellation: What Actually Launched

Project Starweave is not a single satellite type but a layered architecture. The first 120 satellites launched in June 2026 carry edge inference hardware: NVIDIA Jetson-class coprocessors capable of running compact AI models directly on orbit. These are not data centers in space; they are smart routers that can process and prioritize data before downlinking it.

The AI1 satellite, unveiled on June 8 during SpaceX's IPO roadshow, is the second layer. Each AI1 unit delivers 150 kW peak compute payload power (120 kW average) from a 70-meter wingspan of solar arrays, with deployable liquid radiators spanning 110 m². The compute payload is chip-agnostic; SpaceX designed the satellite to accept modules from any vendor, a pragmatic hedge given the company's well-documented difficulty securing AI silicon supply.

The third layer, Starmind, is the long-range vision: a megaconstellation of AI satellites manufactured at SpaceX's Gigasat factory in Bastrop, Texas, a 1,000-acre site targeting production output by late 2027. The FCC filing from January 2026 requests authorization for up to one million satellites. To put that number in context: there have been roughly 7,000 orbital launches in all of human history, and there are approximately 14,500 active satellites in orbit today. Scaling to one million requires not just a factory but a manufacturing revolution.

The Physics Problem That Won't Go Away

The central tension in orbital AI compute is not capital or regulation. It is thermodynamics. Every watt of compute generates an equivalent watt of heat, and in the vacuum of space, the only way to shed that heat is radiative cooling. IEEE Spectrum's July 2026 cover story calculated that cooling a single NVIDIA H100 GPU (700W) requires 1.4 square meters of radiator surface at 60°C. Scaling that to a 150 kW satellite implies roughly 300 m² of radiator, the area of a tennis court, deployed in orbit.

Starcloud, the Washington-based startup that launched the first orbital AI data center in November 2025, confirmed the scale of the challenge: its H100 payload reached orbit and ran inference workloads, but the radiator was undersized to let the chip run at full power for sustained periods. The company has since raised $170 million and is preparing Starcloud-2 with 100x the power generation, integrating NVIDIA's Blackwell platform.

SpaceX's AI1 design addresses this with deployable liquid radiators and a 110 m² thermal management surface, but the unit has not yet flown. Two prototype satellites are scheduled for early 2027, with commercial deployment following no earlier than 2028. Until then, every performance claim is a paper specification, not a measured result.

The Competitive Landscape: Six Players, One Unproven Category

SpaceX is the most visible entrant, but it may not be the first to deliver a commercially operational orbital compute node.

Kepler Communications commissioned the first distributed on-orbit computing fabric in March 2026, deploying 40 NVIDIA Jetson Orin modules across 10 interconnected satellites in its Tranche 1 optical data relay constellation. Kepler's approach is distinctive: rather than building purpose-designed compute satellites, it adds AI payloads to an existing optical communications network, creating a space-based edge compute fabric where workloads can scale dynamically across nodes. The architecture supports both single-node execution and clustered distributed computing. If any node fails, workloads shift to others automatically. This is the only commercially operational orbital compute network in existence today.

Orbital (the startup, not the industry category) raised a $5 million pre-seed round from a16z speedrun in June 2026 to build purpose-designed AI inference satellites. Its Pathfinder mission is slated for 2027, with a long-term vision of 100,000+ satellites delivering 10 GW of orbital compute. The company's differentiation is to match the solution to a specific problem: small satellites running inference workloads at lower launch cost than the megaconstellation approach.

Starcloud achieved unicorn status in March 2026 with a $170 million raise, following its November 2025 H100 demonstration. The company has filed for an 88,000-satellite constellation. Blue Origin entered the race in March 2026 with Project Sunrise (51,600 satellites), targeting government clients. And Google's Project Suncatcher, a test constellation of 80 data-crunching satellites, is targeting demo missions in 2027.

What This Means for Compute Infrastructure Investors

The orbital AI compute market is projected to grow from approximately $1.77 billion in 2029 to $39.1 billion by 2035, a 22x increase in six years, per market forecasts cited in the SpaceX FCC filing. But these projections depend on a chain of assumptions that is unusually fragile: Starship reaching full reusability and launch cadence, satellite manufacturing scaling by two orders of magnitude, and the thermal engineering of multi-kilowatt orbital compute nodes being solved for sustained operation.

What is real today: Kepler's operational compute network, SpaceX's $76 billion in signed compute contracts, and a regulatory race at the FCC that has already drawn filings for over 1.1 million orbital data center satellites across multiple applicants. The FCC's orbital debris mitigation rules, updated in 2024, impose a 25-year deorbit limit, meaning every satellite in these constellations has a built-in replacement cycle that drives recurring manufacturing and launch revenue.

What remains unproven: whether any of these systems can operate at full commercial power, whether the economics pencil out at scale, and whether the orbital debris and reentry concerns — astronomers have calculated that a megaconstellation this size would produce a satellite reentry approximately every three minutes — trigger regulatory intervention before the first commercial node goes live.

Development scenarios

Sources