The most compute-intensive industry on Earth is moving to the most hostile environment for electronics. That is the paradox driving the orbital data center race — and it might be the only way to keep AI scaling alive.
Nvidia and a16z-backed startups are racing to put data centers in low Earth orbit, where solar power is continuous and heat dissipates into vacuum.
The economics remain unproven: launching compute into orbit costs more per watt than building on the ground. But as terrestrial power constraints tighten, the gap is narrowing faster than most analysts expected.
In January 2026, SpaceX filed with the regulators for a constellation of up to one million satellites to function as an orbital data center. Six months later, the company unveiled AI1 — a satellite with 150kW peak compute, a 70-meter wingspan, and interchangeable compute modules. The timing was not coincidental. AI1 was shown to investors days before the the IPO, the largest in history at $75 billion.
But SpaceX is not alone. Project Suncatcher was published by The search giant, a research paper describing compact constellations of solar-powered satellites carrying custom TPUs. The Vera Rubin Space Module from Nvidia delivers 25x more AI compute for space-based inferencing. Starcloud raised $170 million at a $1.1 billion valuation for its orbital data center network. A startup out of a16z's Speedrun accelerator, closed a $5 million pre-seed round to launch its Pathfinder mission in 2027.
You are looking at the early formation of a new infrastructure category.
The Terrestrial Ceiling
The premise behind orbital data centers is simple in outline and brutal in implication. The International Energy Agency projects global data center electricity consumption will double to 945 terawatt-hours by 2030 — roughly Japan's entire annual consumption. In the US, grid operators are struggling to connect new data centers fast enough to meet demand. Cooling requires billions of gallons of water. Permitting timelines stretch years. Land near transmission corridors is priced for a different economy.
In orbit, none of these constraints apply. Solar panels in low Earth orbit receive 1,361 W/m² continuously — more than five times the average terrestrial solar density. There is no weather, no night in sun-synchronous orbits, and no atmospheric attenuation. Heat rejection uses passive radiative cooling into vacuum at near-zero operating cost. The fuel — sunlight — is abundant and free.
The question is whether the cost of getting there and operating there can ever beat building on solid ground.
SpaceX AI1: The First Draft
The company disclosed detailed specifications during its June IPO roadshow. The AI1 satellite delivers 120kW of sustained compute, with 150kW peak. It deploys to a 70-meter wingspan with a 20-meter height. The thermal management system uses a 110m² deployable liquid radiator with redundant pumping loops. Solar arrays are built in-house at its Bastrop, Texas plant, reusing Starlink V3 technology.
"The AI satellite is much simpler than a Starlink satellite," Elon Musk said during the presentation. "It is essentially a lot of solar cells and some laser links."
The company has signed major compute lease agreements with Anthropic ($12.5 billion per month for xAI data center capacity) and ($920 million monthly for AI capacity). First prototype launches are planned for early 2027, with volume production by late 2027 at its new Gigasat factory in Bastrop — a 1,000-acre facility with over 11 million square feet of production space.
The ambition is a constellation of one million satellites that the company claims will produce the lowest-cost AI compute on Earth within three years. AWS CEO Matt Garman called the economics "just not economical." Morningstar pegged SpaceX's fair value at $780 billion — less than half the IPO valuation.
The Field Beyond the company
The company dominates the narrative, but the competitive picture is broader than one company.
Project Suncatcher, published as a preprint in June 2026, describes a system of networked satellites in dawn-dusk sun-synchronous orbit carrying custom TPUs connected by free-space optical links. The company plans two prototype launches by early 2027 in partnership with Planet Labs, using Planet's Owl satellite constellation as a testbed. The search giant is approaching orbital compute not as a side project but as a long-term infrastructure hedge against terrestrial energy constraints.
The Vera Rubin Space Module launched at GTC 2026 in March, delivering 25x more AI compute per watt than the H100 for space-based inferencing. The module is designed for size-, weight-, and power-constrained orbital environments. Partners include Axiom Space, Starcloud, Sophia Space, and Planet Labs. The move standardizes the compute layer for the orbital data center market before the market exists.
Founded in 2024, raised $170 million in Series A funding in March 2026 at a $1.1 billion valuation — the fastest Y Combinator graduate to reach unicorn status. It launched its first satellite in November 2025 and successfully trained AI models in orbit using an H100 GPU. The company has filed with the FCC for a constellation of up to 88,000 satellites and partnered with Crusoe to launch the first public cloud in space by 2027.
Euwyn Poon (founder of Spin, sold to Ford for $100 million), took a different approach. Instead of large satellites, it is building a distributed network of small, independently deployable compute nodes. Its $5 million pre-seed round from a16z Speedrun funds a Pathfinder mission in 2027 to validate GPU hardware in orbit. "Thinking about building a football field size satellite from a maintenance and physics standpoint seems like a pipe dream," Poon told Payload.
AI1 prototype launch (target: early 2027) — the first real test of orbital compute economics
Planet Labs orbital compute demonstration (target: early 2027)
Its Starcloud-2 launch with commercial edge and cloud workloads
Licensing decisions on million-satellite constellations — sets precedent for the entire sector
Insurance market for orbital data centers — Marsh and Lloyd's have been approached by multiple companies
The Hard Parts
Every company in this space acknowledges the same set of unsolved problems. Launch costs remain the dominant variable. At current Falcon 9 pricing of roughly $2,700/kg to LEO, putting a 150kW data center into orbit costs tens of millions before the hardware is even powered on. Starship, projected to reduce launch costs to under $200/kg, is not yet fully operational. The math does not work until Starship flies at scale.
Radiation is a second-order problem that becomes first-order at scale. Commercial-grade GPUs and TPUs were not designed for the radiation environment of low Earth orbit. Single-event upsets — bit flips caused by high-energy particles — are manageable for today's handful of experimental payloads. For one million satellites running AI inference workloads, the error rate becomes a statistical certainty that must be engineered around.
Thermal management in vacuum is counterintuitive. On Earth, fans and liquid cooling move heat away from chips. In space, convection does not exist. All cooling is radiative — heat must be shed through dedicated radiator panels facing deep space. AI1's design allocates 110m² of radiator surface, but scaling radiative cooling to megawatt-class orbital clusters is untested.
Orbital debris is the existential risk. A million-satellite constellation, even with five-year lifetimes and controlled deorbiting, creates a debris environment that affects every other space asset. Regulators are already reviewing these applications. The outcome will set precedent for the entire orbital computing sector.
Insurance remains an open question. Reuters reported in June 2026 that Marsh and Lloyd's have held preliminary discussions with Starcloud, Lonestar, and Orbital about orbital data center coverage. "The conversations in the market are focused on whether the risk can be modeled, rather than what the premium should be," said Kasey Roh, US head of Upstage AI.
The incumbent response
Several large energy companies are developing dedicated data center power plants, including natural gas peakers and small modular reactors (SMRs). The first SMR-powered data center is expected online in 2029 at the earliest — roughly the same timeline as commercial orbital compute services.
What happens to the AI infrastructure market when space becomes an option?
Probability: 65% — The convergence of Starship launch economics, Space-grade compute modules, and a growing pipeline of venture capital creates a credible path to cost parity for latency-tolerant inference workloads within three years. Training, which requires tight interconnects and real-time human oversight, will remain terrestrial for the foreseeable future.
✅ Arguments for
+ Starlink's manufacturing line is directly transferable to AI satellite production, compressing the learning curve.
+ Terrestrial power constraints are accelerating, not slowing — every year of grid permitting delays strengthens the orbital compute value proposition.
Confirmation criteria: Starship achieves full reusability with launch cost below $500/kg and a prototype AI1 satellite successfully runs inference workloads for 12 months in orbit.
❌ Arguments against
− Orbital debris regulation could cap constellation size far below one million satellites, limiting total addressable compute capacity.
− AWS's Matt Garman publicly dismissed orbital compute economics — if hyperscalers do not buy, the market stays small.
Disconfirmation criteria: A major hyperscaler (AWS, Microsoft, The search giant) announces it will not procure orbital compute services, or the regulators deny the company's constellation application on debris grounds.
Development scenarios
🟢 Optimistic scenario (30%)
Implications: Terrestrial hyperscalers shift a portion of inference capacity to orbit. Data center REITs face structural headwinds from a new compute asset class.
🟡 Base-case scenario (45%)
Implications: the startup compute becomes a specialized vertical rather than a general infrastructure category. The energy constraint on AI scaling continues to be addressed primarily through terrestrial solutions (SMRs, grid upgrades, efficiency gains).
🔴 Pessimistic scenario (25%)
Implications: The AI infrastructure bottleneck stays on Earth. Nuclear and grid investments accelerate as the only available scaling path.