Starcloud, Axiom Space, Lonestar, and others are betting they can stake claims in orbit before Google and SpaceX scale up
Manuel Mazzanti / NurPhoto via Getty Images
A startup nobody had heard of 17 months ago is now valued at $1.1 billion and has already put a GPU in orbit. Starcloud, a Redmond, Washington company that went through Y Combinator's summer 2024 cohort, raised $170 million in a Series A in March 2026 and has filed with the FCC for an 88,000-satellite constellation. In November 2025, it became the first company to run an Nvidia $NVDA H100 GPU in orbit.
The ambition is clear. So is the competition. Google $GOOGL plans prototype satellites by 2027. SpaceX has filed for up to a million. Jeff Bezos has sketched out gigawatt-scale orbital facilities for Blue Origin. A handful of startups are racing to establish positions in orbit before the largest companies in tech arrive in force.
Their strategies differ sharply — and those differences reveal what each company believes the market actually is right now, before launch costs fall far enough to make AI data centers in space viable at scale.
Starcloud has the strongest claim to operational hardware. Its 60-kilogram Starcloud-1 satellite, launched aboard a SpaceX Falcon 9 on Nov. 2, 2025, carried an Nvidia H100 GPU and ran multiple AI workloads in orbit — including training a small language model on the complete works of Shakespeare, according to Introl's February 2026 analysis of the orbital data center market. The company, formerly Lumen Orbit before a rebrand in early 2025 to avoid a trademark conflict, is planning its next satellite for October 2026, carrying several H100 chips alongside Nvidia's newer Blackwell hardware.
CEO Philip Johnston described the company's business model as closer to infrastructure than cloud computing. "We'll offer a box that has power, cooling, and connectivity, and then people can do whatever they want with that box," Johnston said in a March 2026 interview. "We're really like an infrastructure and energy company rather than a cloud provider." Its initial customers are military and government satellite operators needing computing power in orbit. The longer-term plan, as Johnston described in a Sequoia Capital podcast, is to sell capacity to major cloud providers once launch costs fall far enough.
Heat management is one of the hardest engineering problems Starcloud has to solve. On the ground, data centers use air or water to cool their chips. In orbit, neither is available, but heat escapes efficiently by radiating off flat panels into the cold vacuum of space. Starcloud estimates that a one-square-meter panel can shed about 838 watts, so the cooling system needs to cover less than half the area of the solar arrays. That keeps the total structure compact enough to be practical. For a five-gigawatt facility, which Starcloud is targeting, solar and cooling panels would only have to stretch about four kilometers.
The other startups have each found a different angle. Axiom Space, the Houston-based company building commercial space station modules, deployed two orbital computing nodes in January and is targeting a fully connected orbital computing node aboard the International Space Station by 2027. Its partner ecosystem includes Kepler Communications, which launched 10 optical relay satellites in January, each carrying multi-GPU computing modules and terabytes of storage. The Kepler constellation operates as a mesh network, providing the connectivity backbone that Axiom and others need to link orbital computing nodes.
Lonestar Data Holdings is focused on lunar and near-lunar data storage, selling the idea that a backup stored off-planet can survive anything that happens on Earth. In February 2025, Lonestar successfully tested its data storage hardware aboard an Intuitive Machines lunar lander, becoming the first company to operate a commercial data center on its way to the Moon. It has since signed a $120 million agreement with Sidus Space to build six data storage satellites. The first will carry 15 petabytes of capacity and orbit a stable point between Earth and the Moon, with launches planned between 2027 and 2030.
OrbitsEdge and Aetherflux round out the field with narrower bets. OrbitsEdge, based in Cocoa Beach, Fla., sells hardened computing hardware designed to withstand the radiation and extreme temperatures of orbit, targeting satellite operators who need to process data in space rather than transmit it raw to the ground. Its first orbital demonstration is planned for 2026, according to Introl. Aetherflux, founded by Robinhood co-founder Baiju Bhatt and backed by Index Ventures, Andreessen Horowitz, and Breakthrough Energy Ventures, is combining orbital computing with power-beaming technology that transmits energy to Earth via infrared laser.
The large tech companies set a formidable benchmark. Google's Project Suncatcher envisions solar-powered satellite constellations equipped with its own TPU chips and connected by optical links that can transmit 800 gigabits per second between satellites — roughly 100 times faster than a typical home fiber connection. Two prototype satellites, built in partnership with Planet Labs, are slated for early 2027.
The economics remain the binding constraint. Google's researchers estimate that orbital computing becomes cost-competitive with ground-based data centers at a launch price of about $200 per kilogram. SpaceX's Falcon 9 rocket currently runs about $2,700 per kilogram. Analysts at Citigroup $C project costs could reach about $100 per kilogram by 2040, with an optimistic case near $33.
The question for all of them is whether they can build enough of a business at current costs to survive until launch prices fall. Johnston told the Sequoia podcast that Starcloud's all-in energy cost at scale will be far cheaper than anything available on the ground today, but only once Starship — SpaceX's next-generation rocket, which is still ramping toward commercial operations — brings launch costs down dramatically. Bezos offered a more cautious read, estimating the full transition from ground-based to space-based data processing would take more than 10 years.
For the startups, the bet is that being first to orbit — even at small scale — buys the customer relationships, operational data, and regulatory filings that will matter when that moment arrives.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
