In November, a 60-kilogram refrigerator-sized satellite called Starcloud-1 entered low Earth orbit aboard a SpaceX rocket carrying the first data-grade GPU ever to operate in space—an Nvidia H100 nearly 100 times more powerful than any previous orbital computer. Within weeks, Starcloud, a satellite company, announced that it had trained a language model on the complete works of Shakespeare and operated Google’s Gemini from about 200 miles above Earth.
“The spacecraft is performing better than we expected,” Philip Johnston, founder and CEO of Starcloud, told the Observer. Johnston founded the company in Redmond, Wash. at the beginning of 2024, supported by a case that attracted 200 million dollars from investors, including Nvidia, In-Q-Tel, Sequoia Capital and Y Combinator: Rather than building a better infrastructure on Earth to fill AI’s apperage energy, remove the infrastructure around the world. Today (March 30), the company announced that it has raised $170 million in Series A funding at a valuation of $1.1 billion, helping it reach unicorn status in just 17 months (and the fastest in Y Combinator history).
Johnston, 39, brings a clear financial pragmatism to the universe. Before moving into space, Johnston worked as an algorithmic trader at BNP Paribas, consulted with national space agencies at McKinsey and founded Opontia, an e-commerce aggregator that raised $46 million before being acquired.
A self-confessed astronaut, Johnston got the idea for Starcloud after a solo weekend trip to Starbase, the Texas city where SpaceX is headquartered, to see the Starship rocket. What struck him was across the street from the rocket: two gigafactories “like Tesla’s production lines,” he shared in a TED Talk in October, designed to each produce a new Starship every day. He went home, called his co-founders Ezra Feilden, a materials engineer he grew up with in the UK, and Adi Oltean, former lead software engineer at SpaceX, and started running numbers on what the launch force could do. Data centers kept coming as the answer.
Starcloud-1 is the first data center in space
The numbers behind his conviction are hard to argue with. An AI-focused data center uses as much electricity annually as 100,000 households, according to the International Energy Agency. In space, Johnston says, one square meter of solar panel produces eight times the power of an equivalent panel on Earth.
The Starcloud-1 is a 60-pound aluminum box packed with a total of five Nvidia GPUs, with the H100 getting the most attention, and other supporting hardware. It orbits the Earth every 90 minutes in a dawn-dusk orbit designed to keep its solar panels in constant sunlight. Starcloud is already running edge computing workloads in space for Earth observation and military satellites: Active inference—say, real-time applications like ChatGPT that don’t require as much computing power or data transfer as training LLMs—is on the way. “We can shorten the time to practical understanding from a few days to a few seconds,” Johnston said. For a long time, he has applied to the Federal Communications Commission (FCC) for a constellation of up to 88,000 satellites and envisions a 5-gigawatt orbital hypercluster. Starcloud does not stand as the next AWS, but as a neutral infrastructure. “I would consider us Equinix,” the CEO said, referring to the global data center developer and operator giant.
Starcloud plans to launch its second satellite, Starcloud-2, in October this year, and the third, Starcloud-2.1, in 2027. The commercial inflection point will happen with Starcloud-3: a 200-kilowatt, three-ton satellite. It is designed in such a way that 50 satellites, each weighing 3 tons, can be launched by SpaceX’s Starship rocket, which is still under development. Johnston projects mid-2028 as the time when orbital data centers are competitive with terrestrial costs, so Starcloud-3 will launch sometime between now and then, although it’s an ambitious bet on a launch vehicle. which has multiplied to exceed the coming programs. Further in the future is Starcloud-4: a satellite installed on the Starship with a four-kilometer solar system that powers a 5-gigawatt data center, a structure that can exceed the production capacity of the largest power plant in the US At the moment, however, it is a conceptual video more than an engineering blueprint.
“By moving AI compute into space, we open access to unlimited solar energy and completely remove the energy bottleneck,” said Johnston in relation to the latest fundraising, which will go mainly to the development of Starcloud-3, the expansion of the population, production capacity and business development. “This funding allows us to quickly scale our orbital infrastructure and meet the huge commercial demand for sustainable AI computing.”
Orbital’s data centers have a cooling problem
Many experts consider thermodynamics to be the most difficult bottleneck for orbital data centers. Since space is empty space, containing almost anything that gives off heat, it is naturally cold. It may seem like it’s good for cooling high-performance electronics, but it’s also protective.
“How do you get heat out of a vacuum? It’s like your thermos,” Josep Miquel Jornet, distinguished professor of electrical and computer engineering at Northeastern University, told the Observer. “It’s cold outside, great. It’s hot inside—great if you want tea, but not if you want to run a data center in space.”
Without air or
In Sequoia Capital’s latest episode The Training Data PodcastJohnson said cooling consumes about 70 percent of his engineering team’s attention. Starcloud-1 is equipped with a proof-of-concept cooling system that can operate continuously. The production version—a liquid-loop radiator with a custom heat sink that feeds a large deployable radiator panel aimed at space—will go aboard Starcloud-2 in October. “We tested it in hot rooms and vacuums, and it works,” he said. Johnston added that chip reliability is another concern: even a 10 percent higher failure rate in space than on Earth “will basically wipe out all the energy savings,” he said. Fortunately, the H100 on Starcloud-1 has no chip-level reboot failures so far.
Space data centers are gaining momentum amid regulatory and political risks
Starcloud, which currently employs 13 people, is not alone in this race. In January, SpaceX filed plans with the FCC for a constellation of up to one million data center satellites. The company cites orbital data centers as a key driver of growth as it prepares for a historic IPO. In the same months, Jeff Bezos’ Blue Origin unveiled its data center effort, Project Sunrise, which plans to launch up to 51,600 satellites (according to its FCC filing) into orbit. Google’s Project Suncatcher plans to launch two test satellites with its AI chips in 2027 in partnership with Planet Labs. Aetherflux, founded by Robinhood founder Baiju Bhatt and backed by Andreessen Horowitz and Breakthrough Energy Ventures, is targeting a 2027 launch of its first orbital data center. Axiom Space has already sent space data to the International Space Station. Against that background, Starcloud’s distinction is concrete: It is the only company that has demonstrated the operation of AI compute hardware in orbit.
Ian Christensen, senior director of private sector programs at the Secure World Foundation, which filed legal comments with the FCC about SpaceX’s data center request, noted that Starcloud’s installation of 88,000 satellites represented a significant increase over Starlink’s current fleet of 10,000 satellites, and that a minimum level of human safety was established. Currently, most satellites burn up when re-entering Earth’s atmosphere. Space policy experts at SWF warn that burning large numbers of aluminum satellites at scale could have unknown effects on the ozone layer. The FCC began revising its 1980s-era exemptions about a year ago without producing any change.
“It’s an area where I think there may be some risk of downstream regulation in some areas that may take a forward-looking approach,” Christensen told the Observer, noting that environmental law may end up working in ways that go far beyond traditional local regulation. “Space is a unique domain,” he said. “There are unique regulatory risks that will affect your business plan in ways you might not think about with a real-world plan.”
Johnston has been building against the odds since before Starcloud-1 left the ground. Now he has a satellite in orbit running real workloads to pay customers, an untested manufacturing cooling system in space and a chip reliability risk he points to as a potential business case. And given how much oil and, subsequently, energy prices have risen due to the escalation of global conflicts in the past few weeks, the appetite for risk that could solve the energy shortage has grown exponentially almost overnight.
While he may not have had specific current events in mind, it’s worth noting something Johnston told a TED audience in San Francisco last October: “The most effective way to save our children and grandchildren from the tree of war is to stop competing for Earth’s limited resources and start harnessing the near limitless potential of our solar system and ultimately our galaxy.
