The pitch sounds almost too science fiction to take seriously: What if the answer to artificial intelligence's spiraling energy crisis wasn't another enormous data center in Iowa or Virginia, but rather thousands of satellites orbiting overhead, powered by unlimited solar energy and cooled by the vacuum of space itself?
That's the wager Starcloud is making. The startup announced a $170 million Series A round on March 30, 2026, at a $1.1 billion valuation—led by Benchmark and EQT Ventures—with plans to build what it calls the first commercially viable orbital data center business. It's an audacious goal, one that depends on technological breakthroughs still years away and launch costs that need to fall by an order of magnitude. But then again, Starcloud has already done something no one else has: it trained an AI model in orbit.
When Starcloud-1, a 60-kilogram satellite carrying an Nvidia H100 GPU, sent its first message from space in December 2025, the transmission included a tongue-in-cheek greeting—"Greetings, Earthlings!"—generated by Google's Gemma language model running on hardware 325 kilometers above Earth. That moment, however whimsical, represented the first time a data-center-grade processor had operated in the harsh environment of low Earth orbit. And it caught the attention of investors willing to fund what might be either the future of cloud computing or an expensive detour into orbital infrastructure.
The new funding brings Starcloud's total raised to roughly $230 million. Benchmark's Chetan Puttagunta joined the board. The round included a sprawling roster of backers: Macquarie Capital, NFX, Nebular, Y Combinator, Adjacent, 776, Fuse, Manhattan West, and Monolith Power Systems. That wide base suggests both enthusiasm for the concept and perhaps some hedging—no single investor is betting the farm, but many want exposure to the idea.
Something Old, Something New
Space-based computing isn't entirely novel. Satellites have carried processors for decades, running navigation algorithms, imaging pipelines, and communications protocols. But those have typically been specialized, radiation-hardened chips designed for the extremes of space—expensive, custom-built, and far less powerful than their terrestrial cousins.
What Starcloud did differently was fly an 80 GB Nvidia H100, the kind of GPU you'd find humming away in an Amazon or Microsoft data center, pulling hundreds of watts and processing the matrix multiplications that underpin modern AI. IEEE Spectrum noted the milestone represented roughly 100 times the compute power of previous space-borne processors, like the edge computing boards used for on-orbit image processing.
Starcloud-1 launched to low Earth orbit in early November 2025, built on Astro Digital's Corvus-Micro spacecraft bus. The satellite settled into a sun-synchronous orbit between 325 and 350 kilometers altitude—low enough that atmospheric drag will pull it back to Earth within about 11 months, eliminating the need for active deorbiting but also limiting its operational lifespan.
In December, the company ran two demonstrations. First, it trained NanoGPT, an open-source language model created by Andrej Karpathy, on a Shakespeare dataset. Training—not just running a pre-trained model—is computationally intensive and involves iterative adjustments to billions of parameters. Doing it in orbit, where radiation can flip bits in memory and thermal management is a constant battle, proved the hardware could handle more than static inference tasks.
Then came the Gemma model inference, generating that playful message. The demonstrations weren't just technical showpieces. They validated that modern AI workloads could run in space despite the radiation environment, thermal challenges, and power constraints that make low Earth orbit a hostile place for consumer electronics.
The satellite has also demonstrated capabilities in processing satellite imagery on-orbit, a more practical application. By analyzing data in space, Starcloud can downlink only the insights—targets identified, changes detected—rather than raw imagery, which can be enormous. That approach reduces bandwidth requirements and cuts the time between image capture and actionable intelligence, which matters in applications like disaster response or defense.
The Economics Depend on Someone Else's Rocket
CEO Philip Johnston told TechCrunch in March that orbital data centers could reach cost competitiveness around $0.05 per kilowatt-hour. For context, that's in the ballpark of what hyperscale cloud providers pay for electricity in regions with cheap power. But—and this is where the business model gets dicey—that figure only works if launch costs drop to roughly $500 per kilogram.
SpaceX's Falcon 9, while more affordable than previous launch vehicles, still prices significantly above that target for most missions. Johnston's timeline hinges on Starship, SpaceX's massive next-generation vehicle, reaching high flight rates and full reusability. He expects that around 2028 or 2029, assuming SpaceX hits its targets.
In other words, Starcloud's path to profitability depends on another company—one known for ambitious timelines that often slip—solving one of the hardest engineering challenges in spaceflight. That's a lot of faith to bake into a $1.1 billion valuation.
Until Starship matures, Starcloud plans to keep launching smaller satellites on Falcon 9, refining its technology and building a track record. Starcloud-2, slated for later in 2026, will carry multiple GPUs including Nvidia's newer Blackwell architecture, an AWS server blade, and—in a curious addition—Bitcoin mining hardware for testing. According to TechCrunch, the mission will sport what the company describes as its largest deployable radiator yet, a critical piece of infrastructure for managing heat in the vacuum of space.
Managing thermal loads is one of the trickier problems in orbital data centers. On Earth, you can blow air over heatsinks or circulate chilled water. In space, the only option is radiation—literally emitting infrared energy into the void. That requires large surface areas, which means deployable structures that fold up for launch and unfurl in orbit.
Starcloud-3 represents the company's first genuine attempt at scale. The design calls for a 200-kilowatt, three-ton spacecraft specifically engineered for Starship's "Pez dispenser" deployment system, which can release multiple satellites in rapid succession. That mission is meant to be the first orbital data center that can compete on cost with terrestrial facilities—assuming, again, that launch prices cooperate.
In October 2025, cloud computing partner Crusoe announced it would become the first public cloud operator to run workloads in outer space, with a dedicated module on a Starcloud satellite targeting limited capacity by early 2027. That partnership offers Starcloud a revenue stream and a real customer, even if the economics aren't yet compelling for most applications.
A Suddenly Crowded Sky
Starcloud isn't alone in eyeing the opportunity overhead. In February 2026, the FCC accepted SpaceX's application to deploy up to one million satellites for an orbital data center system. One million. The filing suggested Elon Musk's company views space-based computing as more than a side project—perhaps as a natural extension of Starlink's infrastructure or a hedge against future data center constraints on Earth.
Google announced Project Suncatcher in March 2026, a research initiative to scale machine learning in space using solar-powered satellites equipped with TPUs. A two-satellite learning mission with Planet is planned by early 2027. Google's involvement lends credibility to the concept, though the company has a habit of launching ambitious research projects that don't always materialize into products.
Kepler Communications revealed in mid-March that it had deployed distributed GPU compute across its optical-relay constellation, positioning itself as a space-based scalable cloud. Axiom Space has been developing orbital data center units for the International Space Station, including AxDCU-1 running Red Hat Device Edge, with plans for broader ODC nodes. The ISS work has the advantage of a stable platform and easier access for hardware upgrades, but it's limited in scale and carries its own costs.
Even Nvidia is taking the concept seriously. At GTC in March 2026, the company launched a "Space Computing" initiative and introduced the Vera Rubin Space Module, claiming up to 25 times the compute of an H100 for future orbital inference. Nvidia's announcement listed Starcloud among six companies using its platforms for space and ground workflows. When the world's leading AI chip maker starts talking about space-specific hardware, it's a signal the market may be real.
The competitive landscape suggests the industry has moved past the question of whether AI can run in space. Now it's about scale, cost, and commercial viability—questions that won't be answered for years.
The Pitch: Infinite Sun, Infinite Cold
The environmental case for orbital data centers is straightforward, at least in theory. Terrestrial data centers consume enormous amounts of water for cooling—billions of gallons annually for large facilities—and strain local power grids. Starcloud-1 flies in a sun-synchronous orbit at roughly 83 degrees inclination, maximizing solar exposure throughout its orbit. In space, energy is effectively limitless, and cooling is free via radiation to the void.
But the environmental argument has a counterpoint: launching rockets to orbit isn't exactly green. Each Starship launch will burn hundreds of tons of methane and liquid oxygen. If orbital data centers require continuous launches to replace deorbiting satellites, the carbon math gets complicated quickly.
The economics only work if launch costs collapse. At today's prices, sending computing hardware to orbit remains prohibitively expensive for most workloads. Starship changes that calculation—if it reaches high flight rates and reusability targets. Johnston's timeline of 2028 to 2029 for commercial Starship access means Starcloud's biggest infrastructure bets are contingent on another company's engineering roadmap.
There's also the question of what workloads make sense in orbit. Some applications—processing satellite imagery on-site, for instance—have obvious advantages. Others, like training the next generation of large language models, seem less suited to platforms with limited bandwidth to Earth and no possibility of physical repair. Finding applications where the benefits of space-based processing outweigh the costs and constraints will be critical to the business model.
Betting on a Timeline
The $1.1 billion valuation prices in considerable optimism. It assumes Starcloud can navigate the technical risks of operating data center hardware in an environment where repair is impossible and deorbiting is inevitable. It bets that the company which proved AI could train in space will be the one to make space-based computing economically viable. And it wagers that SpaceX will deliver on Starship's promise within the next few years.
That's a lot of variables. But Starcloud's investors have reason for confidence, or at least cautious enthusiasm. The company has demonstrated technical capability. It has partnerships with credible players like Crusoe. And it's entering a market where demand for AI compute seems insatiable, with terrestrial data centers already struggling to keep pace.
Whether that translates into a sustainable business—one that can compete with the hyperscalers and deliver returns to investors—remains an open question. For now, Starcloud is flying higher than most startups, literally and figuratively. The question is whether it can stay aloft long enough for the economics to catch up.
