Euwyn Poon spent the last decade watching Ford write him a check for close to $100 million after acquiring Spin, the electric scooter company he co-founded. That was November 2018. Now he's back with a pitch that makes dockless scooters look quaint by comparison: artificial intelligence servers floating in space.
Orbital Compute, Poon's Los Angeles-based startup, announced a $5 million pre-seed round on June 9, 2026—oversubscribed, the company says—to pursue what may be the most capital-intensive bet in the current AI infrastructure gold rush. The plan? Launch GPU-equipped satellites into low Earth orbit, where solar panels the size of half a tennis court will harvest continuous sunlight and radiators will dump excess heat directly into the vacuum of space.
No power grid. No land disputes. No water consumption for cooling. Just, the thinking goes, raw solar energy converted straight into machine learning inference.
Whether it's visionary or wildly premature depends entirely on launch costs—a dependency Orbital readily admits.
Betting on Starship
a16z Speedrun led the financing, joined by a constellation of smaller funds: Basis Set, Human Element, Wayfinder, Antler, Anti Fund, Ascent, Rubik, Zero Knowledge Ventures, LYVC, Feld Ventures, New Legacy, FNDR, UpHonest, and Asterisk. That's a long cap table for a pre-seed, which tells you something about both the ambition and the risk here.
The capital will bankroll "Pathfinder," a hosted GPU payload scheduled to hitch a ride on a SpaceX Falcon 9 rideshare mission sometime in 2027. Pathfinder is essentially a proof-of-concept: Can a GPU survive radiation in orbit? Will thermal management work as modeled? Can you get enough bandwidth back to Earth to make inference services commercially viable?
Meanwhile, Orbital is constructing Factory-1, a satellite assembly and testing facility in Los Angeles, and beginning early work on Orbital-1—a dedicated compute satellite the company hopes to launch in 2028, assuming Pathfinder doesn't blow up or cook itself.
"AI progress is being constrained by the grid," Poon said in the company's announcement. "Orbital is turning energy directly into intelligence." It's a memorable line, though one that glosses over the minor detail of needing tens of millions of dollars per launch to make the unit economics pencil out.
The Starship Problem
Here's the uncomfortable truth Orbital isn't hiding: at current launch prices, none of this makes financial sense. The company conceded as much in April coverage by The Register, acknowledging that today's economics simply don't work. The entire venture is a leveraged bet on SpaceX's Starship reaching operational cadence and driving launch costs down by an order of magnitude—or more.
If that happens, Orbital envisions scaling to a constellation of perhaps 10,000 satellites delivering a cumulative gigawatt of compute capacity. Each satellite would carry roughly 100 kilowatts of power and run on NVIDIA's Space-1 "Vera Rubin"-class GPU architecture, unveiled at the company's GTC conference in March 2026. (Yes, NVIDIA is now designing chips explicitly for the space environment. We live in strange times.)
The pitch has a certain elegance. Low Earth orbit offers 24/7 sunlight on properly positioned satellites, and radiative cooling in a vacuum is theoretically more efficient than any terrestrial data center chiller system. There are no utility interconnection queues, no community opposition to new power plants, no fights over water rights in the desert.
But there's also no room for error. A satellite that overheats or suffers a component failure can't exactly get a service call.
A Crowded Orbit
Orbital isn't alone in this particular fever dream. Starcloud, another space-based compute startup, raised a $170 million Series A at a $1.1 billion valuation in March 2026. Cowboy Space—formerly known as Aetherflux—closed a $275 million Series B in May to build both its own rockets and orbital data centers, presumably hedging against the Starship timeline.
That's nearly half a billion dollars chasing the same thesis across three companies in less than six months, which suggests either extraordinary foresight or the kind of exuberance that historically precedes a reckoning. Perhaps both.
Orbital's team of roughly a dozen employees includes veterans from Amazon's Project Kuiper LEO satellite group, SpaceX, and Northrop Grumman—serious aerospace credentials. That matters when you're asking investors to fund hardware that needs to survive launch vibration, thermal cycling, and years of radiation exposure while executing matrix multiplications.
The company says if Pathfinder validates the core technology, it will begin offering commercial inference services. Orbital-1, the 2028 follow-on, would carry multiple GPU nodes and high-bandwidth optical inter-satellite links, theoretically enabling distributed training workloads across the constellation.
The Waiting Game
For now, Orbital has runway to reach its 2027 test flight. After that, the company faces the same gauntlet every hard-tech startup eventually confronts: Does the physics work? Can you manufacture at scale? Will customers pay enough to justify the capital expenditure?
And hovering over all of it, the Starship question. SpaceX's fully reusable heavy-lift rocket remains in testing, with regulatory and technical hurdles still ahead. If it reaches operational cadence on the timeline Elon Musk has promised—well, Musk's timelines have a certain flexibility to them.
Poon's last company put electric scooters on sidewalks in dozens of cities before most people thought the idea was remotely viable. Ford clearly saw something in that bet. Whether orbital data centers prove equally prescient, or end up as expensive space debris, will likely take most of a decade to resolve.
Until then, Orbital is building satellites in South Bay and watching the skies.
