By StarcloudLearning Objectives
- Understand what a space-based (orbital) data center is and why anyone would build one
- Separate Starcloud's genuine, verified milestone from its aspirational long-term vision
- Evaluate the hard engineering problems — heat, radiation, launch cost — that decide whether the idea works
What Is Starcloud?
Starcloud is a venture-backed startup building space-based data centers: satellites that carry GPUs, run them on large solar arrays, and get rid of waste heat by radiating it into space. Founded in January 2024 as Lumen Orbit and renamed Starcloud in March 2025, the company is based in Redmond, Washington, and came out of Y Combinator's Summer 2024 batch.
The pitch rests on two things orbit offers that Earth increasingly rations for AI. In the right orbit a satellite sees the sun almost continuously, so its solar power is near round-the-clock and never competes for grid electricity. And in the vacuum of space there is no need for water or air cooling — a radiator can dump heat straight to the cold of space. Starcloud markets this as dramatically cheaper power and always-on solar energy, though those specific numbers are company claims, not independently verified results.
In November 2025, Starcloud reached a real first: its Starcloud-1 satellite carried the first NVIDIA H100 GPU into orbit aboard a SpaceX Falcon 9 rideshare. In December the company reported running language-model inference and even small-scale model training on that GPU in space — genuine milestones for the field.
In August 2026 the company added $250 million to its March Series A, led by Manhattan West, valuing it at $2.3 billion — roughly double the $1.1 billion it carried five months earlier. NVIDIA put in $25 million and Cisco also participated. What makes the round worth reading closely is the reason management gave for raising it: not heat, not chips, but launch capacity. SpaceX plans to retire the Falcon 9 program in 2028, and the rockets meant to replace it — from Blue Origin, United Launch Alliance and Rocket Lab — are either not yet flying regularly or still in development. Part of the money is simply to book rides. That is a useful corrective to how orbital-compute stories are usually told: the near-term constraint is a logistics queue, while heat rejection remains the constraint that decides whether the gigawatt vision ever works at all.
💡Key Concept
Space-based data center: A satellite (or, eventually, a cluster of them) that does the same job as a ground data center — running AI training and inference on GPUs — but powered by solar panels in orbit and cooled by radiating heat into space instead of by grid power and water. The appeal is escaping Earth's limits on electricity, land, and cooling; the difficulty is that everything in space is expensive to launch, hard to cool, and impossible to physically repair.
Key Capabilities
- Orbital GPU compute — Starcloud-1 put a working NVIDIA H100 in low Earth orbit, reported as roughly one hundred times more compute than any prior space-based system
- In-orbit AI workloads — ran inference on Google DeepMind's Gemma model and fine-tuned a small model on-orbit, plus processed satellite radar data in space
- Near-continuous solar power — a Sun-synchronous orbit keeps the panels illuminated far more than ground solar, the company's core cost argument
- Radiative cooling — waste heat is shed to space by large radiators rather than water or air
- Roadmap — Starcloud-2 satellites in 2027 carrying about eight kilowatts of compute for both training and inference, a Starcloud-3 design with a 200-kilowatt power footprint intended for mass production on SpaceX's Starship, a partnership with Crusoe for limited orbital GPU access from 2027, and a long-term vision of gigawatt-scale orbital data centers
⚠️Warning
This is a frontier roadmap, not a product you can rent today. What is real is a single roughly 60-kilogram demonstration satellite with one H100 GPU and an expected lifespan of under a year. Everything else — commercial GPU access via Crusoe in 2027, a filed constellation of tens of thousands of satellites, and gigawatt data centers spanning kilometers of solar array — is planned and unproven. The binding problem is heat: in a vacuum the only way to lose heat is to radiate it, and outside experts at the World Economic Forum, IEEE Spectrum, and Scientific American note that even a one-megawatt center would need a radiator roughly the area of a hockey rink, scaling enormously for the gigawatt vision. Radiation damage to chips, launch cost, and the impossibility of hands-on repair are all unsolved at scale. Treat Starcloud's cost and power claims as marketing, and the grand vision as years away.
Company Details
| Detail | Info |
|---|---|
| Company | Starcloud (formerly Lumen Orbit) |
| Founded | January 2024; renamed to Starcloud March 2025 |
| Headquarters | Redmond, Washington |
| Founders | Philip Johnston (CEO), Adi Oltean, Ezra Feilden |
| Accelerator | Y Combinator (Summer 2024); NVIDIA Inception |
| Funding | Roughly $420 million total; $170 million Series A (March 2026) plus a $250 million extension (August 2026) |
| Valuation | $2.3 billion (August 2026 extension, led by Manhattan West); $1.1 billion at the March Series A |
| Milestone | First NVIDIA H100 GPU in orbit (Starcloud-1, November 2025) |
| Status | Demonstration satellite operating; commercial access targeted for 2027 |
| Website | starcloud.com |
Key Takeaways
- Starcloud is building solar-powered, radiatively-cooled data centers in orbit, betting that space solves AI's electricity and cooling limits
- The verified milestone is real and significant — the first NVIDIA H100 GPU running, and a small model training, in orbit — and investors have backed it accordingly, with a $250 million extension in August 2026 taking the valuation to $2.3 billion, roughly double its March mark
- Management says the near-term bottleneck is launch capacity rather than technology, because SpaceX retires Falcon 9 in 2028 and the replacement rockets are not yet flying regularly; heat rejection is still the constraint that governs the long-run vision
- The honest caveat is that today's hardware is one demo satellite; gigawatt orbital data centers are years away and gated on unsolved heat-rejection, radiation, launch-cost, and servicing problems, so treat the grand vision as aspirational
