Stay connected with KayaToday—follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
There is a particular moment in the development of a genuinely new technology when the conversation shifts from whether something is possible to how quickly it can be scaled. Orbital computing crossed that threshold this year. In November 2025, a satellite the size of a small refrigerator, carrying a single Nvidia H100 chip, was launched into low Earth orbit by a Redmond, Washington-based startup called Starcloud. Within weeks, that satellite had trained a large language model and queried an open AI system from Google while circling the planet at 28,000 kilometres per hour.
It was the first time AI training had been demonstrated to function reliably outside Earth’s atmosphere. The company has since raised $170 million, reaching a $1.1 billion valuation just seventeen months after its Y Combinator demo day — the fastest unicorn ascent in the accelerator’s history.
The pitch, stripped of its undeniable novelty, is straightforward and addresses a problem that has become impossible to ignore on Earth: AI data centres require enormous amounts of electricity and water for cooling, and both resources are becoming structurally scarce in the jurisdictions best positioned to host them. Space, Starcloud’s founders argue, offers an answer to both constraints simultaneously — uninterrupted solar power and a near-infinite thermal sink with no risk of depleting a local aquifer or straining a regional grid.
What Has Actually Been Demonstrated
It is worth being precise about what Starcloud-1 proved, because the gap between demonstration and infrastructure-grade deployment remains substantial. The satellite operated with roughly one kilowatt of solar power — sufficient to run a single H100-class GPU on a duty-cycled basis, not the continuous, multi-megawatt clusters that define terrestrial AI training today. Nvidia’s own framing of the achievement was characteristically expansive — the company describes the milestone as unlocking “a new era of space innovation” through what it calls orbital data centres, or ODCs — but the realistic distance between a single chip running NanoGPT in orbit and a commercially viable AI training cluster is measured in years, not months.
Read also this: https://kayatoday.com/news/silicon-valley-has-discovered-nuclear-power-it-did-not-have-a-choice/
The company’s next milestone, planned for October 2026, narrows that gap considerably. Starcloud-2 will carry Nvidia’s Blackwell B200 chip — the same processor powering the largest terrestrial AI clusters today — alongside multiple H100s, a server blade from Amazon Web Services, and seven kilowatts of solar generation, a sevenfold increase over the demonstration satellite. The inclusion of AWS hardware is the detail that deserves the closest reading: it signals that Starcloud is working to make orbital compute accessible through the same cloud APIs that enterprise customers already use, treating a satellite in low Earth orbit as, functionally, another availability zone. If that integration succeeds, the adoption barrier for customers — who would otherwise need entirely bespoke infrastructure to use orbital compute — falls dramatically.
The Physics That Makes the Case, and the Physics That Complicates It
The argument in favour of orbital computing rests on genuine physical advantages. Solar panels in space, unobstructed by atmosphere, clouds, or the day-night cycle in the right orbital configuration, generate considerably more consistent power than terrestrial solar installations. The vacuum of space, counterintuitively, can support certain cooling approaches without the water consumption that has made terrestrial data centres a source of local political controversy — a concern made more acute this month by reporting that SpaceX itself has flagged water scarcity as an investor risk for its own terrestrial infrastructure ambitions.
The complicating physics receives less attention in the promotional material than it deserves. Radiation exposure degrades semiconductor performance over time in ways that require either expensive radiation-hardened components — which lag commercial chips by multiple generations — or a tolerance for shorter operational lifespans than terrestrial hardware enjoys. Heat rejection in vacuum, while avoiding water consumption, requires large radiator surfaces that add mass and complexity to satellite design. And the economics of launch, even as costs have fallen substantially over the past decade, still impose a capital intensity on orbital infrastructure that has no equivalent on the ground. Nvidia’s newly announced Space-1 Vera Rubin Module, delivering up to 25 times the AI compute of the original H100 demonstration, is explicitly designed for “size, weight, and power-constrained environments” — an acknowledgement, in the engineering specification itself, that space remains a fundamentally harder environment to build in than a warehouse in Virginia.
What This Means Beyond the Novelty
The serious strategic question raised by orbital computing is not whether it is technically possible — November’s demonstration settled that — but whether it can scale fast enough and cheaply enough to meaningfully relieve the terrestrial power and water constraints that are now visibly limiting AI infrastructure expansion in the United States, Europe, and parts of Asia. Local governments rejecting data centre permits, utilities warning of grid strain, and corporate disclosures flagging water access as a material business risk are all symptoms of a terrestrial capacity ceiling that companies like Amazon, Google, Microsoft, and Meta are spending tens of billions of dollars annually to push back, including through the nuclear power commitments examined elsewhere in this publication.
Orbital data centres will not replace that terrestrial buildout within any near-term planning horizon. But the fact that Nvidia, AWS, and a venture-backed startup with a billion-dollar valuation are now treating space as a credible extension of cloud infrastructure — rather than a speculative curiosity — signals that the industry’s capacity constraints have become severe enough to justify capital allocation toward solutions that, eighteen months ago, would have been dismissed as implausible. That shift in seriousness, more than the technical demonstration itself, is what makes this development worth tracking closely over the next several launch windows.
Read Also: Silicon Valley Has Discovered Nuclear Power, It Did Not Have a Choice
