Google, SpaceX et Starcloud investissent dans les centres de données en orbite. Entre IA, cloud spatial et défis techniques, un nouveau marché se dessine.

The growing interest in orbital data centers shows that the cloud industry is already preparing to expand into space. Google is experimenting with artificial intelligence processors in orbit through its Suncatcher project, SpaceX is working on a constellation that could include up to one million satellites, and start-up Starcloud is attracting hundreds of millions of dollars from investors. In March 2026, the latter raised $170 million before announcing another $250 million funding round a few months later, bringing its valuation to $2.3 billion.

宇宙データセンターとは、人工衛星や宇宙ステーション内にサーバーやストレージ、通信設備などを設置し、データ処理をできるようにした施設のこと。米SpaceX（スペースX）や米Starcloud（スタークラウド）といった企業が構想を打ち出している。

Gemma 3 runs on Nvidia hardware to analyze satellite images

Two things became true about orbital computing in the past five weeks, and they point in opposite directions. The capital went vertical. Starcloud is reportedly valued at $2.3 billion. Cowboy Space closed a $275 million Series B at a $2 billion post-money valuation, then showed a rocket whose upper stage stays in orbit as a megawatt-class data center. Google put four data-center TPUs on a prototype and got a haiku back. The only company in the category publishing a price list is a 2024 startup with $8 million in seed money and software on three spacecraft it does not own. On October 30, Satlyt said it had tested task-planning software on hardware in orbit. One computer was real. Three were simulated. That gap between the balance sheets and the test bench is the actual investment question. ## The case for owning the whole stack The vertical argument is straightforward. Compute in orbit is a power and thermal problem before it is a software problem. A rack that cannot shed heat is a rack that does not run. Google's Project Suncatcher prototype carries four Trillium TPUs and about 1 kilowatt of solar generation. It can run one TPU for roughly 20 minutes at a time. The satellite is online, and the question has moved from whether inference works in orbit to how long the chips survive thermal stress and radiation. Cowboy Space attacks the same constraint from the launch pad. Its Big Iron vehicle is designed to carry more than 15,000 kilograms per flight, with an upper stage that becomes the data center. The skin doubles as a radiator. The ammonia propellant doubles as coolant. First launch is targeted for no earlier than December 2028, and the company has filed an FCC application for the constellation. Starcloud, backed by Benchmark and valued at $2.3 billion, follows the simpler version: inference on individual satellites. SpaceX is expected to use laser links, though its nodes are projected to orbit too far apart for coherent distributed workloads. This is the standard infrastructure trade. Own the power, the thermal loop, and the ride to orbit, and you capture the full margin when demand arrives. The risk is equally standard: capex lands years before revenue, and every schedule slip compounds. ## The case for renting the layer Satlyt starts from the opposite claim. Most satellites already carry more compute than they use, and most of that compute reaches a ground station for five to ten minutes per orbit. The bottleneck is not FLOPs. It is routing. The company sells software that sits between the spacecraft and the application. Operators keep their hardware. Developers buy a deployment slot. Published pricing runs from $5,000 for a single application on one compatible host to $10,000 for a larger pilot, with managed missions starting at $25,000. There is measured evidence behind the pitch. Satlyt ran Google's Gemma model on a Momentus spacecraft and cut the size of an onboard software-error transmission by more than 60%. CEO Rama Afullo says that kind of reduction can save hundreds of thousands of dollars per satellite each year. The company also holds a NASA STTR contract worth $4 million to design delay-tolerant edge protocols. > If they are the iPhone, we will build Android as a horizontally integrated, open ecosystem.— Rama Afullo, founder and CEO, Satlyt The seed round was led by non sibi ventures, with TLCOM, Antler, Launch Africa Ventures and others participating. Afullo told Payload the total raised is now well past $10 million. Payload also reports an MoU with xOrbita and Symphony Space to test a joint software-defined architecture, which matters because it puts Satlyt inside the hardware camp rather than against it. 🎯 **The two poles are not substitutes.** Vertical stacks are financing power, heat rejection, and launch capacity. Satlyt is financing a routing layer across spacecraft that already fly. The first model needs billions and a Starship cadence. The second needs operator permissions and a working multi-spacecraft demo. Only one of them can be underwritten today. ## What the October test actually proves The test modelled four computers, four users, and eight services. Satlyt's software decided where each task should run, based on available compute and communication links, and sent the plan back to Earth. That is a real result. It is also narrower than the headline. Only one node was physical. The other three were simulations, so the experiment did not move data between separate spacecraft or execute work across them. Afullo says the best-performing scheduling approach also consumed more computing power, and the team found a logging error: the software recorded when a delay arrived instead of how long it lasted. NASA funded the research through the University of Houston. NASA has not selected Satlyt for a mission. The next milestone is a demonstration of computing shared across separate spacecraft, which the company expects to attempt next year. As we wrote in October, Cowboy's Big Iron makes the opposite trade: it treats the data center as a launch vehicle problem. That piece mapped the power and thermal math. Satlyt's result maps the scheduling math. Neither closes the other's gap. ## Where the margin accrues Dimension| Vertical stack| Software layer ---|---|--- **Capital** | $2B+ valuations, $275M rounds | $8M seed, $4M NASA STTR **Revenue unit** | Compute capacity by the hour | Pilot packages from $5,000 **Moat** | Power, thermal, launch cadence | Installed base on non-Starlink buses **Proof point** | Prototypes in orbit | 3 spacecraft, 1-node planning test **Principal risk** | Capex years before revenue | Operator permissions, untested routing Company disclosures, TechCrunch and Payload reporting, October 2026. The vertical players are selling a future capacity market. Satlyt is selling a present software service, priced low enough that a satellite operator can test it without a board decision. The share target is where the stories diverge. TechCrunch reports Satlyt aims for about 20% of satellites by 2030. Payload quotes Afullo at 30% of all non-Starlink satellites by the end of the decade. The second number is the one that would justify a venture outcome. The first is the one currently supported by three hosted missions. For an investor, the layer is an option on every bus that is not built by SpaceX. It is cheap, it has revenue, and it does not require a rocket to work. The option stays cheap until a multi-spacecraft demo lands. ## The physics both models share Neither business escapes the same three constraints. Power is the first. A ground rack draws tens of kilowatts from a grid that never sets. An orbital rack competes with instruments, communications, and attitude control for a solar budget measured in single-digit kilowatts, and it loses half of that budget in eclipse. Heat is the second. Vacuum removes convection from the toolbox. Every watt of compute has to be radiated away, which means surface area, which means mass. Cowboy's decision to turn the rocket's skin into a radiator is an admission that the thermal loop, not the GPU, sets the ceiling. Bandwidth is the third. Processing data onboard only helps if the answer is smaller than the raw input and arrives inside a pass window. Satlyt's 60% reduction in error-transmission size is valuable for exactly that reason. It buys back downlink time on a spacecraft that already launched. The models diverge on which constraint they own. A vertical stack can design power and heat into the vehicle from the first drawing. A software layer has to live inside somebody else's thermal envelope and negotiate for its share of the link. That is cheaper to start and harder to guarantee. It also explains why the two camps keep signing agreements instead of only competing. Hardware operators need workloads to fill capacity. Software vendors need hosts. The MoU with xOrbita and Symphony Space is a small signal that both sides see the same dependency. ## What to underwrite next year Four checkpoints separate the pitch from a position. First, a demo that shares work across at least two physical spacecraft, not one real node and three simulations. Second, paying customers beyond pilot packages, with retention measured in mission cycles rather than logos. Third, evidence that vertical operators will license routing instead of building it. Fourth, a NASA mission selection that converts the STTR relationship into a flight commitment. Until then, the honest comparison is not Satlyt versus Starcloud. It is a $5,000 software pilot versus a $2 billion power plant. Both can be right, and they will not be right on the same timetable. Satlyt tests orbital task-planning on a single live spacecraft TechCrunch's October 30 report supplies the test design: four computers, four users, eight services, one physical node, and the caveats Satlyt itself acknowledges. TechCrunch Primary account of the milestone, including the limits the company did not lead with. Satlyt raises $8M to expand its virtual orbital data-center network Payload details the seed round, the total raised, the Gemma inference result, and the commercial positioning against hardware-first orbital data centers. Payload Space Best source for the funding structure and the CEO's own framing of the software-versus-hardware split. Satlyt's published orbital deployment packages The company's own pricing page lists $5,000, $10,000, and from $25,000 packages, which is the clearest evidence of a commercial model rather than a demo program. Satlyt Used for the revenue model; official company source, not independent reporting.

Firefly stock hub: Starcloud, lunar missions, Alpha, SciTec, contract margins, liquidity covenants and dilution, with dated financial evidence.

Firefly stock hub: Starcloud, lunar missions, Alpha, SciTec, contract margins, liquidity covenants and dilution, with dated financial evidence.

Starcloud's SC-1L GPU payload will fly aboard Firefly's Elytra orbiter, targeting a lunar AI computing demonstration no earlier than 2028.

YouTube video by Starcloud

Firefly stock hub: Starcloud, lunar missions, Alpha, SciTec, contract margins, liquidity covenants and dilution, with dated financial evidence.

修正😉 ねえ 聴こえますか 流れ星が溢したアリア(孤独な唄) ねえ 知ってますか 消えた 星の逝く先 探しに行こう 赤色の木箱は意図の美談 まるでこの世は Like a monster(化物みたい) 曾ては星渦の色 永彗 繰り返した週明けの悪魔 コインは落ちる… 空っぽに成った心 諦める事が正義ですか 散らばった星座 君を捜した 涙色の空 夜の隙間 雨風の中でも輝いた 僕たちの想いはきっと...

The Conversation article discusses the growing trend of corporate proposals for orbital 'data centres' using massive satellite constellations, which experts warn pose significant risks to Earth and the night sky. SpaceX has initiated this trend with a proposal for one million satellites, followed by other companies like Blue Origin, Starcloud, Cowboy Space, and Orbital Compute. These proposals, which invoke the Kardashev scale and describe creating a 'Kardashev II-level civilization,' have been criticized as detached from reality by aerospace and technology experts. The article highlights several critical concerns: the unprecedented scale of these satellite constellations, the risk of satellite debris re-entering Earth's atmosphere, and the impact on the night sky. SpaceX's Starlink, already the most numerous satellite constellation with over 11,000 satellites, experiences daily re-entries of car-sized debris, with one piece discovered on a Saskatchewan farm in 2024. The article notes that if the proposed million-satellite constellation follows similar re-entry patterns, 500 satellites could drop debris on Earth every five years. Additionally, the proposed satellites are significantly larger than the International Space Station, increasing collision risks. The article also discusses how these constellations would alter the night sky, with predictions that one in 15 visible points could be a moving satellite, a prediction now seen as conservative. The Federal Communications Commission (FCC) has been criticized for its 'licensing assembly line' approach, approving proposals with minimal scrutiny despite public opposition and scientific evidence of potential harm. The article concludes that the future of low-Earth orbit and the shared night sky is at stake, with the FCC holding the deciding vote on whether it becomes a dystopian reality.

Orbital data centres are computers on satellites that process data in space using solar power. Prototypes are flying: China's Three-Body constellation (May 2025), Starcloud-1 with an NVIDIA H100 (Nov 2025), Axiom's nodes (Jan 2026) and Google's Suncatcher TPU satellite (Oct 2026). From 2017 to 2026, this orbital data centres guide traces 11 defining moments, including AI data centres in space and space-based computing. Explore the full timeline.

Starcloud signed with Firefly Aerospace to test an AI compute payload around the Moon from 2028. Here is what each side gets, who competes and what could go wrong.