On 01/10/2026 a satellite about the size of a fridge went up on a Falcon 9 rideshare carrying four of Google's Trillium AI chips. It runs on about a kilowatt of solar power and is expected to operate for about a year. It is the first hardware from Project Suncatcher, and it arrived in a year in which the US Federal Communications Commission accepted for filing four applications for data-centre constellations in low Earth orbit (LEO), a few hundred to 2,000 km up: SpaceX for up to one million satellites, Starcloud for 88,000, Blue Origin for 51,600 and Orbital Compute for 100,000. None has been granted. The numbers are ceilings, not build plans. A filing asks for the most an operator could ever deploy, so that it can secure spectrum rights and orbital shells long before the technology is ready to launch. Of the four, only SpaceX, with more than 10,000 Starlink satellites launched, has built a constellation at scale.
The question we hear from people who finance aircraft and ships is whether compute in orbit becomes a leasing market. Our answer: the chips will not be leased this decade, but the long-life platform they sit on could be, the way aviation finances airframes separately from engines. That depends on choices operators have not yet made, and it will not happen on the timetable the filings imply.
Why anyone wants to do it
The case is about power and permits. The four largest US cloud companies are on course to spend about US$760bn on capital expenditure this year, most of it on AI data centres, and the binding constraint is now the grid connection. Ireland knows this better than most. Data centres used 23% of the country's metered electricity in 2025, and the CRU's connection policy of 12/12/2025 now requires new sites of 10 MVA or more to bring dispatchable generation or storage at least equal to their import capacity. A satellite in a dawn-dusk sun-synchronous orbit sees the sun almost continuously, needs no planning permission and draws on nobody's grid.
Why it is harder than it looks
The demonstrators are testing whether the chips survive. Most of what sits between a demonstrator and a business is still unsolved, and each problem bears on what the asset would be worth to a financier.
| Problem | Where it stands, October 2026 | Why a financier cares |
|---|---|---|
| Heat | In a vacuum heat leaves only by radiation, so radiator area grows with compute. Google's prototype runs its chips in 15-minute bursts, then shuts them down to cool. Starcloud's engineers are still working the trade between chip temperature and radiator size. | Radiators are mass, and mass is launch cost. Chips that must rest earn less than a capacity contract assumes. |
| Radiation | Google's TPUs took 15 krad(Si) in a proton beam with no hard failures, about 20 times the expected five-year dose. The high-bandwidth memory showed errors from 2 krad(Si), which Google judges acceptable for inference only. | The payload is limited to inference work, and error rates in orbit are not yet measured. Both feed availability. |
| Links between satellites | A cluster that works as one data centre needs tens of terabits per second between satellites. Google's best bench result is 1.6 Tbps. Its laser-link test pair flies in 2027. | Without that bandwidth the constellation is many small computers rather than one large one, and it earns accordingly. |
| Formation flying | Google's reference design holds 81 satellites within a 1 km radius, 100 to 200 m apart, at 650 km. It has not been flown. | Close formation raises collision and debris exposure, and insurers will price it. |
| Getting data down | Google lists high-bandwidth ground links among the unsolved problems. | The ground stations become part of the asset and part of the security package. |
| Launch price | Falcon 9 costs about US$1,000 per kilogram. Google puts the threshold at about US$200 per kilogram and expects it by the mid-2030s. The only disclosed commercial Starship price is US$90m a flight. | The cost case rests on a vehicle not yet flying commercial payloads. |
| Launch availability | Starcloud's chief executive says launch is constrained now because Falcon 9 is due to end in 2028, and that failing to book SpaceX capacity for 2029 would be a problem. | Delivery dates depend on someone else's manifest. |
| The chips themselves | Starcloud is designing around Nvidia's Vera Rubin Space-1, a space GPU not yet built, which it hopes to fly in late 2028. Starcloud-2, two 8 kW compute satellites on rideshares, has moved into 2027. | The equipment a lessor would be asked to finance does not exist yet. |
| Repair and replacement | No operator can swap a board in orbit today, and in-orbit servicing is at demonstration stage. | Without it the whole satellite has the life of its chips. |
Obsolescence sits on top of all of this. On the ground, operators depreciate GPU servers over five to six years (Meta 5.5, CoreWeave six), Nvidia's chief executive on 21/10/2026 argued for eight or more, and sceptics say two to three. A chip in orbit cannot be refreshed, so a lender will take the short end of that range.
None of this rules the idea out. It does mean the next three years are demonstrators financed by equity.
The leasing question
Lessors finance assets that outlive the lease, that someone else can use if the lessee fails, and whose value a third party can appraise. An orbital data centre today fails all three tests. The lessees are venture-backed; the best-funded raised a US$170m Series A in March 2026 and a US$250m extension in August, at a US$2.3bn valuation. The compute payload is obsolete before a normal lease term ends. There is no secondary market and no appraiser.
The asset does split in two, though, and one half looks familiar.
Aviation solved a similar problem decades ago. The airframe lasts 25 years. Engines are removed, overhauled and swapped, and a whole leasing industry finances them separately. We set out where the aviation analogy holds for satellites in From aviation to orbit. An orbital data centre has the same split. The platform (structure, solar arrays, radiators, power, propulsion and laser terminals) can be built as long-life infrastructure. The compute modules are short-life equipment that should be refreshed.
That first clause is the weak joint. Satellites in low Earth orbit have mostly been built for five to seven years, and SpaceX plans on about five for Starlink. At 600 to 650 km, atmospheric drag uses up propellant, and the FCC rule adopted in 2022, which requires disposal within five years of the end of a mission, gives operators little reason to over-build. A platform that barely outlives its chips gives a lessor nothing to hold. The thesis needs operators to choose long-life platforms, with the propellant, refuelling interfaces and servicing fixtures to match, and to pay more up front for them.
If the platforms are built to take replacement compute modules, and in-orbit servicing makes the swap routine, a lessor can own the platform on a long lease while the operator or a cloud company owns and refreshes the chips. That is a financeable structure, provided the lessor can take the platform back if the operator fails. With the Space Protocol to the Cape Town Convention not yet in force, that right has to be written into the lease and the ground-control arrangements rather than relied on from a registry. Without in-orbit replacement it is a disposable satellite with a three-year life, and the right money for that is the operator's equity or a cloud company's balance sheet.
Will it happen?
Four things have to be true at once:
| Signpost | What it proves | Earliest realistic date |
|---|---|---|
| Starship flying commercial payloads at cadence | Launch cost starts down the path to about US$200 per kilogram | Cadence 2028–29; that price by the mid-2030s on Google's estimate |
| A cloud company signing a multi-year capacity contract with an independent orbital operator | A creditworthy offtaker exists | Possible from 2027 |
| A platform designed for in-orbit compute replacement, and a servicer able to do it | The platform outlives the chips, so the platform can carry a lease | 2029–30 at the earliest |
| The FCC licensing a constellation on the merits rather than pausing for review | The regulatory path is open at all | Unresolved |
That last row is newer than the rest. On 08/07/2026 a coalition of environmental and scientific groups, represented by Earthjustice, petitioned the FCC for a programmatic environmental review of satellite-based data centres, and asked it to grant no licences until that review is done. The Commission has historically authorised satellites without environmental review and has not weighed cumulative effects. Whatever one makes of the merits, a lessor reading an early-2030s delivery schedule has to price the possibility that the licence arrives years after the hardware is ready, and Earthjustice has said it will sue if it has to.
The second row matters most to a lessor. On the ground, the strongest GPU financing does not rely on the chips at all. CoreWeave's US$8.5bn facility won an A3 rating from Moody's because Meta's contracted payments stand behind it. Orbit will work the same way: the lease is only as good as the capacity contract under it.
The largest players will not need a lessor in any case. SpaceX is raising its own capital and Google is funding Suncatcher from its balance sheet. The leasing opportunity, when it comes, is with the second tier: operators that win a cloud contract but cannot carry a fleet of platforms on their own books. That is where aircraft lessors made their money too.
Orbital compute will not be leased in this decade in the form the filings describe. The long-life platform underneath it probably will be, from the early 2030s, once launch is cheap, the chips can be swapped and someone with an investment-grade rating has signed for the capacity. Until then the financeable part sits on the ground: the gateways and optical ground stations that bring the data down are conventional infrastructure, and we covered how those are financed in ground station financing.
Caelum finances satellites with long lives and creditworthy lessees, and orbital compute meets neither test yet. If and when the platform does, we intend to be there, on a realistic timetable rather than the one in the filings. The airframe-and-engine split is standard practice in aircraft finance, which is where Caelum's structures come from. How a satellite lease works is set out in Space Assets 101, and our approach for investors is on the investor page.
Sources: Google Research, "Exploring a space-based, scalable AI infrastructure system design" (Project Suncatcher), and Google and Planet launch coverage, 10/2026; FCC Space Bureau public notices, including DA-26-113A1; WilmerHale client alert, 02/10/2026; Earthjustice petition on behalf of DarkSky International, Public Employees for Environmental Responsibility and Environment America, 08/07/2026; Starcloud funding announcements, 30/03/2026 and 21/08/2026, TechCrunch 21/08/2026, and the company's own Starcloud-2 description; Voyager Technologies 10-K (Starship price); ARK Invest via Fortune, 10/10/2026 (Falcon 9 cost per kilogram); Statista, 31/07/2026 (capital expenditure); CSO, Data Centres Metered Electricity Consumption 2025, 07/07/2026; CRU decision CRU/2025236; company annual reports and Yahoo Finance interview with Nvidia's chief executive, 21/10/2026 (GPU useful lives); DatacenterDynamics, 31/03/2026 (CoreWeave). Figures in US dollars as reported. Nothing here is legal, tax or investment advice.