Part IThe energy wall is real. The relief valve is unproven.
Begin with the demand, because it is genuine. Terrestrial AI data centers have run into a power and thermal ceiling that is now a macro constraint rather than a facilities problem. Global data-center electricity consumption is projected to roughly double by 2030, and on longer-dated forecasts to approach a tenth of all electricity consumed on Earth by mid-century. The near-term North American build alone implies somewhere between 50 and 100 gigawatts of new capacity inside three years, the rough energy equivalent of 50 to 100 nuclear stations standing up in a single capital cycle. The International Energy Agency places global data-center consumption near 1,000 terawatt-hours in 2026, comparable to the entire national draw of Japan.
That is the bull setup, and our house position is that it does not, by itself, make orbital compute investable. Space genuinely removes three terrestrial constraints at once: land, water, and permitting. It offers near-continuous solar exposure and a 2.7-kelvin sink to radiate into. But in exchange it substitutes four harder physics-and-cost problems, and the entire category is, at root, a derivative bet on a single variable, launch cost per kilogram, which we treat in Part II. The disciplined way to read this sector is to walk into the stack rather than into the headline. The satellite, the thing every explainer counts, is the wrapper. The value is underneath it.
Part IIThe stack: four layers, one bottleneck
State the thesis plainly. An orbital data center is four subsystems in a trench coat. The bus that carries them is integration work, not differentiation, and the AI silicon inside is a commodity catalog part. The investable questions are therefore not “who launches the most satellites,” but “who owns the scarce enabling layer,” and “how many vendors exist to supply it.” We walk each layer below, then render the full vendor map in the Supply-Chain Matrix.
Layer 1, Power (deployable solar at kilometer scale). A gigawatt-class orbital node needs photovoltaic arrays measured in square kilometers, launched folded and unfurled on orbit, then sun-tracking continuously. The materials race is silicon versus perovskite. Nanyang Technological University in Singapore, under the national Space Technology Development Programme, is testing roll-to-roll perovskite cells printed as a chemical ink and crystallized in vacuum: cheaper, lighter, and flexible enough to roll for launch. On the public side this maps cleanly to Redwire’s ROSA roll-out arrays and to Rocket Lab’s SolAero space-grade cells. Vendor density: moderate and rising.
Layer 2, Thermal (heat rejection in vacuum). No air and no water means every watt of compute heat leaves the system by one mechanism only: radiation off a physical radiator. At constellation scale this is the hardest unsolved subsystem in the stack, and it is the one with no public pure-play vendor. Primes and the compute operators (Starcloud, SpaceX) are solving it in-house, which is itself the tell. This is our most differentiated call in this brief: whoever vendorizes deployable radiators at scale collects a toll on every orbital compute node ever built, and that vendor does not yet trade. Vendor density: effectively zero. Highest structural opportunity in the category.
Layer 3, Optical interconnect (laser ISL and downlink). Inter-satellite and satellite-to-ground laser links carry on the order of a thousand times the bandwidth of radio frequency, and in vacuum there is no weather to attenuate them. This is the connective tissue of any orbital compute fabric, and it is the layer that just consolidated. Rocket Lab’s completed Mynaric acquisition folds the CONDOR optical-terminal line, already flying on Rocket Lab’s $1.3B Space Development Agency satellite contracts, into a public launch-and-systems prime. Singapore’s Transcelestial is pushing ground-station laser downlink and a roughly 40-satellite equatorial constellation it calls The Ring. Vendor density: consolidating into public hands.
Layer 4, Launch (cost per kilogram, the master variable). Everything above is gated here. Partially reusable Falcon 9 breakeven sits near $500 per kilogram. Fully reusable Starship targets $10 to $20 per kilogram, a 25-to-50x reduction that, if it holds durably, is the only thing that clears orbital compute against terrestrial on total cost. If it does not hold, this is a 2035-and-beyond story, not a fundable-now one. Adjacent capacity comes from Rocket Lab Neutron, Firefly’s MLV, Blue Origin New Glenn, Stoke Space, and Relativity’s Terran R. Vendor density: high, but concentrated in the economics of a single vehicle.
Analyst estimate, pending reconciliation with the live platform recompute. Capital-flow and deployment signals are accelerating, but maturity and concentration penalties hold the composite out of the 80s. Index values (SERD / CDI / CFI) are left as placeholders for the live dashboard recompute and are not asserted here.
Part IIIThe compute layer is already commoditizing
The satellites themselves are where every camera is pointed and where the least durable advantage lives. Starcloud (rebranded from Lumen Orbit) put a single Nvidia H100 in orbit on Starcloud-1 and has filed with the FCC for 88,000 satellites. Axiom Space is developing orbital data-center nodes (its AxDCU units). Lonestar Data Holdings is targeting lunar data storage. OrbitsEdge and Ramon.Space supply space-grade edge compute. The hyperscalers have entered on paper: Google’s Project Suncatcher, SpaceX’s roughly one-million-satellite AI constellation, Blue Origin’s Project Sunrise, and Amazon Leo’s edge ambitions. In every one of these, the chip is Nvidia’s and the differentiation is integration of the four subsystems above, not the silicon. That is precisely why our trade is the stack, not the satellite operator: the operator is buying commodity compute and renting scarce enabling hardware, and the scarcity sits upstream of the operator.
Part IVThe sovereign axis: why China’s structure matters more than its count
China’s edge-computing-first approach, processing satellite data in orbit and downlinking only the results, is a precursor to full orbital data centers rather than the thing itself. But its structure is the story. The Three-Body Computing Constellation, led by Zhejiang Lab with ADA Space (Guoxing Aerospace, Chengdu), launched its first 12 of a planned 2,800 satellites on 14 May 2025, targets 100 by 2027, and at full scale aims for 1,000 peta-operations per second across 2,400 inference and 400 training satellites, linked by 100-gigabit-per-second optical inter-satellite links. It is ITU-coordinated and backed commercially by Alibaba, Kepu Cloud, and iSoftStone. Unlike the US path, which is privately led and gated on launch economics, China’s is funded, mandated, and written into national policy. The European Union runs its own publicly funded feasibility track in Thales Alenia Space’s ASCEND study. The control-of-the-next-internet stakes are not rhetorical: Starcloud (88,000), SpaceX (~1,000,000), and Blue Origin (~51,600) together have placed more than 1.14 million orbital-data-center satellites before the FCC, against roughly 15,000 active satellites in orbit today. The regulatory queue alone is two orders of magnitude larger than the deployed base.
Part VOED assessment
Lead with the counter-thesis, because it is strong. The cleanest case against the entire category is that it is a launch-cost mirage: absent durable sub-$50-per-kilogram Starship economics, orbital compute never beats terrestrial on total cost, and the bull case collapses into a pure derivative bet on one company’s reusability cadence. We take that objection seriously, and our resolution is an asymmetry that makes the sector fundable today regardless of how the objection resolves. The four enabling subsystems (optical, solar, launch, and eventually thermal) get paid whether or not the orbital-data-center thesis fully matures, because the same hardware already serves proliferated-LEO, Space Development Agency, and commercial-constellation demand that exists right now. You are not underwriting the moonshot. You are underwriting the chokepoints that the moonshot, and everything adjacent to it, must pass through.
The satellite is the commodity. The toll booths are the trade. Three of the four are public; the fourth, thermal, is the largest unpriced position in the category.
Three actionable reads follow from that. First, the cleanest public expression of the subsystem-roll-up thesis is now Rocket Lab post-Mynaric: launch, optical inter-satellite links, and space solar inside one entity. Second, the unpriced layer is thermal, where no public pure-play exists; watch for a financing round, an M&A move, or an in-house Starcloud or SpaceX radiator program that could spin out. Third, treat the satellite operators (Starcloud and the hyperscaler filings) as the demand signal, not the trade itself.
| Layer | Constraint | Public vendors | Adjacent private / foreign | Density |
|---|---|---|---|---|
| Compute | Radiation-tolerant, thermally bounded AI silicon | NVDA | Starcloud (ex-Lumen Orbit), Axiom Space (AxDCU), Lonestar Data Holdings, OrbitsEdge, Ramon.Space, Ubotica, Little Place Labs, Exo-Space | Moderate |
| Power | km-scale deployable, sun-tracking solar | Redwire (RDW, ROSA), Rocket Lab (RKLB, SolAero) | Solestial, mPower Technology, Aetherflux, Star Catcher, NTU Singapore (perovskite, STDP) | Moderate |
| Thermal | Heat rejection by radiator in vacuum | None (pure-play) | Advanced Cooling Technologies, Boyd; in-house at Starcloud / SpaceX; prime IP | Gap |
| Optical | >1,000x-RF laser ISL and downlink | Rocket Lab (RKLB, Mynaric/CONDOR), CACI (CACI, SA Photonics) | Transcelestial (The Ring), Kepler Communications, Skyloom, Aalyria, BridgeComm, Warpspace, Tesat-Spacecom (Airbus) | Consolidating |
| Launch | Cost per kilogram to orbit | SpaceX (private), Rocket Lab (RKLB, Neutron), Firefly (FLY), Blue Origin (private) | Stoke Space, Relativity Space (Terran R) | High, concentrated |
| Sovereign / edge | State-funded orbital compute | None (state programs) | ADA Space / Guoxing + Zhejiang Lab (Three-Body), Thales Alenia (ASCEND, EU) | Policy-led (CN/EU) |
Density chips are OED vendor-density reads, not investment ratings: green = high vendor count, amber = moderate, red = structural supply gap, blue = policy-led / non-commercial. Public tickers denote listed exposure; private and foreign names denote the broader supplier field tracked in the OED database.
Public-market exposure cards
Eight listed names with direct, moderate, or indirect exposure to the orbital compute stack. The exposure-confidence tag reflects how load-bearing the read-through is, not a price view.
Card footnotes: 1 Starcloud-1 H100 in-orbit demonstration, late 2025. 2 Rocket Lab 8-K and press release on the Mynaric close, 14 April 2026 ($155.3M aggregate consideration). Cards are exposure maps, not recommendations.
Watch list
- STARSHIP $/KG. Cost-per-kilogram prints and launch cadence. The master variable that gates the entire category.
- THERMAL GAP. Any deployable-radiator vendor financing, prime spinout, or M&A. The unpriced layer.
- THREE-BODY. Next launch batch and the 100-satellite-by-2027 milestone.
- RKLB x MYNARIC. Integration progress and CONDOR optical-terminal volume ramp.
- FCC DISPOSITIONS. Rulings on the more than 1.14 million proposed orbital-DC satellites.