
An orbital data center puts AI accelerators on satellites. Solar arrays power them, radiators reject their heat, and laser links connect them to each other and to the ground. The usual orbit is a dawn-dusk sun-synchronous orbit, where the arrays see the sun almost all the time.
The pitch rests on three claims:
The counter-claims are just as concrete. Heat leaves only by radiation, so radiators get large. Nobody can repair the hardware, so satellites are disposable on a five-year life. Radiation corrupts memory and logic.
Google's Suncatcher paper (arXiv 2511.19468) compares the launch cost per kilowatt of satellite power with the cost of ground electricity.
| Launch price to LEO | Launched power cost ($/kW/yr) | Ground power cost ($/kW/yr) |
|---|---|---|
| $3,600/kg (Falcon 9 today) | about 14,700 | 570–3,000 |
| $200/kg (Google's mid-2030s projection) | about 810 | 570–3,000 |
At $200/kg, launch matches the electricity bill. That comparison leaves out the spacecraft itself. Space engineer Andrew McCalip says satellites cost "almost $1,000 a kilo right now" (TechCrunch, 2026-02-11). It also leaves out the five-year life and the cost of radiation errors.
Electricity is only one line of an AI data center's cost. Chips, buildings and networking make up the rest. Jeff Bezos made the same point: energy must become a bigger share of data center cost before orbit pays (CNBC, 2026-05-20). Orbit saves money on the electricity line and adds launch, bus, radiator and reliability costs.
| Source | Orbit vs ground | Parity view |
|---|---|---|
| SemiAnalysis (2026-06-03) | $8.64 vs $2.37 per GPU-hour, about 3.6x. More than 4x on cost of compute. | Base case about 2040. About 30% more expensive by the early 2030s. |
| IEEE Spectrum | About $51B vs $16B for 1 GW, with five years of operating cost | Not given |
| Andrew McCalip calculator | About $42.4B for 1 GW, about 3x ground | Not given |
| Wood Mackenzie (search summary) | More than 3x ground, about $170B for 1 GW | Not given |
| ABI Research | Upward of 78x, a rough TCO estimate | Not given |
| Layer | What it does | Key US-listed names | Who captures the value |
|---|---|---|---|
| Upstream: compute silicon | GPUs, TPUs, CPUs, memory | NVDA, GOOGL (captive TPU), MU, ARM | Merchant chip vendors |
| Upstream: power | Solar cells and arrays, power electronics | RKLB (SolAero), RDW (ROSA), VICR, BA (Spectrolab) | Mostly integrators, which build in-house |
| Upstream: thermal | Radiators, pumped liquid loops | RDW (concept stage) | Integrators. No listed supplier at scale. |
| Upstream: optical links | Laser terminals, intra-cluster optics | RKLB (Mynaric), CACI, HONA, COHR | Defense demand today, not data centers |
| Midstream: launch | Mass to orbit | SPCX, RKLB, FLY | SpaceX alone sets the price |
| Midstream: buses and platforms | Satellite structure, avionics, propulsion | PL, LUNR (Lanteris), YSS, RDW | Buses sold to operators that are not SpaceX |
| Midstream: integrators and operators | Design, own and run the constellation | SPCX, GOOGL | Integrators |
| Downstream: ground and demand | Ground stations, relay, compute buyers | KTOS, VSAT, AI labs, cloud and defense buyers | Not yet a revenue pool |
Orbital data centers use leading-edge commercial chips with screening, error correction and spot shielding. They do not use classic radiation-hardened parts. Starcloud ran its first Nvidia H100 in orbit from November 2025.
By our estimate, 1 GW of orbital compute needs 0.7–1.0 million accelerators. At $30,000–40,000 each, that is $21–40B, the largest single pool in the chain.
Gigawatt-scale orbit needs square kilometers of solar array. Legacy triple-junction space cells cost more than $200 per watt, and their global capacity is in single-digit megawatts a year (commentary sources). The industry answer is commodity silicon or perovskite cells at a few dollars per watt. That removes the pricing power of legacy space-cell makers.
GAO's April 2026 report, Jensen Huang and Starcloud all call cooling the hard problem. The hardware is aluminum or composite panels, fluid loops and pumps. The know-how sits with integrators. Starcloud calls its radiators "core IP" and builds them in-house. We found no US-listed radiator supplier at scale. Redwire is the only listed name, and its work is at concept stage.
SpaceX makes its own Starlink lasers. It also sells Starlink Mini laser terminals to other operators, including Muon Space.
SpaceX (SPCX) listed on Nasdaq on 2026-06-12 at $135 a share. It acquired xAI in February 2026. At the 2026-11-03 close of $141.31, its market cap was $1.92T and its trailing revenue $32.0B.
Every third-party operator therefore buys launch from its largest competitor.
The Q3 release had no orbital compute line. On the same call, management said it was "fairly confident" of 10 GW of new ground compute in 2027. That is about ten times the orbital target. The FCC has not ruled on SpaceX's filing for up to one million satellites. Amazon has petitioned the FCC to deny it.
The downstream layer has no revenue pool yet. Kratos (KTOS) virtualizes ground-station functions and could serve data center constellations. Viasat (VSAT) says it sees a role in communications links but has no contract. Optical relay through Starlink is the more likely path.
Demand comes from three groups:
| Publisher | Figure | Our view |
|---|---|---|
| SpaceX S-1 | $28.5T total addressable market | Whole company. Says nothing specific about orbital compute. |
| BIS Research | $1.77B (2029) to $39.09B (2035) | Paid report, seen only in summaries. Method not shown. |
| Strategic Revenue Insights | $645M (2030), $3.47B (2035) | Paid report, seen only in summaries. Method not shown. |
| ABI Research | Up to 18,600 nodes and about 1.5 GW by 2035 | Capacity, not revenue |
| Musk | 1 GW/yr by end-2027, about 100 GW/yr about 3.5 years later | Needs weekly Starship flights and a new factory |
Our realistic view for 2026–2027 is tens of millions of dollars a year. Most of it is demonstration and R&D money. External orbital compute services could reach $0.1–2B a year by 2030, mostly defense and edge work. If SpaceX flies at scale, most of the value stays inside SpaceX. It then shows up as SpaceX capital spending and as Nvidia and supplier revenue.
Prices are 2026-11-03 closes from stockanalysis.com. EV/Sales is trailing twelve months. Exposure columns are our estimates. No company discloses orbital compute revenue.
| Ticker | Layer | Mkt cap | TTM revenue | TTM growth | EV/Sales | Exposure today | Exposure 2030 | Dated tie to orbital compute |
|---|---|---|---|---|---|---|---|---|
| SPCX | Launch, operator | $1.92T | $32.0B | +128.6% | 58.3x | 0% | 3–10% | S-1 "as early as 2028" |
| NVDA | Compute | $4.99T | $303.0B | +83.4% | 16.4x | <0.1% | Not estimated | Space-1 module (2026-03-16) |
| GOOGL | Captive TPU | $3.44T | $471.1B | +22.2% | 7.1x | ~0% | Not estimated | Suncatcher in orbit (2026-10-01) |
| MU | Memory | $967.5B | $133.2B | +256.3% | 6.8x | <0.1% | Not estimated | None. HBM is the weak link. |
| RKLB | Solar, lasers, buses, launch | $32.6B | $769M | +52.5% | 39.4x | <1–3% | 2–8% | Data center solar arrays (early 2026) |
| COHR | Intra-cluster optics | $40.9B | $7.12B | +22.5% | 6.0x | <0.1% | Not estimated | None |
| VICR | Power conversion | $8.26B | $561M | +41.4% | 13.8x | <1% | Not estimated | None |
| CACI | Laser terminals | $15.1B | $9.75B | +10.1% | 2.1x | ~0% | Not estimated | SDA demand, not data centers |
| PL | Bus partner for Google | $4.92B | $378M | +44.1% | 12.0x | 0% booked as revenue | 3–10% | Suncatcher prototype |
| FLY | Launch, host vehicle | $2.85B | $287M | +179.2% | 7.9x | <1% | 1–5% | Starcloud lunar payload (2026-09-30) |
| LUNR | High-power buses | $2.40B | $490M | +116.9% | 5.1x | <1% | 2–6% | Stated target (2026-02-25) |
| RDW | Power, thermal, structures | $2.02B | $426M | +63.1% | 3.6x | <1% | 3–10% | Sophia Space MOU (2026-09-30) |
| VOYG | Station host | $1.52B | $223M | +41.8% | 8.3x | <1% | 1–5% | ISS cloud demonstration |
| YSS | LEO buses | $0.84B | $405M | +59.7% | 1.2x | 0% | 1–5% | None |
| ECHO | SpaceX equity proxy | $29.7B | Distorted | n.m. | n.m. | 0% operating | 0% operating | Owed SPCX shares |
Notes:
| Bull | Bear |
|---|---|
| Ground power, permits and interconnection are the binding AI constraint. Orbit avoids them. | Full-system models put orbit at about 3x to more than 4x ground cost today. ABI's rough estimate is 78x or more. |
| SpaceX controls launch, laser relay, chip co-design and an in-house customer in xAI. It needs only internal parity. | Google's $200/kg parity covers electricity only. Spacecraft cost, five-year life and radiation losses are missing. |
| TPUs survived about 3x their expected dose. An H100 trained a model in orbit. | Google's prototype can power a TPU for only about 15–20 minutes at a time. |
| FCC Chair Brendan Carr publicly dismissed Amazon's petition against SpaceX. The FCC has not ruled. | Schedules slip. Starcloud-2 moved from October 2026 to January 2027, and Neutron to 2027. |
| Nvidia, Google and Cisco are investing capital. | Altman and Bezos both doubt near-term timelines. Astronomers and Amazon oppose the filings. |
| Date | Event |
|---|---|
| 2026-11-04 to 11-12 | Q3 results: RDW (11-04), KRMN (11-05), RKLB (11-09), FLY (11-10), LUNR and YSS (11-12) |
| No earlier than 2026-11-12 | Starship Flight 15, booster and ship catch attempt |
| 2026-11-15 | Comment period closes on China's proposed perovskite technology export ban |
| 2026-11-17 | Nvidia FQ3 FY27 results. Watch for a Space-1 ship date. |
| About December 2026 | First SPCX 180-day lock-up releases (estimate from reported terms) |
| 2026-12-09 | Planet Labs results |
| January 2027 | Starcloud-2 launch, the first test of a powered radiator at about 100x Starcloud-1 power |
| 2027 | First Starmind satellites (SpaceX Q3 call). Google's two-satellite laser link test. |
| Q4 2027 | Space-optimized Vera Rubin NVL72 launch (Musk) |
| 2027-11-30 | Target date for EchoStar's SPCX share delivery |
| 2028 | SpaceX S-1 deployment start "as early as 2028". Falcon 9 rideshare booking window ends. |
| Not scheduled | FCC decisions on the SpaceX, Blue Origin and Starcloud filings |
Main sources:
This report draws on company filings, earnings calls, public market data, research papers and media coverage available on 2026-11-04. It is for information only. It is not investment advice or a recommendation to buy or sell any security. Readers should conduct their own due diligence and consult a licensed financial adviser before making any investment decision. Past performance does not indicate future results.