Starlink and Orbital Data Centres: How SpaceX Plans to Move AI Compute Into Space

Starlink started as a way to make money from SpaceX's reusable rockets. By June 2026 it was the company's largest business, with 12.0 million subscribers and $4.3 billion of quarterly revenue. SpaceX now wants to use the same satellites, lasers and launch system for something new: data centres in orbit that run AI workloads on solar power.

This primer has two parts. Part A explains how Starlink works: the satellites, the lasers that connect them, the dishes on the ground, and the physics that makes it faster than older satellite internet. Part B covers SpaceX's orbital compute plan: what it has filed with regulators, the engineering problems it has to solve, and the launch-cost arithmetic that decides whether the idea pays.

Sources are SpaceX's IPO prospectus (Form S-1/A) and Q2 2026 Form 10-Q, Starlink's 2025 Progress Report, two Federal Communications Commission (FCC) documents, and Google's Project Suncatcher research paper. Calculations are ours unless stated, with the inputs shown. For SpaceX's segment revenue and profit, see our SpaceX business breakdown.

---

Part A. How Starlink Works

1. Four Pieces of Hardware

A Starlink connection passes through four kinds of equipment.

Sources: 10-Q for the quarter ended June 30, 2026; Starlink 2025 Progress Report.

1.1 The Dish

The Starlink dish is a phased-array antenna. A traditional satellite dish is a curved reflector that has to point at one fixed satellite. A phased array is a flat panel of many small antenna elements. By adjusting the timing of the signal at each element, it can steer its beam electronically in any direction within milliseconds, with no motor.

This matters because Starlink satellites are always moving. A satellite in low orbit crosses the sky in minutes, so the dish has to hand over from one satellite to the next many times a minute. The Progress Report says each terminal also builds a map of obstructions such as trees and buildings, and picks satellites with a clear line of sight. SpaceX's prospectus describes it as the first company to make consumer-grade phased-array terminals at scale, starting in 2022.

1.2 The Satellites

Starlink uses a "V" naming system for satellite generations. The prospectus lists these:

Sources: S-1/A, "Our Satellite Names"; Q2 2026 earnings release; Starlink 2025 Progress Report.

SpaceX does not publish a V2 Mini's mass or power. Google's Suncatcher paper cites an optimised V2 Mini at about 575 kg with roughly 105 square metres of solar panel. Assuming 22% panel efficiency, it estimates about 28 kilowatts of power per satellite. We use that estimate again in Part B.

SpaceX depreciates broadband satellites over five years and first-generation mobile satellites over three. That is the period over which their cost is charged against profit. The prospectus says it often deorbits satellites before then, bringing them down to burn up in the atmosphere, to reduce the risk of a failing satellite it cannot steer. Replacing the fleet is therefore a permanent cost of the business. SpaceX's satellite factories in Redmond and Woodinville, Washington, delivered more than 70 satellites a week in 2025, according to the Progress Report.

1.3 The Laser Mesh

Each V2 Mini carries three laser links to neighbouring satellites. Traffic can hop from satellite to satellite across oceans or remote regions, then come down at a gateway thousands of kilometres from the user. During the April 2025 power outage in Spain and Portugal, the Progress Report says, Starlink routed traffic over lasers to gateways elsewhere in Europe.

The 2025 report states the mesh had more than 24,000 lasers. Faster link acquisition cut latency by 30 to 40 milliseconds in Asia and Africa, and a hardware upgrade was planned for 2026 to allow 400 gigabits per second per link. SpaceX's FCC filing for orbital data centres describes the mesh as "high capacity (petabit)". This laser network is the part of Starlink that orbital compute depends on most directly.

2. Why Low Orbit Is Faster

Older satellite internet uses geostationary satellites 35,786 km above the equator. At that height a satellite takes exactly one day to orbit, so it appears fixed in the sky and a dish can point at it permanently. The cost is distance. Every request goes up and down, and every reply goes up and down again.

Latency is the delay between sending a request and receiving the first byte of the answer. Light travels at about 299,792 km per second. The minimum round-trip delay from the satellite hop alone is:

Our calculation: four times altitude ÷ speed of light. This assumes the satellite is directly overhead and ignores ground-network and processing time, so real latency is higher in every case.

The geostationary delay is fixed by physics and no engineering can remove it. For Starlink, most of the measured delay comes from routing and ground networks, not altitude. The Progress Report gives median global latency of about 26 ms in 2025, against a stated goal of a stable 20 ms. Median download speeds were above 200 Mbps.

The trade-off is coverage. A geostationary satellite can see about 42% of Earth's surface (our calculation from its altitude). A satellite at 480 km sees a small patch and moves across it in minutes. Continuous global service therefore needs thousands of satellites instead of three, and that is only affordable with cheap launches.

2.1 Where the Satellites Fly

In January 2026 the FCC granted a partial approval (order DA 26-36) of SpaceX's "Gen2 upgrade" applications. The order:

• authorised another 7,500 Gen2 satellites, taking the approved Gen2 constellation to 15,000
• approved new orbital shells at 340, 345, 350, 355 and 365 km, with up to 144 satellites in each of up to 72 planes
• required SpaceX to lower Gen2 satellites from 525 to 535 km down to 475 to 485 km

SpaceX had asked for up to 29,988 Gen2 satellites. The FCC noted that lower altitudes carry less collision risk, because a failed satellite falls back into the atmosphere sooner. Separately, the prospectus says Starlink accounted for about 75% of all active, manoeuvrable satellites in orbit at March 31, 2026, and made more than 1,000 automated collision-avoidance manoeuvres a day in 2025.

3. Starlink as a Business, Briefly

Source: SpaceX Q2 2026 Form 10-Q.

Two features of the economics carry into Part B. First, Starlink does not pay for its launches in cash. SpaceX launches its own satellites and adds the cost to the satellites' book value, so launch cost appears later as depreciation. Second, the five-year satellite life means that cost comes back every five years. The full segment analysis is in the business breakdown.

---

Part B. Data Centres in Orbit

4. The Problem Orbital Compute Is Meant to Solve

On Earth, the limit on building AI data centres is increasingly electricity, not chips. A new campus needs hundreds of megawatts, and the queue for a grid connection in many regions runs for years. We cover this in detail in the AI energy bottleneck and in who builds AI power infrastructure.

SpaceX's prospectus states its answer directly: "The logical path forward is to move power-intensive AI workloads into orbit, where solar energy is near-constant and uninterrupted."

The specific orbit matters. A sun-synchronous orbit passes over the poles and slowly rotates so that it keeps the same angle to the Sun all year. A "dawn-dusk" version follows the line between day and night, so a satellite in it is almost never in Earth's shadow. Google's researchers write that in such an orbit a solar panel receives "up to 8× more solar energy per year than a panel located on Earth at mid-latitude." There is no night, no cloud and no atmosphere, and little battery storage is needed.

5. What SpaceX Has Committed To

5.1 The FCC Application

On January 30, 2026, SpaceX applied to the FCC for the "SpaceX Orbital Data Center system" (ICFS File No. SAT-LOA-20260108-00016). The FCC's public notice (DA 26-113, February 4, 2026) summarises it:

• Up to one million satellites, at altitudes of 500 to 2,000 km
• Orbits at 30-degree and sun-synchronous inclinations, in shells up to 50 km thick, with different hardware versions for different shells
• Communications mainly by optical inter-satellite links, connecting to each other and to the existing Starlink Gen1 and Gen2 satellites, which relay traffic to the ground
• Waivers requested from the FCC's deployment milestone rules and its surety-bond rules. Normally a licensee must deploy half its constellation within six years and post a bond as a guarantee.

SpaceX describes the system in the application as the "first step towards becoming a Kardashev II-level civilization," a reference to a hypothetical scale in which a Type II civilisation uses all the energy of its star.

5.2 The Prospectus

The S-1/A sets out targets and timing:

The prospectus also ties part of Elon Musk's pay to the project. 302,072,285 restricted Class B shares vest only if SpaceX meets market-value milestones and completes "non-Earth-based data centers capable of delivering 100 terawatts of compute per year."

5.3 Checking the Headline Numbers

The prospectus figures can be checked against each other and against SpaceX's current launch rate.

• Mass. 100 GW at 100 kW per ton is 100,000,000 kW ÷ 100 kW = 1,000,000 tons a year, which matches the prospectus.
• Launches. Starship V3 is designed to carry 100 tons to orbit fully reusable, so that is 10,000 Starship launches a year, about 27 a day.
• Against today. In the first half of 2026, SpaceX put 1,041 metric tons into orbit across 78 launches. Annualised, that is about 2,080 tons. The 100 GW goal is about 480 times SpaceX's current mass to orbit.
• Power. 100 GW running all year is 100 × 8,760 hours = 876 terawatt-hours. The prospectus compares this with US electricity production of 4.4 thousand TWh in 2025: 876 ÷ 4,400 = 19.9%, the "approximately one-fifth" it cites.

6. The Engineering Problems

6.1 Getting Rid of Heat

A chip turns almost all the electricity it uses into heat. On Earth, data centres remove that heat with air, water and chillers. In a vacuum there is no air to carry heat away. The only way out is to radiate it as infrared light from a surface, which is what the prospectus means by "leveraging the space environment for cooling."

How much heat a surface can radiate depends steeply on its temperature (the Stefan-Boltzmann law: power per square metre equals emissivity × a physical constant × temperature to the fourth power). Using an emissivity of 0.9, typical of radiator coatings, and counting both faces of a flat panel:

Our calculation. It is a best case: it ignores sunlight and Earth's infrared falling on the radiator, and the temperature lost between chip and radiator surface. Real radiators would need to be larger.

For comparison, the solar panels for 1 GW would cover about 3.7 km², assuming 1,361 W per m² of sunlight, 22% efficiency and 90% panel coverage (the same inputs Google uses). An orbital data centre is mostly panels and radiators, and the chips are a small part of its mass. Google's paper calls thermal management "a critical optimization challenge for power-dense TPUs operating in a vacuum."

6.2 Radiation

Outside the atmosphere, chips are hit by charged particles. Two effects matter. Accumulated dose (total ionising dose) degrades a chip slowly. Single particle strikes can flip a bit in memory or crash a processor.

Google tested its Trillium TPU chips in a proton beam. It estimated a five-year dose in its planned orbit of about 750 rad(Si). The chips survived far more than that "without permanent failures". The high-bandwidth memory showed irregularities only after about 2,000 rad(Si). Bit flips remained, which means designs must detect and correct errors in software. SpaceX's prospectus says its satellites use "radiation-hardened components, shielding, and redundancy" but that "these measures may not be sufficient."

6.3 No Repairs

On Earth, a failed GPU is swapped by a technician. In orbit that is impossible. Google's paper says "the simplest solution is redundant provisioning," meaning launching spare capacity. The prospectus lists the same risk. Once deployed, orbital compute "will not be easily repaired or upgraded," and failures "could result in permanent capacity loss."

The prospectus adds that satellite life is "inherently shorter than that of the information technology systems and infrastructure they host." Starlink satellites are written off over five years. AI chips on Earth are commonly depreciated over five to six years, but their useful economic life is being debated because each new chip generation is faster. Our data centre economics piece covers that debate. In orbit, the chip and the satellite reach the end of their lives together.

6.4 Getting Data Down

A data centre in orbit is only useful if results reach users on the ground. SpaceX plans to route traffic from compute satellites over the Starlink laser mesh to existing gateways. That is why it states Starlink is "another crucial enabler of orbital AI compute."

This suits some workloads better than others. Inference, running a trained model to answer a query, sends little data in and out per request and is the workload SpaceX names first. Training a large model needs thousands of chips exchanging data constantly at very high speed. Google's design handles that by flying 81 satellites within a 1 km radius so the laser links between them can reach about 10 Tbps. The fastest demonstrated space-to-ground optical link Google cites is NASA's TBIRD mission, at 200 Gbps in 2023.

6.5 Regulation and Debris

One million satellites would be about 100 times the size of today's Starlink fleet. The prospectus says the plan depends on "spectrum authorizations, orbital debris mitigation approvals" and international coordination, and "there can be no assurance that such approvals will be obtained." The FCC has taken comments on SpaceX's application. Its request for waivers of the deployment milestones and surety bond is part of what regulators must decide.

7. The Cost Arithmetic

The comparison that decides orbital compute is launch cost against the cost of electricity. At SpaceX's target of 100 watts per kilogram, a 1 GW orbital data centre weighs 10,000 tons. The table prices that launch at different costs per kilogram and spreads it over a five-year satellite life.

Our calculation: 10,000,000 kg × price per kg, divided by 5. Launch only; the satellites and chips cost extra.

At today's price, the annual launch cost ($7.2 billion) is about seven times the electricity it replaces ($1.0 billion). At $200 per kilogram it falls to $0.4 billion, below the terrestrial power bill. Google's paper reaches a similar conclusion by a different route. It expresses launch as a "launched power price," the cost per kilowatt per year of power put in orbit. For a Starlink-type satellite that price falls from $14,700/kW/year at current launch prices to about $810/kW/year at $200/kg. US data centres pay about $570 to $3,000/kW/year for electricity.

Google estimates when $200 per kilogram might arrive. SpaceX's own price history shows a learning rate of about 20%: the price per kilogram has fallen about 20% each time cumulative mass launched doubled. If that continues, the paper projects prices "could fall to less than $200/kg by ∼2035," which it says would need about 180 Starship launches a year. Starship has not yet delivered a payload to a lasting orbit. Flight 13 in July 2026 released 20 V3 satellites on a suborbital path, and SpaceX expects orbital payload delivery to begin in the second half of 2026.

The comparison leaves out the largest terrestrial cost, hardware. Chips cost the same on Earth and in orbit, plus the added cost of solar panels, radiators and a satellite body. The case for orbital compute is not that it is cheaper per chip. It is that it could add capacity without waiting years for grid connections, and SpaceX's prospectus argues it would get newer chips running before competitors for that reason.

The prospectus lists the risks to this case itself. If "advances in nuclear energy" or other terrestrial sources cut energy costs or ease grid constraints, "the viability of our orbital AI compute infrastructure may be materially diminished." It also warns that SpaceX "may prioritize our own launch payloads" over government and commercial customers to meet its orbital compute goals.

8. Who Else Is Working on It

Sources: Google Research blog and arXiv:2511.19468; IEEE Spectrum; FCC DA 26-113.

Starcloud-1 launched on a SpaceX rocket. Whoever builds orbital compute has to buy launch capacity, and SpaceX is the largest seller of that capacity as well as a competitor. Google's paper names Blue Origin as a potential future competitor on launch price. The chips inside all three projects come from the suppliers covered in who competes with NVIDIA.

9. What to Watch

1. Starship payload flights. Every figure in section 7 depends on Starship carrying 100 tons at a much lower cost per kilogram. The first orbital payload delivery is expected in the second half of 2026.
2. The FCC decision on the one-million-satellite application, and whether the milestone and bond waivers are granted.
3. The first compute satellite. SpaceX says "as early as 2028." Google's Planet prototypes are due in early 2027.
4. Watts per kilogram. Any disclosed compute satellite mass and power, measured against the 100 kW-per-ton target.
5. Terrestrial power. Grid connection waits and new nuclear capacity on Earth, which SpaceX names as the main threat to the orbital case.

---

Key Takeaways

• Starlink combines phased-array dishes, more than 10,200 satellites, a laser mesh of over 24,000 lasers and a gateway network. Under the FCC's January 2026 order, its Gen2 satellites fly at 340 to 485 km.
• Low orbit cuts the minimum round-trip delay from 477 ms (geostationary) to about 4.5 to 6.4 ms. Starlink's measured median latency was about 26 ms in 2025.
• SpaceX has applied to the FCC for up to one million orbital data centre satellites at 500 to 2,000 km, and expects to start deploying them as early as 2028.
• The stated goal of 100 GW a year at 100 kW per ton means one million tons and about 10,000 Starship launches a year, about 480 times SpaceX's current mass to orbit.
• A gigawatt of orbital compute needs roughly 3.7 km² of solar panels and at least 0.65 to 1.2 km² of radiators. Heat, radiation and the lack of repairs are the main engineering problems.
• At today's roughly $3,600 per kilogram, launching 1 GW costs about $7.2 billion a year over five years, against about $1.0 billion a year of electricity on Earth. At $200 per kilogram, launch becomes cheaper than that electricity.
• SpaceX's own prospectus lists the risks: the technology is untested, satellites wear out faster than chips, and cheaper terrestrial power, including nuclear, could undermine the case.

---

Disclaimer

This article is for educational purposes. The author holds no position in SpaceX (SPCX). It is not investment advice. Please do your own research and consult a SEBI-registered investment adviser, or a registered adviser in your jurisdiction, before making investment decisions. Calculated figures are identified as such in the text.

---

Sources

1. Space Exploration Technologies Corp. Registration Statement on Form S-1/A (Amendment No. 2), June 2026. sec.gov. Orbital AI compute targets, satellite generations and lives, risk factors, executive compensation.
2. Space Exploration Technologies Corp. Form 10-Q for the quarter ended June 30, 2026, filed August 4, 2026. sec.gov.
3. Space Exploration Technologies Corp. SpaceX Reports Second Quarter 2026 Results, August 4, 2026. sec.gov.
4. Starlink. Progress 2025. starlink.com. Laser mesh, latency and speeds, gateways, production rates, Direct to Cell.
5. Federal Communications Commission, Space Bureau. Space Bureau Accepts for Filing SpaceX's Application for Orbital Data Centers, DA 26-113, February 4, 2026. docs.fcc.gov.
6. Federal Communications Commission, Space Bureau. SpaceX Gen2 Upgrade Applications, Partial Grant, DA 26-36, January 9, 2026. docs.fcc.gov.
7. Agüera y Arcas, B., Beals, T., Biggs, M., Bloom, J. V., et al. (Google). Towards a future space-based, highly scalable AI infrastructure system design, arXiv:2511.19468v2, June 2026. arxiv.org.
8. Google Research. Exploring a space-based, scalable AI infrastructure system design, November 4, 2025. research.google.
9. IEEE Spectrum. Nvidia H100 in space (Starcloud-1 launch report), November 2025. spectrum.ieee.org.