What could SpaceX’s orbital data centers be worth? This cockpit compares the cost of AI computing in orbit with data centers on the ground, turns the saving into a value today, and compares it with SpaceX’s share price. Move the sliders to test your own view. The sections further down show what matters most, how wide the range of outcomes is, and how the model works. Built on the Laniakea Research orbital data center report.
This interactive cockpit models SpaceX's orbital data center (ODC) business as it could look in 2040, once its space-based AI compute constellation is fully deployed. It estimates the annual cost savings of orbital versus terrestrial data centers in 2040, values that business with a terminal P/E multiple (applied after tax), and discounts the result back to today as a Present Value (PV). Use the sliders to stress-test the assumptions, see which ones move the valuation most in the Tornado chart, and compare the PV with SpaceX's current market capitalization.
How much of SpaceX’s share price the Laniakea base case accounts for. The shading runs from dark green at the orbital data center line to dark red where the share price is 50% of the price or more above it: 0% · 25% · 50%+. Context, not a forecast.
How much of SpaceX’s value should orbital data centers explain?
News tied to your assumptions. Labels describe what it suggests for the assumption, not for the share price. Tap an assumption to jump to its slider.
The space-based AI compute opportunity of $1.22T (at current model settings) represents approximately 63.4% of SpaceX's public market valuation of $1.925T.
Each bar shows how far the value today moves when one assumption is set lower or higher, with everything else unchanged. Red means a lower value, green a higher value. The longer the bar, the more that assumption matters.
Impact of changing each variable to its Low / High value (others at base) on PV Valuation ($ Trillions). The dashed line marks the current SPCX market cap. Red shows a lower value and green a higher value than the current case.
Nobody knows the exact numbers, so the model runs 10,000 scenarios, each with slightly different assumptions around your settings. The chart shows how often each value came up: red bars are below today’s market value, green bars above it. Switch on the Advanced view to see and change the ranges.
Illustrative only. 10,000 simulated outcomes. Each draw picks the inputs from skewed ranges around your current slider settings, keeps related inputs consistent (more flights, lower launch cost), and treats the discount rate, the tax rate and the share of the saving SpaceX keeps as fixed at your slider values, because it already prices in execution risk. The spread reflects the assumed ranges (shown in the Advanced view), not data. The chart's right edge is the 99th percentile; the last bar collects everything above it.
The value built up step by step, from one satellite to the whole fleet.
A quick look at what the cockpit can do.
Only one possible business: SpaceX running AI data centers in orbit, on satellites launched by Starship. It estimates what that business could be worth today once the full network is in place.
Starlink, launch services and everything else SpaceX does are left out. So the value shown is not a value for SpaceX as a whole, and the cockpit never says what SpaceX shares "should" cost.
Mainly for power. On Earth, electricity is 50% to 70% of a data center's running costs, and new data centers wait years for grid connections and permits. In a sun-synchronous orbit, solar panels are in sunlight 95% to 99% of the time (against 15% to 25% for solar panels on the ground), and the power is produced right where it is used.
Space also removes the need for land (over $8 million per acre in Northern Virginia, the largest data-center market), water for cooling and property taxes. In the base case, a megawatt of computing in orbit costs about $1.05 million a year (satellite, launch and operations), against about $3.3 million for the same computing in a data center on the ground.
Only with a fully and rapidly reusable Starship (version 4), which is not yet a reality. Today a Falcon 9 launch costs SpaceX about $1,400 per kg internally, and an expendable Starship test flight works out at about $20,680 per kg.
The cost falls the way airline costs did once planes were flown thousands of times. A Starship v4 carries about 150 tonnes to the sun-synchronous orbit used here. In the 2040 end state, with thousands of flights a year, a flight costs roughly $4 million:
Wear on the vehicle: about $1 million, if a ship costing about $100 million flies around 100 times. Propellant: about $1.3 million, once SpaceX produces its own. Refurbishment and ground operations: about $1 million. Fixed programme costs: about $1 million, because about $3 billion a year is spread over around 3,000 flights.
About $4 million for 150 tonnes is about $27 per kg. Getting there requires catching and reusing both the booster and the ship many times, including the heat shield, which has not yet been demonstrated. If you expect reuse to be slower or more costly, set a higher launch cost in the Standard view: the simulation tests $20 to $100 per kg.
Then orbital data centers would very likely not pay off. With the base-case satellite, computing in orbit only becomes cheaper than on the ground once launch costs fall below roughly $650 per kg. Falcon 9 at about $1,400 per kg is far too expensive. Full reusability is the single most important condition behind this business, which is why launch cost has its own slider.
The $0.56 million covers the satellite itself: structure, solar panels, radiators, power and communication systems. For a satellite of about 3.3 tonnes, that is about $170 per kg, which assumes SpaceX builds these satellites on a production line at very high volume, as it already does with Starlink. An AI satellite needs no user-facing antennas or direct-to-phone hardware, so per kilogram it should be simpler to build than a Starlink satellite.
The AI chips are not included, on purpose: the same chips would be needed in a data center on the ground, so they cancel out in the comparison. Likewise, the ground data-center costs exclude the computers themselves. The simulation tests about $0.28 million to $1.11 million per satellite.
Heat is carried by liquid to large radiator panels that send it out into space as infrared light. The AI1 satellite pairs 250 kW of peak power with about 840 m² of solar panels and about 160 m² of double-sided radiators. The radiators run at about 75°C, with the chips at about 90°C. They must be kept edge-on to the Sun and Earth so they do not soak up extra heat.
Google tested its Trillium TPU chips for its Project Suncatcher and found no permanent failures up to 15 krad of radiation. A satellite in this orbit is expected to receive about 750 rad over five years, about 20 times less. The chips' high-bandwidth memory showed minor errors from 2 krad, which still leaves a margin of about three times.
They are not repaired: each satellite is built to work reliably for its life and is then replaced. Five years matches how often AI chips are replaced in ground data centers anyway, because newer chips are much faster.
A longer life spreads the satellite and launch cost over more years and leaves more satellites in orbit at the same launch rate: try the lifetime slider in the Standard view (3 to 8 years).
Through laser links to the Starlink network, which then connects to ground stations. Each AI satellite needs far more bandwidth than a Starlink satellite (the aim is about 1 terabit per second), and more ground stations will likely be needed. This is one of the open engineering questions behind the business.
Each Starship flight carries 45 AI1 satellites (150 tonnes at about 3.3 tonnes each). At 3,200 flights a year and a 5-year life, about 720,000 satellites are in orbit at any time, delivering about 180 GW of computing power, roughly the output of 180 large power plants.
The base case uses 3,200 flights because it is a less aggressive cadence than SpaceX's own plan. SpaceX's plan of 1 million satellites (about 4,400 flights a year) is the Believer case, and 2,000 flights a year is the Skeptic case.
The flight rate also stands for demand: about 180 GW of AI computing in orbit would be more than all data centers on Earth are expected to use around 2030, so 3,200 flights a year assume AI demand keeps growing strongly for another decade. If it slows, fewer AI satellites are launched.
The model compares the yearly cost of computing in orbit with the same computing in ground data centers, and treats the difference as the business's profit. That yearly saving in the end year (2040 in the base case), grown with 2% inflation a year from 2026 prices, minus 21% tax, times a P/E multiple of 35 (as in our robotaxi cockpit), gives the value in 2040, which is discounted back to today.
The discount rate is built the same way as in our robotaxi cockpit: the 10-year US Treasury yield plus beta times a 5% market risk premium, plus a 5% extra premium for the risks of the project. SpaceX has traded only since June 2026, too short a history for a reliable beta, so we use the median beta of nine listed peers in space and AI infrastructure (Rocket Lab, AST SpaceMobile, Planet Labs, Iridium, CoreWeave, Nvidia, Oracle, Digital Realty and Equinix), updated weekly. The higher the rate, the lower the value today. The business inputs can be changed in the Standard view, the rate in the Advanced view.
In the base case, yes: SpaceX sells computing in orbit at the price of computing on the ground and keeps the whole saving as profit. We see no other company with both fully reusable rockets and satellite production at this scale within the next 10 to 15 years, so the price ceiling is set by data centers on the ground, whose costs we hold flat to 2040 although land, grid connections and power are getting scarcer. If you expect customers to bargain for part of the saving, or ground computing to get much cheaper, lower the "Share of the saving SpaceX keeps" slider: the value moves one for one with it.
A P/E of 35 is high for a mature infrastructure business. We use it, as in our robotaxi cockpit, because it also stands for the growth businesses that cheap, frequent launches would make possible beyond AI computing, which are not modelled separately. The simulation tests about 23x to 47x.
It is not in the value. The model leaves AI chips out on both sides, so it compares only the data center around the chips: building, power, cooling and launch. SpaceX and Tesla plan a chip factory in Texas, Terafab, with Intel as a partner. It is at an early stage: a new chip factory usually takes three to five years to build, and SpaceX's IPO filing reportedly warns that it may not become commercially viable.
If it works, chips made in-house instead of bought from suppliers such as Nvidia would add a second saving per megawatt on top of the one modelled here, growing with the same fleet. Those chips would save just as much in data centers on the ground, so we would value them as a separate chip business rather than add them to this cockpit.
Nobody knows the exact outcome. The "How sure can anyone be?" section runs thousands of scenarios, each with slightly different assumptions around your settings, and shows the range that 8 in 10 of them fall into. Our base case is our view, not the truth: the range shows how wide the reasonable outcomes are.
No. The cockpit is a scenario tool built on stated assumptions, not a forecast, a price target or a recommendation. It is meant to help you see what your own view means in money terms.
Laniakea Research holds positions in Tesla and SpaceX and may benefit if their share prices move. We show this next to every result so you can weigh our base case with that interest in mind.
When you change the assumptions, we record your settings and the resulting value anonymously, with the date and your country, to improve the cockpits. We do not record your name, IP address or any cookie or device ID, and nothing is recorded if your browser sends a Do Not Track or Global Privacy Control signal. Details are in our privacy policy.
These answers use the figures and assumptions in this cockpit. For background on the business, see our research. Not investment advice.