
- What Did SpaceX Starship Flight 15 Actually Achieve?
- Why Would SpaceX Put AI Data Centers in Orbit?
- How Many Starship Launches Would An Orbital Data Center Need?
- Could the Electricity Savings Pay for the Launch Bill?
- What Could Still Prevent Orbital AI Data Centres from Working?
- How Could Orbital Data Centres Create Value for SpaceX Stock?
- What Would Orbital Computing Need to Earn to Move SPCX Shares?
- Which Starship Milestones Would Make the Investment Case Stronger?
Imagine building an expensive delivery aircraft, sending it around the world and throwing it away after unloading the cargo. Every delivery would have to pay for another aircraft.
Recovering it changes the calculation. But only if you can inspect it, repair it cheaply and send it out again.
That is the business story inside the latest SpaceX Starship rocket launch. After deploying 26 Starlink V3 satellites, Starship's upper stage returned from an orbital flight and was caught by the giant arms of its launch tower. It was the first time SpaceX had brought a Starship upper stage back this way.
For India, the Starship catch happened at about 2:50 am IST on November 15, after a flight lasting roughly 22.5 hours. It is an extraordinary recovery, although the rocket's Super Heavy booster was lost after its own catch attempt was aborted.
The bigger question is what happens if this eventually becomes routine. Could SpaceX use a fleet of reusable rockets to put power-hungry AI data centres above the electricity grids constraining them on Earth?
Flight 15 strengthens one necessary part of that argument. Whether orbital data centres become a valuable business still depends on how much useful computing SpaceX can deliver for each kilogram it launches, how long that hardware earns money and how much cash remains after replacing it.
What Did SpaceX Starship Flight 15 Actually Achieve?
The mission launched from Starbase, Texas, on November 13 in the US, or early November 14 in India. Ship 42, the upper stage, completed approximately 14 orbits before returning to Starbase and being caught by the tower commonly called Mechazilla.
This was Starship's first upper-stage tower catch after an orbital flight. Its previous mission, Flight 14, had already reached orbit and deployed satellites.
The difference between the two rocket stages also matters. The booster provides the initial push and returns much earlier. The upper stage travels into orbit, then has to survive atmospheric reentry and arrive at the tower in a condition that permits a controlled catch.
Flight 15 recovered that upper stage. It did not recover both stages, and it did not demonstrate that the recovered ship can fly again with little refurbishment. Those are the next tests of the commercial argument.
After inspecting Ship 42, Elon Musk estimated that it could probably fly again after a few weeks of refurbishment. He also acknowledged that further fixes are needed to achieve immediate reuse with only scheduled maintenance. The condition of the recovered ship is encouraging; its actual turnaround time and reflight performance remain to be demonstrated.
There was an immediate business purpose, too. SpaceX says each of the 26 Starlink V3 satellites adds approximately one terabit per second of network capacity. Together, the batch represents about ten Falcon 9 launches carrying V2 Mini satellites in added downlink capacity.
That comparison concerns communications capacity, rather than ten times the payload mass or ten times the revenue. Starlink V3 also remains a communications satellite; deploying it does not demonstrate a working Starmind AI data centre.
Still, this gives Starship a useful stepping stone. SpaceX can improve the rocket while deploying infrastructure for an existing business, before asking it to support a new one.
Why Would SpaceX Put AI Data Centers in Orbit?
An AI data centre needs more than chips. It needs somewhere to put them, electricity to run them and equipment to remove the heat they produce. Building the servers can be faster than securing the power connection that makes them useful.
The International Energy Agency's 2026 analysis projects global data-centre electricity consumption of about 945 terawatt-hours in 2030, against 485 TWh in 2025. That is its base case, rather than a guaranteed outcome, but it explains why companies are looking for alternatives to conventional expansion.
Space offers a tempting possibility: satellites with large solar arrays, computers and radiators, connected through laser links. Carefully chosen orbits can provide near-continuous sunlight, reducing dependence on batteries and removing the need for a terrestrial grid connection.
SpaceX has already taken concrete steps towards that architecture. Its January application to the US Federal Communications Commission sought authority for an orbital data-centre system of up to one million satellites, at altitudes ranging from 500 km to 2,000 km. The FCC accepted the application for review and public comment; that February notice was not permission to deploy the entire proposed constellation.
There is also a named computing platform. In August, Nvidia confirmed that SpaceXAI's planned first-generation Starmind satellite would use an optimised Vera Rubin NVL72 system. The companies are adapting that architecture for orbital conditions.
Musk's August announcement set a target of launching that space-optimised system in the fourth quarter of 2027, with substantial expansion in 2028.
SpaceX's advantage is the combination: rockets, satellite manufacturing, a communications network and an AI business that could use the computing capacity. It can work on several parts of the system together. It still has to make the complete system cheaper or more useful than the alternatives.
How Many Starship Launches Would An Orbital Data Center Need?
Let’s start with one gigawatt, or one million kilowatts, of computing power. This describes how much electrical power the computers can use, rather than how much AI work they necessarily complete.
SpaceX's prospectus sets out a design assumption of approximately 100 kilowatts of compute power per metric tonne, alongside a Starship payload assumption of 100 tonnes. Both are planning parameters; Flight 15 did not validate that complete orbital-computing configuration.
Using those assumptions, each kilowatt requires about 10 kg of deployed hardware. One gigawatt therefore requires approximately 10,000 tonnes in orbit, or 100 launches carrying 100 tonnes each.
| Computing capacity deployed | Hardware at 10 kg per kW | Launches at 100 tonnes per flight |
| 1 GW | 10,000 tonnes | 100 |
| 10 GW | 100,000 tonnes | 1,000 |
| 100 GW | 1 million tonnes | 10,000 |
Source: SpaceX's June 2026 prospectus for the design assumptions; INDmoney calculations. These are idealised deployment estimates. They exclude replacement launches, failed missions and competing demands on the launch fleet. Payload available to a particular orbit can differ from the assumed capacity.
The table explains why catching a ship is relevant. SpaceX's ambition to deploy 100 GW annually would, under these assumptions, require about 27 launches every day. The rocket, satellites, launch sites and ground operations would all have to support that pace.
It also exposes a less obvious risk. If the complete satellite needs 20 kg per kilowatt instead of 10 kg, the transport requirement doubles. Heavier radiators, larger solar arrays or extra protection can offset part of the benefit from cheaper launches.
The important measure is therefore the cost of delivering a working kilowatt to its required orbit. A rocket's advertised cost per kilogram tells only part of that story.
Could the Electricity Savings Pay for the Launch Bill?
Now let’s give that gigawatt an electricity bill.
For a transparent illustration, assume the computers run continuously and terrestrial electricity costs $0.10 per kWh. That is close to the EIA's preliminary US industrial average of 9.64 cents in August 2026, although an individual data centre's contract can be very different.
One gigawatt running for a year consumes 8.76 billion kWh. At ten cents each, the bill is $876 million a year, or $4.38 billion over five years, before discounting future spending. This covers the computers' electricity; it excludes additional facility power for cooling and other equipment.
At the assumed 10 kg per kilowatt, the initial transport bill looks like this:
| Assumed effective launch cost | Launch bill for 1 GW | Compared with the $4.38 billion five-year IT electricity bill |
| $100 per kg | $1 billion | Transport consumes about 23% |
| $200 per kg | $2 billion | Transport consumes about 46% |
| $500 per kg | $5 billion | Transport already exceeds it |
Source: INDmoney calculations using the stated operating, mass and electricity assumptions. These are sensitivities, rather than observed Starship launch costs or SpaceX guidance. Effective launch cost should include the economic cost of maintaining the launch system.
This is why a successful catch cannot establish a profitable orbital data centre. At $500 per kg, getting the hardware upstairs already costs more than five years of the IT electricity it could avoid at this tariff. At $200 per kg, there is more room, but solar arrays, radiators, spacecraft systems, financing and failures still need to be paid for.
Put another way, $438 per kg is the transport-only ceiling supported by this five-year electricity comparison. At 20 kg per kilowatt, it falls to $219. Neither is a commercial break-even price: both leave nothing for the additional orbital equipment and ignore the cost of money.
Orbital computing could also avoid some building and cooling expenses, or create value by bringing capacity online sooner than a delayed ground project. Those benefits have to be measured alongside its extra costs. The sunlight itself does not arrive with an electricity bill, but collecting it reliably requires paid-for hardware.
Google's independent Project Suncatcher research offers a useful reference point. Its modelling suggests launch prices could fall below $200 per kg by the mid-2030s, under a sustained cost-learning assumption. At that level, it sees potential for launch and operating costs to approach terrestrial energy costs on a per-kilowatt-year basis.
That supports taking the idea seriously. It also shows how much cost reduction the argument requires, and why a projected future price should not be treated as today's demonstrated economics.
What Could Still Prevent Orbital AI Data Centres from Working?
The remaining problems interact. Solving one by adding more equipment can make another more expensive.
Cooling still requires equipment
On Earth, air or water can carry heat away from computing equipment. In a vacuum, there is no surrounding air to do that job. NASA's spacecraft thermal guidance explains that heat exchange with the external environment relies on radiation.
The heat therefore has to travel from the chips to surfaces that radiate it into space. Cooling loops, radiators, coatings and their orientation become part of the design. A cold environment does not automatically provide cheap heat removal.
Larger radiators can help control temperature, but they also add mass and deployment complexity. That feeds directly back into the kilograms-per-kilowatt assumption driving the launch bill.
A computer that works in orbit must also earn its replacement cost
Running advanced AI hardware in space is already possible on a small scale. Starcloud's CEO has described operating an H100 in orbit, training a small nanoGPT model and running Google's Gemma model aboard Starcloud-1.
The harder commercial question is how much useful work the hardware completes over its economic life. Radiation-related errors, component failures and interruptions can reduce output. Hardware can also become less competitive as newer chips produce more work for each dollar and watt.
SpaceX's prospectus says it does not anticipate servicing or repairing processors in space. The business therefore needs to recover its investment while allowing for failures and eventual replacement. A satellite surviving for five years does not establish that its original processors remain commercially attractive for all five.
Different AI workloads need different networks
Training a large model can require many processors to exchange data extremely quickly. Running an already trained model, known as inference, can offer workloads that are easier to divide between independent machines, although some inference systems also need tightly connected clusters.
Google identifies inter-satellite links supporting tens of terabits per second as a requirement for performance comparable to terrestrial data centres. Maintaining those links between moving spacecraft adds another task.
Connections to Earth create a separate constraint. Customer data still has to reach the computers and results have to get back, with enough bandwidth and acceptable response times for the service being offered.
An early orbital business may therefore find customers in selected inference jobs or processing data already collected by satellites. Success there would be meaningful, while leaving the much larger task of reproducing a terrestrial training cluster's performance unresolved.
The comparison keeps changing on Earth
SpaceX would be competing against improving chips, better cooling systems and new terrestrial power capacity. Its orbital system must compete with the ground alternatives available when it launches, including their construction schedules.
Permissions and orbital operations matter as well. A proposed constellation has to secure the applicable authorisations, manage collision risks and provide credible disposal arrangements. The FCC's acceptance of an application does not settle those questions.
This makes routine, reliable launch capacity necessary. It also makes it insufficient on its own.
How Could Orbital Data Centres Create Value for SpaceX Stock?
There are three plausible routes, with different economics.
First, SpaceX could launch computing satellites for outside customers. It would earn transport revenue, while those customers funded and operated the computers. This offers exposure to the industry's growth without owning every deployed processor.
Second, SpaceX could own computing satellites and charge outside customers for their output. The potential revenue pool becomes larger, alongside the investment, utilisation risk and replacement bill.
Third, it could use orbital computing within its own AI business. The benefit could appear as lower costs or faster access to capacity. An internal payment from the AI business to the launch business would not create new revenue for the consolidated company.
That last point is easy to overlook. One internally used satellite cannot simultaneously create a full outside launch sale and a second outside computing sale. The shareholder benefit must ultimately come from external revenue, lower costs or a better return on capital.
SpaceX already operates substantial businesses against which that benefit must be measured. Its Q3 results reported revenue of $14.575 billion, including $5.248 billion from Connectivity and $7.560 billion from Super Intelligence. The company also reported 1.7 GW of nameplate computing capacity at September 30.
Those existing AI revenues are not evidence of orbital revenue. Equally, dividing the entire Super Intelligence segment's sales by its computing capacity would not produce a reliable orbital revenue-per-gigawatt forecast: that segment includes software and other activities.
The funding requirement deserves equal attention. Across the first nine months of 2026, SpaceX generated $13.097 billion in operating cash flow and spent $48.069 billion on property, plant and equipment. Subtracting those two lines gives a cash shortfall of about $35 billion, before other investing and financing items.
This is company-wide spending. It is not an orbital data-centre budget, and existing financing helps fund it. But it shows why an attractive future margin cannot substitute for examining the cash needed to build the business.
What Would Orbital Computing Need to Earn to Move SPCX Shares?
At its November 13 US close, SpaceX stock was at $139.11. The latest quarterly filing reported approximately 13.58 billion shares across Classes A and B as of October 15. Together, those inputs imply an equity value of about $1.89 trillion.
Against that starting point, an additional $100 billion of equity value represents about $7.36 per share, or 5.3%. A successful new business needs considerable scale to materially change the value of a company already this large.
Rather than invent future chip prices and customer contracts, we can work backwards. Assume investors value the orbital business in 2032, roughly six years away, at 20 times its sustainable annual incremental free cash flow to equity. Discount that future value by 12% annually.
Both the multiple and discount rate are illustrative choices. Free cash flow here means cash available to shareholders after operating expenses, taxes, debt service and the investment needed to maintain competitive computing capacity. It also means additional cash compared with what SpaceX could earn without the orbital project.
| Illustrative added equity value today | Dollar value added | Equivalent per existing share | Annual incremental free cash flow needed in 2032 |
| 10% of the starting valuation | $188.9 billion | $13.91 | $18.6 billion |
| 25% of the starting valuation | $472.3 billion | $34.78 | $46.6 billion |
| 50% of the starting valuation | $944.6 billion | $69.56 | $93.2 billion |
Sources: SpaceX Q3 2026 Form 10-Q for shares outstanding; Nasdaq Dorsey Wright; INDmoney calculations. Formula: required 2032 annual cash flow = added equity value today × 1.12⁶ ÷ 20. The table assumes no net investment outlay before 2032 and no additional dilution, making it a favourable starting hurdle. These are valuation sensitivities, rather than price targets or forecasts.
The first row is already demanding. Even 10 GW running continuously would avoid only $8.76 billion a year in IT electricity costs at our assumed tariff, before paying for the orbital system. Electricity savings at that scale, on their own, would fall short of the table's $18.6 billion annual free-cash-flow hurdle.
A larger contribution would therefore need some combination of more capacity, profitable external computing demand, avoided construction costs and valuable access to power sooner. It would also need those benefits to survive competition and recurring hardware replacement.
Construction spending raises the hurdle further. As a sensitivity, every $50 billion of net investment in today's money before 2032 adds approximately $4.93 billion to the annual 2032 cash flow required under this model. That $50 billion is an illustrative input, rather than a SpaceX budget. Future share issuance would spread the value across more shares.
There is downside as well. A failed programme with a $50 billion net present cash cost would consume about $3.68 per existing share, before any reduction in the market's expectations for future profits. Conversely, a credible path to commercial success could affect the stock well before the cash arrives.
One final constraint applies to every upside calculation: some expectation of orbital success may already be included in today's price. An estimated value for the orbital business cannot simply be added in full to SpaceX's current valuation. The stock responds to how the outlook changes relative to what investors already expect.
Which Starship Milestones Would Make the Investment Case Stronger?
The next useful evidence should connect the engineering achievement to its cost and earning power.
Watch whether SpaceX can recover both stages, refly the upper stage and disclose how much work that takes. Then watch the payload it can repeatedly deliver to the required orbit, the time between flights and the full economic cost of maintaining that service.
For Starmind, the important evidence will be useful computing output sustained in orbit, measured power and cooling performance, reliable links and a credible hardware replacement cycle. Commercial contracts and returns on the cash invested would carry more weight than a larger satellite count alone.
Flight 15 makes the transport case more credible. Starlink provides a nearer-term use for that progress, while orbital AI offers a larger and much less proven opportunity.
The tower catch is a first step towards recovering an expensive asset. The business breakthrough comes when SpaceX can send it out again cheaply, deliver computers that work reliably and charge enough for their output to cover the entire cycle. That is the distance still left between catching Starship and making orbital data centres valuable to shareholders.