SpaceX Innovation Explained: Space Before Reusable Rockets and the Technology Behind Orbital Data Centres

For most of the space age, reaching orbit meant throwing away a rocket that cost tens or hundreds of millions of dollars. SpaceX's central bet was that a rocket could fly back, be refuelled and fly again, the way an aircraft does. Whether that bet fully works is now the hinge of its plan to run AI data centres in orbit, because that plan needs launch at a scale and price no expendable rocket could approach.

Before SpaceX, a launch was closer to a government project than a product. Most American rockets were built by large defence contractors under cost-plus contracts, where the customer reimburses the contractor's costs and adds an agreed profit margin. That gives the contractor little reason to cut costs, because a saving reduces the reimbursement instead of adding to profit. A fixed-price contract works the other way: the customer agrees a sum in advance, and the contractor keeps any saving or absorbs any overrun.

Nearly every rocket was also expendable, meaning it flew once and its engines, tanks and computers fell into the ocean or burned up. The exception was the Space Shuttle, which flew 135 times between 1981 and 2011.

The Shuttle was reusable, but reuse did not make it cheap. NASA analyst Harry Jones, in a 2018 paper on launch costs, concluded that reuse had not so far cut the price of reaching orbit. The Shuttle's design raised development costs, and the savings did not cover the cost of refurbishing it between flights. Some sources give a lower figure for one more flight, between $409 million and $450 million in 2010 and 2011, but that leaves out the fixed cost of keeping the programme alive. When the Shuttle retired in July 2011, the United States had no way to launch its own astronauts. That lasted until May 30, 2020, when Crew Dragon flew the first crewed orbital launch from American soil in nine years.

A market where the United States was a small player

The US Department of Commerce counted 74 orbital launches worldwide in 2010, of which 23 were commercial, meaning paid for by a private customer. Russian launch providers, mostly on the Proton rocket, flew 13 of them, 57% of the market. Europe's Arianespace flew 6. The United States flew 4, two on Boeing's Delta rockets and two on SpaceX's first Falcon 9 flights. The same report found that prices for medium and heavy rockets had risen from about $50 million to nearly $100 million in the previous four years.

Inside the United States, the two big rocket makers, Boeing and Lockheed Martin, merged their launch businesses in December 2006 into a joint venture called United Launch Alliance (ULA). Its Atlas V rocket used a Russian engine, the RD-180, which became a political problem after Russia annexed Crimea in 2014. Congress later barred the military from buying launches that use it after 2022.

Cost figures from this period are disputed. In 2013 SpaceX claimed ULA's launches cost about $460 million each, and ULA answered that its average price was $225 million. In 2015 testimony to Congress, SpaceX's president Gwynne Shotwell cited a Government Accountability Office report showing the unit cost of the military's launch programme rising from $101.7 million to $376.4 million per launch, a 3.7 times increase, while the number of planned missions fell from 181 to 163. That is SpaceX's account of an audit, given while it was competing for the same contracts, so treat it as an argument. The direction it describes, fewer launches at a higher unit cost, is consistent with the incentives of cost-plus contracting.

Satellites built for rare launches

Expensive, infrequent launches shaped the satellites too. A 2020 industry article gives a typical example of a communications satellite in geostationary orbit (a high orbit 35,786 km up, where a satellite stays over one spot on the equator). It weighs 2,500 kg, costs $100 million or more to build, lasts a planned 15 years, and needs about $22 million of insurance for launch and another $22 million for its time in orbit. Launching to that orbit cost up to $30,000 per kilogram. Launch and launch insurance came to about $95 million of a $300 million lifetime cost, roughly a third (95 ÷ 300).

When a launch is a rare event and a failed satellite cannot be repaired, the sensible strategy is to build few satellites, make each one large and bespoke, and test it to exhaustion. That was the industry SpaceX entered.

SpaceX was founded on March 14, 2002 with the stated aims of cutting the cost of spaceflight and making access to space more reliable. Its first three Falcon 1 launches, between 2006 and 2008, all failed and nearly bankrupted the company. The fourth, on September 28, 2008, made it the first private company to put a liquid-fuelled rocket into orbit.

Building most of the rocket in-house

The prospectus credits SpaceX's "extensive vertical integration and end-to-end control over the entire value chain, from design to launch to operations" for its speed and cost. Vertical integration means a company makes its own parts instead of buying them. Jones's paper reports SpaceX's own estimate that subcontracting one dollar of in-house work would cost three to five dollars. The reasoning is the one behind any economic moat built on cost: a supplier adds a margin, a delay and a communication gap at every hand-off, and a company that owns the whole chain can redesign across those boundaries.

Selling for a fixed price

NASA's crew transport programme gives a clean comparison. Its Inspector General estimated in 2019 that NASA would pay about $55 million per seat on SpaceX's Crew Dragon and about $90 million per seat on Boeing's Starliner, both under fixed-price contracts, against an average of $55.4 million per seat on Russian Soyuz rockets. The contracts told the same story: SpaceX's was worth $2.5 billion against Boeing's $4.3 billion, and SpaceX's development and test-flight cost was $1.2 billion against Boeing's $2.2 billion.

Source: NASA Office of Inspector General, report IG-20-005 (November 2019). Percentage differences are our calculation.

Under a fixed price, the risk of overruns sits with the company. SpaceX's lunar lander contract with NASA, worth $2.89 billion, is also fixed-price.

Landing the first stage, and flying it again

Reuse is the change that matters most. A Falcon 9 has two stages. The first stage does the heavy lifting off the pad and carries nine of the rocket's ten Merlin engines, so throwing it away discards 90% of the engines (9 ÷ 10). SpaceX's answer was to fly the first stage back and land it upright on legs or on a ship at sea. It did that for the first time in December 2015 and reflew a landed stage in March 2017. The prospectus states that Falcon 9 first stages have flown 34 times, and by mid-2026 the rocket had flown 650 missions with a 99.5% success rate. Since 2023, SpaceX says it has launched more than 80% of the world's mass to orbit each year.

The scale of the problem explains why nobody had done it. On the pad a Falcon 9 weighs 549,000 kg. Its payload to low orbit is 22,800 kg if the first stage is thrown away, and 17,500 kg if it is saved for landing, which is 4.2% and 3.2% of liftoff mass (22,800 ÷ 549,000 and 17,500 ÷ 549,000). The first stage alone carries about 411,000 kg of propellant, 75% of the total weight. Everything else, including the stage that gets saved, is the small remainder. Saving a stage costs payload, because it must carry extra fuel to slow down and land.

What reuse did to price. The story here has a twist. If reuse had passed savings on to customers, list prices should have fallen. They have not.

Sources: NASA (2018 list price and payload); Wikipedia's Falcon 9 page (2026 commercial price, payloads and third-party estimates of internal cost). SpaceX does not disclose its internal cost per launch, and the prospectus contains no such figure.

The 2018 figure matches Jones's $2.7 thousand per kilogram. In nominal dollars the list price has risen 19% since 2018 (74 ÷ 62). The savings from reuse appear to have gone into SpaceX's margins and into launching its own satellites, not into lower prices for outside customers. This matters for the orbital data centre argument, because what decides the outcome is SpaceX's own cost per kilogram, and the list price says little about that. Other analyses use a current cost near $1,400 to $1,800 per kilogram (SemiAnalysis) or $1,500 (Forethought), which sit close to the internal-cost range above and well below the list price.

Giving itself a customer that never stops buying

A rocket that flies every few days needs a customer that buys just as often. SpaceX built one: itself. In the second quarter of 2026, 28 of its 38 launches carried its own Starlink satellites (see the business breakdown). Starlink also turned the satellite model upside down. Where a geostationary satellite is one large machine planned to last 15 years, Starlink satellites are built in large numbers, launched in batches and replaced on a cycle of about five years, as the primer explains. Cheap launch made that design rational.

Raptor and Starship, the second generation

Starship is SpaceX's attempt to make the whole vehicle reusable, and to make it much bigger. The Block 3 version stands 124.4 metres tall and 9 metres wide, is built from welded stainless steel about 4 mm thick, and burns liquid methane and liquid oxygen. Its prospectus says Starship V3 is designed to deliver 100 tonnes to orbit fully reusable, against 17.5 tonnes for a Falcon 9 that saves its booster (a factor of 5.7).

The engine is the Raptor. Rocket engines use turbine-driven pumps to force propellant into the combustion chamber. In simpler designs the turbines are driven by a side stream of burning propellant that is then exhausted, wasting some of it. In staged combustion that gas goes on into the main chamber. In full-flow staged combustion, the design Raptor uses, all of both propellants pass through the turbines before reaching the chamber. Before Raptor only two such designs had reached a test stand, and none had flown. Raptor 3 produces about 250 tonnes of thrust at sea level at a chamber pressure of 330 bar, against 185 tonnes and 250 bar for the first version. Methane was chosen partly because it burns cleaner than kerosene and could be manufactured on Mars from carbon dioxide and water.

Sources: SpaceX Form S-1/A; Wikipedia pages for Falcon 9, Falcon 9 Block 5, SpaceX Raptor and SpaceX Starship.

SpaceX's prospectus sets a target of 100 gigawatts of orbital compute launched each year. It then states what that needs: "thousands of launches per year and the transport of approximately one million metric tons to orbit annually," and adds that "the fully reusable nature of Starship positions us to be capable of launching this level of mass." The primer works through the power, heat and radiation problems. This section asks a narrower question: of everything that has to work, which piece decides whether the plan can be paid for?

Launch cost is the gate

Two independent analyses land in the same place. SemiAnalysis estimates that launch is about $1.6 million of a $3.1 million infrastructure cost in its model, and that launch costs must fall from roughly $1,400 to $1,800 per kilogram to about $250 for orbital compute to make sense. Forethought puts the threshold for overall cost competitiveness at about $100 per kilogram, with space solar power matching Earth-based off-grid solar at $250. Both name Starship as the vehicle on track to get there. Forethought adds that it would need 10 to 25 reuses per vehicle.

They disagree on timing. SemiAnalysis projects cost parity around 2040 in its base case and the early 2030s in a scenario where power on Earth runs short. Forethought considers orbital data centres unlikely to be a meaningful share of computing before 2030. Neither treats launch cost as sufficient. Chip production, chips that cannot be repaired once in orbit, and bandwidth to the ground are their other constraints.

So the answer to the question is fully and rapidly reusable heavy launch. Concretely, that means four things working together on Starship: the Raptor engine that makes the vehicle efficient and rebuildable, the tower catch that returns the booster without landing legs, a ship that survives re-entry with a heat shield that can be inspected and flown again quickly, and enough production volume to fly thousands of times a year.

What has been shown and what has not

Sources: SpaceX Form S-1/A; Wikipedia pages for Starship, Flight 13 and Flight 14; SpaceX plans are as reported before Flight 14 and may change.

The main open question is the heat shield. Starship's re-entry protection is a skin of roughly 18,000 hexagonal tiles, and reports after Flight 13 described cracked tiles and signs of hot gas leaking through seams. Three former NASA experts, including Dan Rasky, who co-invented the material used on Dragon's heat shield, argued in July that a tile system needs individual inspection and repair, which they say cannot support turnarounds of hours, the same limit that held back the Shuttle. Rasky called the current design "a dead-end for all missions that require full and rapid reusability." SpaceX has not announced a replacement and says it is iterating on tile shape, attachment and materials. We have not independently checked the tile count or the damage reports, which come from a trade press account.

How much cost reduction is enough?

The prospectus says Starship is meant to "reduce the cost to reach orbit by 99% or more relative to the historical average launch cost." It does not say what that average is. The answer changes what the target means, as this table shows using the thresholds above.

Our calculation: a 99% cut leaves 1% of the baseline. Cut needed = 1 - target ÷ baseline. For example, 250 ÷ 61,700 = 0.4%, so a cut of 99.6% is needed.

The other pieces around it

Reusable launch is necessary but not the only requirement. Compute satellites would send results to the ground over the Starlink laser network described in the primer, and Starlink's factories already build satellites at scale. Two limits sit outside SpaceX's control. Semiconductor production is the constraint SemiAnalysis expects to bind through about 2034, and the chip designs still have to survive radiation and go years without repair. Meanwhile, how much orbital compute is worth depends on how tight power is on Earth, the subject of the AI energy bottleneck and the data centre economics analysis. Google, an early SpaceX investor, and Anthropic have both signed deals to rent computing capacity in SpaceX's ground-based data centres in Memphis, so demand for compute is not the doubt. The doubt is whether the launch economics get there.

Announcements are cheap and flight results are not. These five checkpoints turn the claim into events that can be observed.

The fifth checkpoint is the first compute satellite itself, which the prospectus places "as early as 2028."

Practice exercises

Key Takeaways

• Before SpaceX, rockets were expendable and largely bought on cost-plus terms. The Shuttle averaged about $1.45 billion per flight over 135 flights, and in 2010 only 4 of 23 commercial launches worldwide were American.
• SpaceX combined three changes: it built most of the rocket in-house, sold at fixed prices (NASA's estimate was $55 million per Crew Dragon seat against $90 million for Starliner), and landed and reflew the first stage. Falcon 9 first stages have flown 34 times.
• The savings show up more in SpaceX's margins than in its prices. Falcon 9's list price rose from $62 million in 2018 to $74 million in 2026, while third-party estimates put its internal cost at $15 million to $28 million.
• Starlink gave SpaceX a customer for its own rockets. In Q2 2026, 28 of its 38 launches carried its own satellites.
• The technology orbital data centres depend on most is fully and rapidly reusable heavy launch, meaning Starship: Raptor engines, tower catch and a ship that can be reflown quickly.
• Booster reuse is proven. Ship reuse is not. No Starship has yet reached orbit, been caught or been reflown, and critics say the tile heat shield may prevent fast turnaround.
• A 99% cost cut is enough only from some baselines. From Shuttle-era cost it leaves $617 per kilogram, above the $100 to $250 that independent analysts say is needed. Their timelines for parity run from the early 2030s to about 2040.

Disclaimer

This article is for educational purposes and 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. Starship schedules and plans were current on September 21, 2026 and may have changed since.

Sources

1. Jones, H. The Recent Large Reduction in Space Launch Cost, NASA Technical Reports Server, 2018. ntrs.nasa.gov. Cost per kilogram by vehicle, SpaceX vertical integration estimate.
2. NASA Office of Inspector General. NASA's Management of the Commercial Crew Program, Report IG-20-005, November 2019. oig.nasa.gov. Price per seat and contract values.
3. Space Exploration Technologies Corp. Registration Statement on Form S-1/A, June 2026. sec.gov. Starship design claims, 99% cost target, launch share, reuse record, orbital compute mass.
4. Shotwell, G. Statement to the House Armed Services Committee subcommittee, March 17, 2015. docs.house.gov. SpaceX's account of military launch costs. An interested party's testimony.
5. US Department of Commerce. Commercial Space 2011 (launch market data for 2010). trade.gov. 2010 launch counts, market shares and price trend.
6. Withington, T. Space on Budget, Armada International, March 24, 2020. armadainternational.com. Cost structure of a geostationary satellite.
7. Wikipedia: Space Shuttle program, United Launch Alliance, SpaceX, Falcon 9, Falcon 9 Block 5, SpaceX Raptor, SpaceX Starship, Starship flight test 13, Starship flight test 14. Secondary references for dates, specifications and third-party cost estimates.
8. SemiAnalysis. To Boldly Go: The Case for Space Datacenters. newsletter.semianalysis.com.
9. Forethought. Will We Really Put Data Centers in Space? forethought.org.
10. Tech Times. Starship Heat Shield Tiles Are a Dead End for Rapid Reuse, Former NASA Experts Warn, July 29, 2026. techtimes.com.
11. TechCrunch. SpaceX doubles revenue on Anthropic and Google compute deals, Starlink growth, August 4, 2026. techcrunch.com.
12. TeslaNorth. SpaceX Details Starship Flight 14 First Orbit and 26 Starlink V3s, September 15, 2026. teslanorth.com. Flight 14 plan, subject to change.