Launching a kilogram to orbit costs 96% less than it did in 1960, and economists expect the fall to continue. That collapse is real and measured. But satellite refuelling, asteroid mining and orbital data centres — the industries cheap launch was meant to unlock — are still mostly demonstrations, failures and announcements, not revenue.
Every idea in this piece — mining an asteroid, refuelling a satellite, building a data centre in orbit — depends on the same underlying trend: it has become dramatically cheaper to get a kilogram of hardware into space. A July 2026 study led by Cambridge’s Bennett School of Public Policy and Turin’s Politecnico Institute, drawing on 4,405 launches between 1960 and 2025, put the average cost of reaching low Earth orbit at $87,023 per kilogram in 1960.1 That fell to roughly $20,000/kg by the early 1970s as Apollo-era production scaled up, then stalled for two decades once the Saturn V was retired and the Space Shuttle became the main US vehicle.1 It only began falling again after 1989, as Cold War-era state programmes gave way to commercial launch providers — and by 2025 the global average had reached $3,868/kg, a 96% fall since 1960.1
The same researchers project the fall will continue: to $1,569/kg by 2030 and $273/kg by 2040, if reusable heavy-lift vehicles such as SpaceX’s Starship reach routine, high-cadence flight.1 Those two later figures are a model’s central projection, not a measurement — they assume the cost curve continues its recent trajectory, which the researchers themselves flag is not guaranteed (see the caveat below).
The 96% fall is a global average, but it is not evenly produced. SpaceX accounted for roughly 80% of all payload mass launched into orbit worldwide in 2025, on the same Cambridge/Politecnico dataset1 — a concentration the study’s authors compare, in a related working paper, to the East India Company’s 19th-century hold on maritime trade.2 That matters for everything that follows: nearly every business case in this piece assumes launch keeps getting cheaper on schedule, and that schedule currently depends heavily on one firm’s reusability programme continuing to work and continuing to sell capacity at close to today’s prices.
Of the three industries this piece covers, orbital data centres have generated the most headlines relative to what has actually flown. Nvidia-backed Starcloud launched a 60kg satellite, Starcloud-1, in November 2025 carrying a single Nvidia H100 GPU — the first time a high-end AI chip has run in orbit — and used it to run inference and train a small language model.3 That is a genuine first, and a real measured fact. It is also a single GPU on a refrigerator-sized satellite, a long way from the 5-gigawatt facility the company has since pitched.4
The constraint is not power — sunlight in orbit is abundant and, outside eclipse, continuous. It is waste heat. In a vacuum there is no air or water to carry heat away by convection, so every watt a chip draws must eventually leave as infrared radiation, governed by the Stefan-Boltzmann law.5 Engineering estimates published this year put the radiator area needed to reject a single megawatt of waste heat at roughly 1,200 to 1,600 square metres — several tennis courts to a hockey rink, per megawatt56 — and that radiator cannot simply be the back of the solar array, because a panel wants to face the sun while a radiator wants a cold, shaded view of deep space.6 A modern terrestrial AI training cluster draws power in the hundreds of megawatts; scaling that to orbit on current radiator physics implies radiator arrays measured in square kilometres, not square metres.
Strip away the market-size forecasts and the mission announcements, and the measured record across the four industries most often bundled into "the space economy" looks like this. Satellite servicing is the one genuine, repeated commercial success: Northrop Grumman’s MEV-1 docked with a commercial satellite in geostationary orbit in February 2020 — the first such commercial docking — extended its life for five years, then undocked in April 2025 to serve another customer; MEV-2 has been docked and providing service since April 2021 and was extended in 2024.7 In-space manufacturing has one clear success (Varda’s first capsule grew crystals of the drug ritonavir in orbit and returned them intact in February 2024) followed by five further missions that have mostly been repurposed toward hypersonic defence-sensor testing rather than pharmaceutical production.8 Refuelling has real hardware built and contracted, but as of this piece’s publication date it has not yet flown: Astroscale U.S.’s APS-R spacecraft is manifested to launch in summer 2026 to refuel two Space Force satellites via an Orbit Fab depot.9 Asteroid mining has the weakest record of the four: AstroForge’s first mission, Brokkr-1, reached orbit in 2023 but its refinery payload never activated; its second, Odin, launched in February 2025 and lost contact in deep space before reaching its target asteroid; a third mission, Vestri, is scheduled for late 2026 or 2027 to attempt the first-ever private landing on a body beyond the Earth-Moon system — landing only, not extraction.10
The Terzi and Nicoli cost projections apply Wright’s Law — the empirical pattern that costs fall a fixed percentage with each doubling of cumulative output — forward from 2025; the paper itself flags that monopoly pricing power or geopolitical fragmentation could break that pattern, and no one can yet say which path is more likely.1 The two radiator estimates in Exhibit 03 come from published engineering analysis, not a built and measured facility at that scale; nothing close to a 1MW orbital data centre exists yet to check them against. No independent, audited figure exists for what it would actually cost to extract and process ore from an asteroid, because it has never been done — industry estimates of asteroid resource value (sometimes quoted in the trillions) describe the metal in place, not a recoverable, sellable quantity net of mission cost, and this piece has deliberately left such figures out rather than repeat them uncited. We also do not know whether Astroscale’s APS-R refuelling demonstration, manifested for summer 2026, will succeed by the time this piece is read; nor whether Varda’s pivot from pharmaceutical crystallisation toward defence and hypersonic-sensor payloads across its more recent missions reflects a permanent change in its business model or a temporary shift toward the more reliably funded customer.8
Credible, informed sources disagree on several distinct questions here. They are not all the same kind of question, so we have labelled each.
The cost curve keeps falling. Reusability is spreading beyond SpaceX — Blue Origin’s New Glenn and China’s emerging reusable boosters both target the same physics — and the study’s authors find the post-1989 rate of decline (44% per doubling of cumulative payload) is accelerating, not slowing, as commercial competition has replaced state monopolies.1
One firm still launches roughly 80% of world tonnage. The same researchers warn that a profit-maximising near-monopoly, combined with rising geopolitical fragmentation as nations build parallel, protected launch capacity, could stall or reverse the trend well before 2040 — the projection is a central case, not a guarantee.12
The falling-cost logic already supports real, if narrow, businesses: satellite life-extension has a five-year commercial track record,7 and in-space pharmaceutical manufacturing has produced at least one verified result.8 On this reading, mining and orbital data centres are simply earlier on the same curve servicing already climbed.
Servicing works because it solves a narrow, well-defined problem (keep an existing satellite flying) using mature docking technology. Mining and data centres face open engineering problems — unproven in-space ore processing, and cooling physics that scale badly with size — that cheaper launch alone does not solve; three years of asteroid-mining attempts have yet to return a qualified success.106
The 2015 US Space Resource Exploration and Utilization Act entitles American citizens to resources they extract in space, and the 69-nation Artemis Accords state that extracting space resources does not, by itself, amount to the "national appropriation" the 1967 Outer Space Treaty prohibits — extraction is framed as permitted "use," not forbidden ownership of the celestial body itself.11
Russia and China, both non-signatories, have called the Accords a unilateral US attempt to write international space law to its own advantage; legal scholars note the Outer Space Treaty is binding as customary international law on all states regardless of signature, and that whether extraction on this scale counts as "appropriation" has never been tested by any court or tribunal.1211
| Measure | Value | Basis / date | Source |
|---|---|---|---|
| Launch cost to LEO, 1960 | $87,023/kg | Dataset average | PNAS Nexus T1 |
| Launch cost to LEO, early 1970s plateau | ~$20,000/kg | Post-Apollo stall to 1989 | PNAS Nexus T1 |
| Launch cost to LEO, 2025 | $3,868/kg | Measured average | PNAS Nexus T1 |
| Fall, 1960→2025 | 95.6% | (87,023−3,868)/87,023 | Noodle News calc. |
| Launch cost to LEO, 2030 (projected) | $1,569/kg | Central model projection | PNAS Nexus T1 |
| Launch cost to LEO, 2040 (projected) | $273/kg | Central model projection | PNAS Nexus T1 |
| Cost fall per doubling of payload, pre-1989 | 17% | Wright’s Law fit | PNAS Nexus T1 |
| Cost fall per doubling of payload, post-1989 | 44% | Wright’s Law fit | PNAS Nexus T1 |
| Global payload launched, 2025 | ~4,900t | Measured | PNAS Nexus T1 |
| Global payload capacity, 2030 (projected) | ~9,100t/yr | Central projection | PNAS Nexus T1 |
| Global payload capacity, 2040 (projected) | ~32,000t/yr | Central projection | PNAS Nexus T1 |
| SpaceX share of 2025 global payload mass | ~80% | Measured | PNAS Nexus T1 |
| Global space economy value, 2024 | $600bn+ | ≈ Sweden’s GDP | PNAS Nexus / Cambridge T1 |
| MEV-1 commercial docking (first ever) | Feb 2020 | IS-901, GEO graveyard | Northrop Grumman T1 |
| MEV-1 service duration before undocking | 5 yrs | Ended Apr 2025 | Northrop Grumman T1 |
| MEV-2 docking | Apr 2021 | IS-10-02, GEO, still docked | Northrop Grumman T1 |
| MEV-2 contract extension | +4 yrs | Announced May 2024 | Northrop Grumman T1 |
| Varda W-1 launch / reentry | Jun 2023 / Feb 2024 | Ritonavir crystals grown | Varda T1 |
| Varda missions flown to date | 6 (W-1–W-6) | Through May 2026 | Varda T1 |
| Astroscale APS-R launch | Manifested, summer 2026 | Not yet flown, 19 Jul 2026 | SpaceNews T3 / Astroscale T1 |
| AstroForge Brokkr-1 | Apr 2023 | Refinery payload never activated | AstroForge T1 |
| AstroForge Odin | Feb 2025 | Lost contact in deep space | AstroForge / Space.com T1/T3 |
| AstroForge Odin build cost | $3.5m | Built in <10 months | Space.com T3 |
| AstroForge Vestri | Scheduled late 2026/2027 | Landing attempt only | Space.com T3 |
| Starcloud-1 launch | 2 Nov 2025 | 60kg, 1× Nvidia H100 | DCD / CNBC T3 |
| Radiator area per 1MW rejected, est. A | ~1,600 m² | ~20°C, WEF analysis | WEF T2 |
| Radiator area per 1MW rejected, est. B | ~1,200 m² | ~20°C, SatNews analysis | SatNews T3 |
| Artemis Accords signatories | 69 | As of 16 Jul 2026 | NASA T1 |