Theses›Space›Space-based datacentres

Space-based datacentres

REBUILT AROUND THE CHEAP-LAUNCH CASE (3 Sep 2026, at Lawrence's instruction: he believes launch costs fall faster than consensus).

  • Bear
  • Early

The callOrbit has constant sunlight, but compute there costs about 4x Earth even at free launch, since the platform outlives the chips by years; the radiator is the open slot.

D1D2R1R2R3C1foragainst D1D2R1R2R3C1foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the restraints.
  1. D1AI's power demand outruns the grid, with about $6.7T of datacentre capex needed by 2030, and orbit has constant sunlight and no permitting queue.
  2. D2Launch cost is collapsing, and SpaceX buying xAI to build orbital datacentres shows how big the operators think it is.
  3. R1Orbital compute costs about 4x terrestrial even at free launch, because the platform cannot be serviced and pays off over a 3-year chip cycle.
  4. R2The operator layer is late and captive: Starcloud raised $170M at $1.1B, SpaceX owns launch, and the 2027 pilots belong to SpaceX and Google.
  5. R3The unowned slices are small, and ACT pulled radiator panels in-house via Canyon Composites in Oct 2025.
  6. C1Starship's first orbital insertion attempt succeeds, but cheaper launch still leaves orbital compute about 4x terrestrial and does not enlarge the startup slice.Sep 2026
Rests onthe beliefs underneath
Supported by · confidence
Threatened by · confidence
What would prove itdated tests, open and resolved
2027202820292030203120322033today123456predictionswatch 2027202820292030203120322033today123456predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1Starcloud-2 flies a B200-class satellite with 100x the power of Starcloud-1.Dec 2026 · open
  2. 2Starcloud or another orbital-compute demonstrator announces a paying customer workload run in orbit between 6 Nov 2026 and 31 Jan 2027.Jan 2027 · open
  3. 3Terafab's first rad-hard 2nm silicon by mid-2027 and Google's capital decision after Suncatcher.Dec 2027 · open
  4. 4No commercially meaningful space-based datacenter is operating before end-2030; kW-scale demonstrators do not count.Dec 2030 · open
  5. 5A radiator flown in orbit above about 150 W/kg, or a demonstrated swap of compute modules in orbit with disclosed per-satellite costs.Dec 2030 · open
  6. 6Cost parity with terrestrial datacenters is not reached before 2032; the binding variable is launch cost reaching ~$100-250/kg on a fielded reusable vehicle.Dec 2032 · open
Expressionslisted, private, technologies
Private · 21 private companies, 14 of them in the US.

Cosmic Shielding Corporation · Metashield (by Metamagnetics) · Ramon Space · Coherent Logix · Vorago Technologies · Apogee Semiconductor · Frontgrade Technologies · NanoXplore (France, rad-hard FPGA) · Solestial (formerly Regher Solar) · Zeno Power Systems · Advanced Cooling Technologies (ACT) · Mynaric · Reflex Photonics · Laser Light Communications · AAC Clyde Space (Hyperion Technologies) · Starcloud · Axiom Space · Kepler Communications · Lonestar Data Holdings · ADA Space (Chengdu Guoxing Aerospace) · K2 Space

3 more private companies tracked, not shown

Technologies
Who has already bet on itprivate rounds, from public sources

$2.3B raised by 6 companies, $1.4B of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a top-tier venture firm: K2 Space, Starcloud, Kepler Communications. Backed by a strategic investor: Starcloud.

What changedthe record, newest first
  • 13 Oct 2026I rebuilt the orbital-datacentre model from the ground up and checked it against SemiAnalysis, Mach33, Starcloud and Google. Tokens from orbit cost about 8x Earth today on hardware that has actually flown, about 2x even if launch were free, and about 2x in 2035 if every paper-design radiator and solar array matures on schedule. Tokens per watt is a chip property, not a location property. Two things everyone gets wrong: dawn-dusk orbits still have eclipses, and GPUs cannot run hot enough to shrink the radiators the bull case needs.
  • 24 Jul 2026The space-datacentre story changed shape this year: SpaceX bought xAI and filed for a MILLION-satellite orbital data centre constellation, and China's state banks have quietly committed ~20x more capital than all Western VCs combined. I still would not underwrite the operators (the maths only closes if a $5K space chip and hourly Starship launches both land), but the demand push is real — the US grid literally failed its first capacity auction — and the picks-and-shovels layer (radiators, optical links) is where the venture-shaped opening sits.
  • 1 Jul 2026The independent SemiAnalysis teardown everyone's been citing is now in my notes in full — same conclusion: orbital compute is ~4x terrestrial today, parity around 2040 (early-2030s only if terrestrial hits a wall), and the quiet killer is that space datacentres live 5 years to Earth's 15 because you can't send a technician up to swap a dead GPU.
3 earlier
  • 28 Jun 2026I ran the orbital-datacentre numbers from first principles and the headline doesn't survive: space compute is ~4x terrestrial even if launch were FREE. Launch cost isn't the gate everyone says it is — it's third, behind being able to service the chips (you can't) and the sheer cost of the solar arrays and radiators you have to haul and rebuild every time the silicon obsoletes. AI datacentres don't move to orbit because it's cheaper, because it isn't.
  • 26 Jun 2026An operator who actually builds in space agrees: orbital datacentres are a real but timing-gated bet — credible at one order of magnitude cheaper launch (~2029-32), not yet two. And the binding cost isn't just $/kg to orbit, it's asset life: radiation kills the chips and you can't send a technician.
  • 23 Jun 2026Space datacentres aren't a hoax and aren't imminent: the serious estimates agree they only pencil out when launch hits ~$100-250/kg (early-to-mid 2030s, contingent on Starship). Radiation is largely a solved part-selection problem; the real gate is launch cost, full stop.
Reading
Show the reading