First-principles comparison · model, not prophecy
This mirrors the framing of Andrew McCalip’s orbital vs terrestrial model: the question is whether useful watts and operable thermals close on cost before counting the glamour. Here the terrestrial side is an underground facility in arid Australia: rock shell, solar plus lithium buffering, cooling by the cheaper of geo-exchange or mechanical A/C, and IT refreshed on a schedule so denser nodes (e.g. future leading-edge silicon) can replace older fleets.
In addition to being non-polluting (once installed! Of course we have arbitraged the bulk of the pollution to China, Brazil etc for mining and manufacture). Once evolutions and revolutions in chip architectures happen we can easily upgrade. During the 5 years we expect not only the chip separation to decrease but also some fundamentals to completely change like the introduction of analog state computing, if successful would reduce the power demands by a factor as high as 10000, or allow an increase of computing power by same.
GPUs/accelerators are costed as generic IT $/W with a refresh cadence—not a die-shrink forecast. All figures are illustrative USD; no financing, tax, or permits. See assumptions to argue with the numbers.
math.js)
Same slider set and formulas as
andrewmccalip.com/space-datacenters
— ported from
thoughts/static/js/math.js
(calculateOrbital). Ops are fixed at 1%/yr of satellite hardware; Starship payload is 100 t; array area scales from the V2 Mini reference (116 m² @ 27 kW).
Coober Pedy–class: dug-in shell, high insolation, cheap land. Cooling picks the better of geo-loop vs chiller path at your sliders.
calculateOrbital implementation verbatim (geometric cell degradation, ceil satellite count, hardware $ on actual initial watts, launch $ on total mass, ops 1% of hardware × years, GPU replacement on hardware × rate × years, NRE). Outback remains our separate model.(IT × PUE) / (CF × 8760 h) nameplate DC watts, then × $/W slider.IT_W × ($/W + step × event_index)—a stand-in for migrating to denser/newer nodes without forecasting a specific process node.grid kWh = IT_MW × PUE × 8,760 h/y × 10; water = IT kWh/y × WUE (L/kWh) × 10 y priced at $/m³; CO₂ = grid kWh × kg/kWh. No capex, staff, taxes, PPA vs retail, REC offsets, or REC arbitrage. See operational costs below.Conventional suburban grid datacentre: same IT nameplate as Global, typical suburban facility on utility power, with evaporative / cooling-tower style water use (tunable). Horizon is fixed at 10 years (independent of the programme slider). Carbon is location-based grid average (Scope 2 style); it excludes embodied equipment and upstream fuel methane.