Orbital AI Datacenter Calculator i

Orbit viewiinertial frame · Earth rotates beneath the orbit · drag to look around · Sun marker shows its true direction

Model notes & what is not modeled

• Eclipse: circular orbit, fraction in shadow ⟦⟦f_e = frac{1}{π} acos frac{sqrt{h^2 + 2 R_E h}}{(R_E + h) cos β}⟧⟧ β = 0° is the worst case (sun in orbit plane); β = 90° is a dawn-dusk sun-synchronous orbit with no eclipse (only achievable near-polar, and the β drifts seasonally).
• Array: sized to carry the IT load plus battery recharge during sunlight: ⟦⟦P_array = frac{P}{η_pcu} (1 + frac{f_e}{f_d η_rt})⟧⟧ Area uses 1361 W/m² × cell efficiency × derate (pointing, temperature, packing, end-of-life degradation).
• Thermal: every watt of electrical power ends up as heat. Two-sided radiator: ⟦⟦A_rad = frac{Q}{2 ε σ (T_rad^4 − T_sink^4)}⟧⟧ The effective sink in LEO is not 3 K: Earth IR (~240 W/m²) and albedo load a radiator, so 200–250 K is typical. ⟦T_rad = T_chip − ΔT⟧: the radiator can only be as hot as the chip junction minus the drop through cold plate, coolant loop and fin, so an 85 °C chip with a 30 K drop gives a 328 K radiator. Hotter-qualified silicon is the main way to shrink the radiator.
• Drag: ⟦F_D = ½ ρ v^2 C_d A_proj⟧ with an exponential-interpolated standard atmosphere (mean solar activity) scaled by the solar-activity multiplier. Projected area = fraction × (array + radiator planform); sun-tracking arrays in a terminator orbit can fly nearly edge-on, in other orbits the orbit-averaged projection is large.
• Propulsion: electric thrusters cancelling drag continuously: jet power ⟦P_thr = frac{F_D v_e}{2 η_thr}⟧ with ⟦v_e = I_sp g_0⟧; propellant rate ⟦ṁ = F_D / v_e⟧. Decay lifetime integrates ⟦frac{d a}{d t} = −ρ frac{C_d A}{m} sqrt{μ a}⟧ from the design altitude down to 150 km.
• Launch: ideal energy per kg ⟦ε = frac{μ}{R_E} − frac{μ}{2 r}⟧ (no rotation credit, no losses). Chemical energy per launch assumes a Starship-class stack burning ≈1000 t of methane (≈50 MJ/kg LHV). Payload per launch is not derated for altitude or inclination.
• Economics: capex (compute hardware + spacecraft hardware + launch) is annualized with a capital-recovery factor over the design life. Upkeep = propellant resupply (charged at launch $/kg; the propellant itself is negligible), replacement of failed compute units (hardware + relaunch, since nothing is repairable on orbit) and operations as a share of capex. The ground baseline uses the same hardware cost and life, a 15-year facility, PUE-scaled electricity, and a 2 %/yr repair allowance. Orbit has no electricity bill — its power plant is in the spacecraft hardware and launch lines.
• Politics & permitting: the direct political bill (lobbying, PACs, community benefit agreements) is a recurring % of facility capex; the whole US industry's known political money is ≈0.1 % of its capex, so this line cannot move the verdict — it is here so you can see that. Tax incentives are the larger, opposite-signed political term and are taken off ground hardware, facility and electricity. The cost that does matter is time: expected ground time-to-power ⟦t_power = t_permit (1 + frac{p}{2 (1 − p)})⟧ for block probability ⟦p⟧, charged as carrying cost on half the facility capex plus 8 % sunk cost per blocked attempt. Orbit time-to-power = licensing lead time + fleet launches ÷ cadence. Hardware is bought when the power is ready, so waiting costs time, not chip generations: compute delivered ⟦D⟧ years late is discounted to ⟦(1 + x)^{−D}⟧ of compute delivered now, so delivering the same value after waiting ⟦D⟧ years costs ⟦(1 + x)^D⟧ as much; the excess, ⟦C_time = C_yr ((1 + x)^D − 1)⟧, is booked on each side as the opportunity cost of time to power. Stations work from the day they are launched, so orbit's factor is averaged over the launch campaign. It is part of annual cost and LCOC, so the Investor's verdict prices time.
• Verdicts: commercial = levelized cost per IT-kWh, orbit vs ground, with "marginal" for anything within ±15 % (well inside the uncertainty of these inputs). Environmental = lifetime CO₂e per IT-kWh with the same band; water is reported but orbit wins it by construction (no cooling towers, no power plant), and debris, re-entry aerosols and night-sky impacts are listed but not scored. There is no separate permitting verdict: time to power enters the commercial verdict as the opportunity-cost line, which is the only place where politics changes the answer; set the discount rate on late compute to 0 to see the plain bill. Neither verdict prices latency, sovereignty, insurance availability, or debris liability.
• Environment: launch CO₂ = methane burned × 2.75 (stoichiometry) × a lifecycle multiplier (upstream leakage, LOX/LNG production) × an optional high-altitude forcing multiplier; it is prorated by mass share of each launch. Embodied carbon is per kW of IT (hardware), per kg (spacecraft) and per kW over a 15-year life (facility). Ground water = on-site WUE + power-plant consumption × PUE; both sides carry fab water for the chips.
• Offsets: both sides may buy credits for the same share of their annual CO₂e at one price; the bill goes into annual cost and LCOC. Only share × integrity is netted off the reported intensity, so cheap low-integrity credits barely move the cost and move the verdict less. Because both sides offset the same share, the orbit/ground ratio never changes; only the "neutralised" state (100 % share at 100 % integrity) retires the Planet's verdict, and the comparison becomes purely a bill. The compliance carbon price is a separate slider: a datacenter in the EU pays ≈€82/t through its power whether or not it buys offsets. The high-altitude forcing multiplier is not CO₂ (soot, water vapour and NOx are short-lived), so "offsetting" it with CO₂ removal is GWP bookkeeping, not physics. Debris, re-entry aerosols, night sky and water cannot be offset and stay unscored.
• Not modeled: radiation hardening and single-event upsets (and the mass of shielding), micrometeoroid/debris protection, deployable-structure dynamics at km² scale, array shadowing of radiators, Earth-link bandwidth (laser downlink), on-orbit servicing or repair (failed units are written off and replaced, never fixed), payload derating for altitude or inclination, collision risk and conjunction avoidance, and end-of-life disposal.