2026 06 03 Feed Semianalysis TO Boldly GO Case FOR Space Datacenters

Summary

SemiAnalysis quantifies orbital datacenter economics and finds space DCs are currently 4x more expensive than terrestrial (LCOC $10.91 vs $2.49/GPU-hour). The report debunks four common pro-space arguments (24-hour free solar, free cooling, low latency, no permitting) and identifies semiconductor manufacturing — not power or permitting — as the binding terrestrial constraint through the 2020s. Cost parity requires ~80% launch cost reduction (Starship target $250/kg vs current $1,400–$1,800/kg) and lands in 2040 base case / early 2030s if terrestrial supply is constrained by permitting. Space DCs are rational only when terrestrial supply is exhausted, not for inherent efficiency advantages. Key wildcard: Terafab (Elon Musk's 1M wafer-starts/month by 2040 target) viewed skeptically — process IP licensing more realistic than greenfield.

Article

Core Economics (2026 Baseline)

For a 30.5kW B300 cluster:

• Space LCOC: $10.91/GPU-hour
• Terrestrial LCOC: $2.49/GPU-hour
• Ratio: ~4.4x more expensive in space

The 17x difference in levelized datacenter costs stems primarily from the 5-year useful life assumption in space (vs 15-year terrestrial). Launch costs represent $1.6M of the $3.1M space datacenter capital expense. 20% spare GPU provisioning required in space vs 5% terrestrial (failure rate in orbit substantially higher).

Four Debunked Arguments

1. "24-hour free solar": LEO satellites receive sunlight only ~60% of the time; Sun-Synchronous Orbits are better but still require batteries for ~35-minute daily eclipses. Not "free."

2. "Free cooling": Space's vacuum makes cooling harder, not easier. Heat dissipation relies entirely on radiation (no atmospheric convection). The ISS radiator system illustrates the scale of engineering required.

3. "Low latency": Satellites pass overhead only 5–7 minutes daily; multi-hop inter-satellite links accumulate 30–80ms delays. Not competitive with ground-based latency.

4. "No permitting": Dawn-dusk SSO capacity is severely constrained vs LEO's broader orbital slots. Regulatory coordination (ITU frequency, debris rules) is a real constraint.

The Five Terrestrial Layers

SemiAnalysis identifies terrestrial power constraints in layers (from most available to least): grid-connected supply → converted capacity (crypto miners) → behind-the-meter generation → industrial production scaling → semiconductor production. Current modeling shows semiconductor manufacturing — not power or datacenter capacity — is the binding constraint through the 2020s. This applies equally to space and terrestrial compute.

Cost Convergence Scenarios

Base case (2040): Requires ~80% launch cost reduction (Starship: $250/kg from current $1,400–$1,800/kg), radiator/solar array cost reductions, and GPU reliability improvements. Space DCs remain optional given ample terrestrial capacity through this period.

"Elon Musk" case (early 2030s): Terrestrial capacity constrained by regulations/permitting (NIMBYism, grid interconnection queues); space becomes necessity as terrestrial supply is exhausted. Near-parity achievable in this scenario.

Terafab wildcard: Musk's 1M wafer starts/month by 2040 (~70% of current TSMC output) is viewed skeptically. Process IP licensing (rather than greenfield fab development) is "the realistic path." Timeline and memory production claims are questioned.

Implications for the Thesis

The analysis supports the orbital DC thesis as a structural long-duration option rather than a near-term trade. The scenario where it matters most (terrestrial supply exhausted by permitting) is exactly the Musk/bull case for SPCX. The article's framing: "Space DCs make economic sense only when terrestrial supply is exhausted" — which, if terrestrial permitting remains a constraint, could arrive faster than the 2040 base case suggests.

Key contradiction to terrestrial-power-flat-to-orbital-dc-arbitrage step 2 (the "5x solar advantage" claim): SemiAnalysis finds SSO solar is better than LEO but requires battery storage for eclipse periods; it is NOT 24-hour. The "5x advantage" requires Sun-Synchronous orbit AND the comparison must be against terrestrial solar (not baseload). This weakens the step-2 claim as stated.

Andrew Feldman's clustering objection (networking GPUs across vacuum) also referenced: "solving GPU reliability/redundancy challenges" is listed as a required innovation for orbital DCs — corroborates Feldman's "last 10% is 80%" framing from the All-In IPO panel.