2026 06 04 TO Boldly GO THE Case FOR Space Datacenters

Summary

SemiAnalysis argues space datacenters are economically unviable in 2026 (4.4× terrestrial cost) and reach parity only ~2040 under the base case, or early 2030s if terrestrial power peaks in 2028 (Musk scenario). The load-bearing causal claim: it's not free solar or cold space cooling that makes space viable, it's terrestrial power-layer exhaustion forcing the hand. Meanwhile SpaceX's S-1 claims 100GW/yr launch target, making this directly relevant to SPCX IPO valuation.

Article

Title: To Boldly Go: The Case for Space Datacenters Authors: Daniel Nishball, Pranav Myana, Ellie Holbrook, and 7 others Publication: SemiAnalysis, June 3, 2026 Note: Paid subscription — free preview extracted below.

Introduction & Thesis

The article examines the economic viability of deploying AI datacenters in orbital space. The authors challenge popular misconceptions while analyzing when space-based infrastructure could become cost-competitive. Core thesis: space datacenters succeed only when terrestrial power constraints force deployment or costs achieve parity (~2040), not because of commonly cited benefits.

Four Debunked Arguments

Solar Energy Myth: LEO satellites receive sun only ~60% of the time (avg 800 W/m² vs 1,361 W/m² max). Sun-Synchronous Orbits offer better exposure (35-min daily eclipses only).

"Free" Cooling Fallacy: Space requires radiators for heat dissipation via radiation (no convection). ISS radiator system covers 325 m² and costs $340–500M to remove just 70 kW.

Latency Concerns: LEO satellites pass overhead 5–7 min/day. Inter-satellite link hops accumulate 30–80ms latency.

Orbital Slot Constraints: Dawn-dusk Sun-Synchronous Orbits are a constrained subset with limited capacity.

Five-Layer Terrestrial Power Supply Framework

1. Grid-connected power (cheapest; interconnection queues now 7 years)
2. Converted capacity (repurposed bitcoin mining ~8–10 GW)
3. Behind-the-meter generation (on-site; reaching 26 GW by 2030)
4. Industrial production expansion
5. Semiconductor production (universal constraint — chips are the binding limit before power)

Key finding: "chip manufacturing will be the global constraint before we even worry about supply." AI-related DRAM demand consumes 70% of total DRAM wafer capacity by 2027.

TCO Analysis (2026, 30.5kW B300 cluster)

Launch costs dominate: $1.6M out of $3.1M total space capex. Space is 4.4× more expensive today.

Cost Parity Timeline

• Base case: ~2040, after 80% Starship launch cost reduction and significant tech scaling
• Musk scenario (terrestrial peaks 2028): early 2030s with 30% cost advantage possible
• Implication for SPCX: SpaceX S-1 claims 100 GW/yr launch target; SemiAnalysis model implies this requires Starship economics to materially improve from current levels

SpaceX Terafab Assessment

SpaceX's proposed 1-terawatt-per-year fab faces constraints:

• 100K WSPM entry = 2.5% of global foundry capacity
• Full 1M WSPM scale = 68% of TSMC's global output
• Memory IP concentrated in Samsung/SK Hynix/Micron — licensing dependency unresolved

Paywall begins at "So When Does Space Actually Get Interesting?" (Part 4 full cost breakdown requires subscription)