To Boldly Go: The Case for Space Datacenters

Daniel Nishball · SemiAnalysis · June 03, 2026 at 13:07 · ⏱ 58 min read  | Read on Substack ↗
Summary
The article argues that space datacenters are not viable today (4x more expensive than terrestrial) but could reach cost parity by ~2040 if launch costs, radiators, and solar arrays decline sufficiently, and if terrestrial power and permitting constraints become severe. For markets, the near-term bottleneck is chip manufacturing (TSMC, HBM), not datacenter capacity, so space compute is a long-duration optionality rather than an immediate investment theme.
  • •Space datacenter LCOC in 2026 is $10.91/hr/GPU vs $2.49/hr/GPU for terrestrial — a 4.4x premium.
  • •Launch costs ($1.6M for a 30.5kW cluster) are the largest single cost driver for space deployments.
  • •Cooling in space is not free; radiation is the primary heat rejection method, and the ISS radiator system ($340-$500M) only handles 70kW.
  • •Chip manufacturing (advanced nodes at TSMC, HBM from SK Hynix/Samsung/Micron) is the universal constraint, limiting AI deployment on Earth and in space alike.
  • •Terrestrial datacenter capacity could quadruple to 338 GW by 2030 via grid connections, bitcoin miner conversions, behind-the-meter generation, and industrial expansion.
  • •Musk's Terafab initiative aims for 1M WSPM by ~2030, but the article notes this would require ~68% of TSMC's current global output and is logistically daunting.
  • •Space-Earth cost parity is modeled for ~2040 in the base case, but the 'Elon Musk scenario' (severe terrestrial constraints) brings near-parity by the early 2030s.
  • •Space datacenter useful life is only 5 years (vs 15 for Earth), amplifying the capex gap; in-space robotics could extend it to 10 years post-2032.
Read time 58 min
Length 58,647 chars
Category finance
Trade Ideas
Daniel Nishball Substack author, SemiAnalysis
The article identifies TSMC's N3 wafer capacity as the binding global constraint on AI compute, noting that AI-related demand will consume 86% of N3 output by 2027. This implies sustained pricing powe
The article identifies TSMC's N3 wafer capacity as the binding global constraint on AI compute, noting that AI-related demand will consume 86% of N3 output by 2027. This implies sustained pricing power and long lead times for TSMC's advanced nodes. Risk: Geopolitical disruption (Taiwan scenario) or a cyclical downturn in non-AI demand could offset pricing gains.
TSM
Daniel Nishball Substack author, SemiAnalysis
HBM supply from Micron (along with SK Hynix and Samsung) is called out as a key bottleneck: AI-related DRAM demand rises from 12% of wafer capacity in 2023 to 70% by 2027. HBM consumes ~3x the wafer c
HBM supply from Micron (along with SK Hynix and Samsung) is called out as a key bottleneck: AI-related DRAM demand rises from 12% of wafer capacity in 2023 to 70% by 2027. HBM consumes ~3x the wafer capacity per bit, amplifying tightness and pricing power for Micron. Risk: Memory cycle downturns or technology transitions (HBM4) could compress margins if oversupply emerges.
MU
Daniel Nishball Substack author, SemiAnalysis
The article references 'EUV tool production constraints' as one of the obstacles that must be overcome for the high-demand scenario, and notes that getting close to 800 GW of AI compute would require
The article references 'EUV tool production constraints' as one of the obstacles that must be overcome for the high-demand scenario, and notes that getting close to 800 GW of AI compute would require 'the entire global EUV fleet dedicated to AI.' This underscores ASML's monopoly position and the criticality of its EUV systems for future compute expansion. Risk: Export controls, geopolitical tensions (China), or technology obsolescence (High-NA EUV adoption delays) could disrupt the demand narrative.
ASML
Daniel Nishball Substack author, SemiAnalysis
The article states that small, efficient chips 'akin to Tesla's FSD chips' are the most likely form factor for space datacenters, and that Tesla's Terafab initiative (alongside SpaceX and xAI) aims to
The article states that small, efficient chips 'akin to Tesla's FSD chips' are the most likely form factor for space datacenters, and that Tesla's Terafab initiative (alongside SpaceX and xAI) aims to produce 1M WSPM with 80% allocated to orbital compute. This validates Tesla's custom silicon capability and its strategic role in space AI. Risk: Terafab's ambitious scale (requiring ~68% of TSMC's global output) faces immense execution risk; Tesla's automotive margins could be strained if capital is diverted.
TSLA
More from SemiAnalysis
AI Dark Output: The Visible Cost of Invisible Output
May 29, 20:02
Finding Miscompiles for Fun, Not Profit
May 28, 00:22
Anthropic Growth and Bedrock Mix Drive AWS Margins Higher While Peers Lag
May 27, 16:25
Inside the 800VDC Revolution – Part 1
May 26, 05:37
EDA Market Primer - Market Dynamics, Cadence, Synopsys, Siemens, China EDA Rise
May 21, 20:53

This newsletter, published June 03, 2026, features Daniel Nishball discussing TSM, MU, ASML, TSLA. 4 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Daniel Nishball  · Tickers: TSM, MU, ASML, TSLA