Space data centres: SpaceX and Blue Origin race to orbit while scientists question the physics

SpaceX filed on Jan 30 to launch up to 1,000,000 low-Earth-orbit satellites for onboard AI compute; Blue Origin proposed 51,600 satellites (Project Sunrise) plus 5,408 TeraWave nodes; Starcloud raised $170M at a $1.1B valuation and filed for up to 88,000 satellites. Key constraints: thermal management requires ~1,200 m² radiator per 1 MW (making several-hundred-MW systems astronomically large), radiation hardening adds 30–50% to hardware costs and cuts performance 20–30%, and orbital latencies (several ms to 60–190 ms) preclude training workloads — orbital compute seems viable only for inference. Economics are unfavorable: IEEE estimates a 1 GW orbital data centre >$50B (~3x terrestrial), launch costs must fall below ~$200/kg (current Starlink economics ~$1,000–$2,000/kg; some analysts argue $20–$30/kg needed), so physics and cost barriers make orbital data centres unlikely to be competitive in the near term.

Analysis

Orbital computing hype is already creating a capital rotation risk: venture and public capital is being bid toward space-first narratives that, if the physics and regulatory frictions persist, will face multi-year realization delays and write-downs. That creates a near-term funding squeeze for terrestrial infrastructure projects (permits and grid upgrades) as capital is redeployed to showier space plays, increasing execution risk and capex timing volatility for cloud providers and chip vendors.

Thermal and radiation tolerance scale as multiplicative rather than additive costs — once you price in replication, shielding, and servicing, the mass-to-orbit loop feeds back into launch economics and insurance costs, producing a regime where marginal dollars buy less compute per dollar than on Earth for the foreseeable cycle. This amplifies regulatory exposure: visible orbital expansions invite stricter licensing and liability regimes that can crystallise value erosion for public investors in space-exposed assets within quarters.

From a product-demand standpoint, the dominant commercial opportunity remains grounded: tightly coupled training workloads stay in terrestrial racks, preserving demand for high-end datacenter GPUs and on-prem power/cooling upgrades for at least the next 3–5 years. The primary long-term reversal would be a structural drop in launch cost or a breakthrough in non-convective cooling or in-orbit servicing; those are multi-year binary catalysts that would re-rate a subset of space plays but are binary and low-probability on a 24-month view.

Key Decisions for Investors

• Long NVDA via a calendar/vertical spread: buy Jan-2028 LEAPS 500C and sell Jan-2028 800C (size 3–5% notional). Rationale: preserves exposure to sustained on-Earth training demand while capping premium; reward skew ~2–3x if enterprise AI capex stays on track, downside limited to premium (stop 35% of premium).
• Long GOOG (Class C) stock or 12–18 month call spread (buy Jan-2028 140C, sell Jan-2028 190C) — conviction trade for cloud & networking moat that benefits if orbital timelines slip and terrestrial capex continues. Target hold 12–24 months; downside: 12-month equity risk if macro slows; expected IRR 15–25% if cloud growth persists.
• Short the space-theme ETF (ARKX) via a 6–9 month put spread: buy 9-month 15% OTM puts and sell 9-month 25% OTM puts to fund position (size 2% notional). Rationale: hedge vs near-term derating of 'space hype' as regulatory pushback and engineering reality checks compress multiples; limited downside due to spread structure.
• Buy optionality on defence/aerospace rad-hard suppliers: MCHP 12–18 month call options (size 1–2% notional). This is a low-probability, high-upside asymmetric play: if firms win durable military/space contracts or modular servicing emerges, they rerate; if not, loss is capped to premium.