Orbital Data Centers

PRE-COMMERCIAL · 30 operators · HHI LOW

Compute and storage infrastructure deployed on orbit that processes data at its source in space and offers solar-powered capacity outside the terrestrial power, water, and land constraints that bound conventional data centers. Intended revenue sources are edge-processing fees for Earth-observation operators, on-orbit hosting for government and enterprise workloads, and subscription compute capacity for artificial-intelligence and analytics customers. The sector is pre-commercial, with activity consisting of small demonstration payloads and concept studies rather than production services, and revenue dominated by early seed and government-study funding.

CURATED RESEARCH · AS OF 2026-06-10 · EVERY FIGURE CARRIES ITS SOURCE INLINE · UNVERIFIED AMOUNTS HEDGED AS “REPORTED”

UNIT ECONOMICS — FORMULA + NUMBER + SOURCE

• Compute working unit: 1 NVIDIA H100 SXM = 700 W TDP (Nvidia spec); an 8-GPU DGX H100 ~ 10.2 kW; a 4-server AI rack ~ 41 kW IT load; use a 44 kW 'rack-equivalent' (41 kW + ~7% overhead) as the working unit (Nvidia spec; spacecomputer.io).
• Solar supply: LEO solar constant = 1,361 W/m2. Array need = 44,000 W / (1,361 W/m2 x 0.30 cell efficiency x 0.85 packing/pointing) ~ 127 m2; at ROSA-class specific power (~130 W/kg), 44,000 / 130 ~ 340 kg of array. Google's Suncatcher paper computes a panel in 650 km dawn-dusk sun-synchronous orbit is up to 8x more productive than the same panel on Earth; non-SSO orbits eclipse ~35% of each orbit, which is why every serious design picks dawn-dusk SSO (research.google).
• Thermal — the binding constraint: Stefan-Boltzmann q = epsilon x sigma x T^4 = 0.9 x 5.67e-8 x (353 K)^4 ~ 790 W/m2 per side ideal at 80 C electronics temperature; ISS EATCS achieves ~166 W/m2 in practice (70 kW rejected over 422 m2 of radiator); industry rule of thumb ~ 2.5 m2 of radiator per kW (spacecomputer.io). The 44 kW rack-equivalent therefore needs ~110 m2 (optimistic) to ~265 m2 (ISS practice) of radiator = 550-2,650 kg at 5-10 kg/m2, i.e. 12-60 kg/kW for thermal alone.
• Mass cross-checks: Starcloud-3 is specced at 3 t for ~200 kW = 15 kg/kW whole-spacecraft (SpaceNews); SpaceX's FCC filing implies ~1M tonnes/yr for 100 GW ~ 10 kg/kW whole-system — inferred from Musk statements, not a filed spec.
• Launch math (June 2026): Falcon 9 dedicated ~ $74M list / 22.8 t to LEO = ~$3,245/kg; F9 rideshare raised to ~$7,000/kg ($350K base for 50 kg SSO) — satbase/Payload reporting of SpaceX's 2026 price increase. At 15 kg/kW x $3,245/kg = ~$48.7/W for launch freight alone, vs ~$10-13/W ALL-IN construction cost for terrestrial AI datacenters (AstraVeris /datacenter/ pro formas on Turner BCI/JLL basis). F9-class pricing is 4-5x too expensive on freight alone — demos only.
• Starship scenario: 15 kg/kW x $200/kg (assumption — no published price) = $3/W launch; add spacecraft hardware at scale (unproven, $2-5/W) = $5-8/W with no ongoing power bill. Terrestrial comparison: $10-13/W build PLUS energy at 8,760 kWh/kW-yr x $0.05/kWh ~ $440/kW-yr forever. That is the entire bull case in one line.
• Insurance line-item: space hull rates run ~4-12% of insured value at launch vs ~0.2-0.5%/yr for terrestrial datacenter property — a 100x+ difference that must appear in any orbital pro forma (orbitalradar / Howden space-insurance market reviews; global premium pool ~$500-600M/yr).

KEY RISKS — INCLUDING THE INSURANCE ANGLE

• Starship $/kg: every business case in the vertical is a derivative of one company's unpriced product. SpaceX publishes no Starship price; all $/kg figures are planning assumptions (Google's <$200/kg by mid-2030s).
• Thermal scaling beyond ~1 MW per spacecraft is unproven. Mach33 analysis argues radiators are only 10-20% of total mass at 100 kW scale — a trade, not a wall, below MW scale — but at GW scale it is unproven (SatNews 'Physics Wall', Mar 2026). Advanced concepts (liquid droplet radiators, ~450 W/kg) exist only in labs.
• GPU obsolescence vs build-to-launch latency: 12-24 month integration cycles mean obsolescence-in-transit — the H100 launched on Starcloud-1 (Nov 2025) was already three generations stale at the Blackwell announcement. Underwrite as full-depreciation DCF with deorbit cost, never an exit cap rate: no stabilized-NOI market, no resale, no residual (5-7 yr bus life vs 3-4 yr GPU relevance).
• Radiation and silent-data-corruption rates are unknown at fleet scale; Google's Trillium TPU total-ionizing-dose data supports survivability to the expected shielded 5-year dose, but it comes from Google's own unreviewed paper.
• Insurance gap (the underwriting angle): launch + first-year in-orbit all-risk is insurable today on conventional satellite-hull terms (global premium pool ~$500-600M/yr; orbitalradar/Howden). NOT insurable today: compute-SLA/business-interruption on orbital capacity, GPU residual value, constellation-scale aggregation (10,000+ sat fleets exceed per-risk capacity ~$500M and concentrate debris-cascade tail risk). Megaprojects will self-insure, so insured premium volume will lag deployed value — an opening for parametric radiation-event and launch-slip products aimed at tier-2 pure-plays who must show insurance to lenders.
• Demand premise: terrestrial datacenter power constraints could ease (gas turbines, SMRs) and remove the orbital premise entirely.
• Hype contagion: vendor TAMs and renders vastly outrun the ~$250M of verified pure-play equity in the vertical; FCC filings are nearly free to make and are not a demand signal.

Sources: Google Research — Suncatcher system design · SpaceNews — SpaceX 1M-satellite FCC filing · SpaceNews — Starcloud Starlink laser order / Starcloud-3 spec · spacecomputer.io — cooling for orbital compute · satbase — SpaceX Falcon 9 2026 price increase · Luminix — data centers in space viability analysis · SatNews — 'The Physics Wall' cooling analysis · orbitalradar — space insurance market · Howden — space insurance outlook · BIS Research — in-orbit data centers market (vendor-grade)

INTEL: curated research as of 2026-06-10 · adversarially verified before publication · single-source amounts hedged or omitted · not refreshed by the 6-hour pipeline

THESIS: deterministic fallback (Gemma unavailable)

HHI estimated from ARI-weighted market-share proxy (ARI × data-coverage, normalized). 0 = perfectly competitive, 1 = single-operator monopoly. Banding: <0.15 Low, 0.15-0.25 Moderate, 0.25-0.50 High, >0.50 Concentrated.

Principal due by year across public sector issuers. Private operators excluded (no 10-K). Source: quarterly 10-K footnote extraction.

• 01 The next quarterly earnings print from the top public Orbital Data Centers operators — backlog, guidance, and any cash-runway inflection will mark comps for the entire peer group.
• 02 Government contract renewals and new-start awards inside Orbital Data Centers — each award validates or resets the anchor-demand assumption that underwrites the sector's valuations.
• 03 Any primary funding round inside Orbital Data Centers, particularly down-rounds or pro-rata waivers — these reset cap tables and tell us where late-stage capital is willing to clear.

WATCH: deterministic fallback (Gemma unavailable)