Orbital datacenters: separating funded reality from renders

SpaceX / xAI. SpaceX filed with the FCC on Jan 30, 2026 for an "Orbital Data Center System" of up to one million satellites at 500–2,000 km (accepted Feb 4, 2026, DA 26-113), days after absorbing xAI in an all-stock deal valuing the combined entity at $1.25T (CNBC) with orbital datacenters cited as the core rationale. The S-1 followed on May 20, 2026: Nasdaq ticker SPCX, ~$135/share, targeting a $1.75–2T valuation and roughly $75B raised. The widely circulated 10 kg/kW system mass budget (1M tonnes/yr → 100 GW/yr) is inferred from Musk's statements, not a filed engineering spec — but if real, it is the most aggressive specific-power target in the industry. SpaceX entering as operator compresses every pure-play's strategic space: it owns the $/kg lever everyone else's model depends on.

Google — Project Suncatcher. The vertical's most-cited economics paper (research.google): TPU satellites in 650 km dawn-dusk SSO, ~1 km clusters, with two prototype satellites built and operated by Planet Labs launching by early 2027 (Planet). Its key published figures — solar panels up to 8x more productive than terrestrial in that orbit, Trillium v6e TPUs surviving 15 krad(Si) with first HBM irregularities at 2 krad versus a ~750 rad expected shielded 5-year dose, 800 Gbps each way per optical transceiver pair, and a <$200/kg launch-cost assumption by the mid-2030s — all come from Google's own unreviewed paper. Accurately quoted everywhere, independently replicated nowhere. Planet (NYSE: PL, FY2025 revenue $244.4M per its 8-K) is the listed-equity proxy.

Blue Origin — Project Sunrise. FCC filing in March 2026 for 51,600 compute satellites (TechCrunch); Bezos publicly frames gigawatt space datacenters as inevitable while calling a 2–3 year timeline "a little ambitious" (CNBC, May 2026). No disclosed budget; New Glenn is the in-house lift lever.

Space Compass (NTT + SKY Perfect JSAT). Japan's sovereign vehicle: distributed compute/storage under NTT's IOWN photonics strategy, JAXA Space Strategy Fund selection for optical data relay, a Japan MoD GEO optical-comms demo contract, and a first commercial GEO relay satellite ordered from Swissto12 in March 2026 (Via Satellite). Slower but better-funded than Western startups.

The hedgers. Eric Schmidt took control of Relativity Space explicitly to support orbital datacenters — the reported ~$800M of fresh funding traces to a single Bloomberg origin and should be treated as unconfirmed (prior raises of $1.2–1.34B are well documented). Microsoft positions Azure as the control plane for others' hardware; AWS's CEO calls orbital DCs "pretty far" out while an AWS server blade flies on Starcloud-2 anyway. Thales Alenia's EU-funded ASCEND study puts Europe's first space datacenter at 2036 — a decade behind the US private sector.

While Western filings stack up, the only multi-satellite orbital-compute constellation in orbit is Chinese. ADA Space (Guoxing Aerospace) and Zhejiang Lab launched the first 12 satellites of the "Three-Body Computing Constellation" (Star-Compute program) on May 14, 2025 — reported at 744 TOPS of AI compute per satellite, 100 Gbps inter-satellite laser links and ~30 TB of onboard storage, with a roadmap of 2,800 satellites targeting ~1,000 POPS of aggregate compute, and Alibaba Cloud-ecosystem partners attached. Western trade coverage is thin and the specifications are program-stated rather than independently measured, so we carry the numbers with that caveat — but the existence of the constellation is not in question, and it predates every funded Western pure-play's first multi-satellite deployment.

The analytical significance is less the TOPS count than the demonstration order: China fielded a networked compute constellation while the US fielded filings. For sovereignty-driven buyers (the customer class Madari Space is courting in the UAE, and Space Compass in Japan), the existence proof matters more than the unit economics. It also means any "first mover" claim in a Western pitch deck is already false.

Working unit: a 4-server AI rack ≈ 41 kW IT load plus ~7% overhead ⇒ a 44 kW rack-equivalent (one NVIDIA H100 SXM = 700 W TDP per NVIDIA's spec; a GB200 NVL72 rack ≈ 120 kW).

That is the entire bull case in one line: orbital wins only if (a) Starship reaches ≤ ~$200–500/kg, (b) hardware survives ≥5 years (Google's Trillium TID data supports survivability), and (c) whole-system mass lands at ≤ ~15 kg/kW. Miss any one and terrestrial wins — which matches Google's own "roughly comparable per kW-year" framing and Luminix's <15% probability of economic viability at scale by 2029–31 (Luminix). Bandwidth is the quieter constraint: 800 Gbps optical pairs exist, but space-to-ground remains the chokepoint, which is why near-term revenue is in-space data processing (EO inference) and batch training, not latency-sensitive serving. Vendor TAMs — BIS Research's $39.1B by 2035 and the like — are marketing-grade; we cite none of them as planning inputs.

AstraVeris underwrites 661 terrestrial datacenter projects with full pro formas at /datacenter/. Mapping that machinery onto orbit:

• Transfers — the pro forma skeleton. Capex → revenue/kW → stabilized cash yield → sensitivity grid works in any gravity. JLL lease $/kW-month comps become the price ceiling: orbital compute must clear at or below terrestrial GPU-hour pricing except for sovereignty or in-space-locality premiums. Crusoe's early-2027 orbital capacity pricing will print the first real $/kW analog — we will capture it the day it exists.
• Transfers — escalators and vintages. BLS PPI and Turner BCI escalators repurpose for spacecraft hardware indexation, and GPU generations map exactly onto our vintage-history concept. The twist is obsolescence-in-transit: Starcloud-1's H100 was already three generations stale at the Blackwell announcement, because build-to-launch runs 12–24 months. That is a structural orbital handicap, not a timing accident.
• Breaks — cap rates. There is no stabilized-NOI market, no resale, no residual: the asset deorbits at end of life (5–7 yr bus, 3–4 yr GPU relevance, versus a 30-year terrestrial shell). Underwrite as a full-depreciation DCF/IRR with a deorbit cost line — never an exit cap. Our SEC-EDGAR REIT cap-rate machinery has no orbital analog this decade.
• Breaks — the moat and the hazards. Land, interconnection queues and power contracts are replaced by launch-capacity access and FCC/ITU filings as the scarce assets. FEMA flood/fire overlays are replaced by debris-flux and radiation-environment maps. "Occupancy" becomes duty cycle: eclipse fraction, safe-mode events, radiation scrubs.

Our house position: track orbital DC as a watchlist vertical with cadence-style leading indicators (compute-payload launches, FCC filings, laser-terminal orders), and do not extend datacenter pro formas to orbit until the first arm's-length $/kW-month contract prints. Earliest candidate: Crusoe, early 2027.

Insurable now: launch plus first-year in-orbit all-risk on conventional terms, inside the existing ~$500–600M/yr space premium pool (Orbital Radar, Howden). An orbital DC node is underwritable today as an ordinary satellite hull — which is exactly how the Axiom/Kepler nodes and Starcloud-2 can be covered.

Not insurable now: compute-SLA or business-interruption cover on orbital capacity (no actuarial base for radiation-induced silent data corruption or GPU degradation rates); GPU residual value; constellation-scale aggregation — a 10,000+ satellite compute fleet exceeds per-risk market capacity of roughly $500M and concentrates debris-cascade tail risk; and third-party liability beyond mandatory FAA/FCC minimums.

Where the market goes: blanket constellation policies, parametric triggers, and SSA data-sharing incentives — Lloyd's and AXA XL already waive deductibles for operators sharing tracking data (NewSpace Economy). The megaprojects (SpaceX, Google) will self-insure as SpaceX always has, so insured premium volume will badly lag deployed value — the commercial opening is parametric radiation-event and launch-slip products for tier-2 pure-plays that must show insurance to lenders. The line item every orbital pro forma must carry: space hull rates run ~4–12% of insured value at launch versus ~0.2–0.5%/yr for terrestrial DC property — a 100x+ difference.