Emerging · Demand catalyst for the orbital book
AI-grade compute is moving to orbit. The clearing layer is what keeps it from collapsing into two vertical stacks.
Starcloud commercial
Apr 2026
FCC accepts SpaceX ODC
Feb 4 · 2026
2030 addressable spend
$15B
Settlement
T+0 · atomic · USDC
What just happened
Eight months. Eight events. The category is forming in real time.
- November 2025Complete
Starcloud-1 launches with the first NVIDIA H100 in orbit
60-kg satellite, ~100× the GPU compute ever previously operated in space. Establishes the substrate. - December 2025Complete
First in-orbit LLM training and inference
Starcloud runs Google DeepMind's Gemma on a high-power GPU on-orbit and performs in-orbit training of nanoGPT. The category is no longer hypothetical. - January 30, 2026Complete
SpaceX files orbital-DC system with the FCC
Up to one million solar-powered satellites between 500 and 2,000 km altitude — a hundred-fold increase over the current LEO population. - February 3, 2026Complete
Starcloud files for 88,000-satellite orbital DC constellation
FCC application for compute capacity at scale. Day before SpaceX's filing is accepted. - February 4, 2026Complete
FCC Space Bureau accepts SpaceX ODC filing — DA 26-113
Public Notice issued. Comment period opens. Secure World Foundation and 1,000+ public commenters raise debris, Kessler, and light-pollution concerns. - March 30, 2026Complete
Starcloud Series A — $170M at $1.1B valuation
Capital markets confirm the category. NVIDIA backing; Crusoe contracted to run the cloud platform on Starcloud-2. - April 13, 2026In progress
Largest orbital compute cluster open for business
TechCrunch report. Starcloud begins commercial sales of H100-equivalent compute time on-orbit. LOIs in hand. - October 2026In progress
Starcloud-2 launches with NVIDIA Vera Rubin Space Module
Claimed 25× H100 inference performance. Crusoe cloud platform operational. Starcloud says the satellite generates more cash than it costs to build and launch.
The vertical-integration wedge
Two stacks are forming. Everyone outside them needs a neutral CCP.
SpaceX and xAI merged into a single corporate entity. Starcloud signed a multi-gigawatt-of-interest signal from Anthropic and a compute partnership with xAI. The shape of the orbital compute market that follows is two vertical stacks: one Musk-aligned (Starlink V3 backhaul → SpaceX/xAI compute → xAI demand), one NVIDIA-anchored (Starcloud GPU clusters → Crusoe cloud → Anthropic and frontier-lab demand). Both stacks are rational. Both will route compute intra-firm where they can.
Vertical integration handles the demand the integrated firms produce. It does not handle the demand the rest of the industry produces. Hyperscalers procuring orbital compute without committing to either stack, defense and intelligence buyers requiring jurisdictional neutrality, sovereign AI programmes that need attestation rather than vendor trust, ground-station operators settling downlink against compute output, insurers underwriting compute-delivery SLAs — none of them want to live inside one of the two walled gardens.
Wavestar is what they trade on. The clearing layer is the only structure that prevents the orbital compute market from collapsing into two zero-sum vertical stacks. Standardised Compute-Hour contracts, observer-signed delivery, atomic cash-and-resource settlement, and CCP novation — applied to compute the same way they apply to spectrum and downlink.
Vertical integration is rational. Neutral clearing is also rational. They serve different counterparties.
Demand catalyst
Compute pulls every other Wavestar market forward.
Downlink uplift · 2030
+30%
Output-data egress drives downlink bookings beyond observation/comms baseline. Output manifests are the observer signal.
ISL fabric demand · 2030
1 Tbps · per-link
Starlink V3 carrier-grade ISL becomes the AI training fabric. Cross-constellation routing becomes liquidity, not a feature.
Spectrum allocations
Ku · Ka · V
ODC operations consume new authorised allocations at the same rate compute footprint grows. EPFD-modernised regime accommodates.
ISAM servicing market
88k–1M sats
If even a fraction of filed ODC capacity ships, recurring propellant top-up and on-orbit servicing scale by an order of magnitude.
What you trade
1 GPU-equivalent × 1 hour delivered to a named orbit shell, named latency class.
- 01
Unit of trade
1 GPU-equivalent × 1 hour at a named compute-provider DID, named orbit shell, named latency class. GPU-equivalent is normalised to a working-group reference (initial draft: H100-equivalent SXM5). - 02
v1 scope — short-block reservation
Reservations of 1 to 168 hours on inventory the provider already operates. Provider-signed result manifest is the delivery primitive. No new authorisation required. - 03
v2 scope — long-dated forward
Forward contracts on capacity that ships in a future quarter. Activates after the contract spec lands and SEC ATS-N covers compute alongside spectrum. - 04
Latency classes
Training-tolerant (asynchronous, hours-to-days latency acceptable) settles distinctly from inference-bound (round-trip latency to a named ground region). Different liquidity pools, different reference curves. - 05
Delivery attestation
Provider DID signs the start-of-job attestation; observer-signed downlink of the result manifest closes the leg. Multi-observer quorum required for inference-bound contracts where round-trip timing is part of the spec. - 06
Export-control posture
Every match passes a counterparty screen (OFAC, BIS Entity List, equivalents) and a jurisdiction screen for the underlying GPU class. H100 in-orbit jurisdiction is unsettled — Wavestar enforces the most restrictive interpretation per match.
Contract specification
Compute-Hour (CMPT) · Rulebook draft v0.1.
CMPT · Short-block reservation contract
- Ticker
- CMPT · [gpu-class] · [orbit-shell] · [latency] · [hour]
- Unit of trade
- 1 GPU-equivalent × 1 hour at a named provider DID
- Reservation tenor
- 1 hour minimum · 168 hours maximum (v1)
Long-dated forwards (> 168 hours) activate in v2 after working-group spec lands and SEC ATS-N covers compute.
- Tick size
- $0.10 per Compute-Hour
Working-group draft v0.1. Calibrated against $3.00 (GCP) – $6.98 (Azure) terrestrial H100 hourly references.
- Minimum block
- 8 Compute-Hours
Aligned with the smallest training-or-inference job that pencils in delivery cost.
- Delivery attestation
- Provider DID signature + observer-signed result manifest downlink
Inference-bound contracts require multi-observer quorum on round-trip timing. Training-tolerant contracts settle on result-hash reconciliation.
- Settlement
- T+0 · atomic · USDC primary · Fedwire / SWIFT optional
- Initial margin
- 20%
Reflects compute-provider performance risk plus jurisdictional uncertainty during reservation window.
- Variation margin band
- 5%
Daily mark on the cleared curve per GPU class and orbit shell.
- Prohibited pairs
- OFAC · BIS Entity List · GPU-class jurisdiction screen
Every counterparty pair is screened on match. Cross-jurisdiction matches subject to the most restrictive applicable export-control regime.
- Position limit
- 5% of cleared open interest per buyer; 25% per provider
- Spec status
- Working group v0.1 · target ratification Q1 2027
Working group convening Starcloud, NVIDIA, Crusoe, hyperscalers, frontier labs. Public draft will be posted to the rulebook.
Counterparties and observers
Providers, buyers, observers, regulators.
- PRV
Compute providers
Starcloud (NVIDIA H100 / Vera Rubin Space Module), SpaceX/xAI (Starlink V3 compute pilot), Axiom Space (orbital DC nodes from Jan 2026), Lonestar Data Holdings (lunar edge), and the wider NVIDIA partner network — Aetherflux, Kepler, Planet Labs, Sophia Space. - BUY
Demand-side buyers
Frontier AI labs (Anthropic published interest; xAI internal), hyperscalers (AWS / Azure / GCP procurement outside their own footprints), defense / intelligence buyers needing jurisdictional neutrality, sovereign AI programmes, brokers aggregating reservations. - CLD
Cloud platform layer
Crusoe runs the cloud platform on Starcloud-2 — the cloud surface customers actually deploy against. Wavestar clears the Compute-Hour underneath the cloud invoice; the cloud platform is the operational delivery layer. - GPU
Silicon class anchors
NVIDIA Space-1 Vera Rubin Space Module is the reference inference fabric. H100 is the training-class baseline. Working group sets the GPU-equivalent normalisation. - MON
Result-manifest observers
Ground-station providers operating downlink for the compute payload sign result-manifest delivery. Independent timing-attestation observers sign round-trip latency for inference-bound contracts. BLS-aggregated quorum forms the delivery proof. - REG
Regulators
FCC Space Bureau (DA 26-113 + Starcloud filing in process), BIS (Oct 2024 space export-control revisions, License Exception CSA), ITAR DDTC (non-cooperative grappling controls), Wassenaar-40 administrations.
Regulatory context
The category is novel. The rulebook is not.
- FCC · DA 26-113 · Feb 4 2026
- SpaceX ODC filing accepted for filing
1M-satellite orbital data-center system. Public Notice issued. Comment period opened. Wavestar tracks the docket; Compute-Hour delivery on SpaceX-side capacity is gated on operational authorisation.
- FCC · Starcloud filing · Feb 3 2026
- 88,000-satellite ODC constellation
Filed by Starcloud. Wavestar's compute book includes Starcloud commercial capacity; new constellation operational dates feed the cleared curves.
- Secure World Foundation · 2026
- 1,000+ public comments · Kessler / scale concerns
SWF and the majority of public commenters opposed the 1M-sat scale citing debris, Kessler syndrome, and light pollution. Wavestar's attestation network mirrors the FCC docket and signs sustainability-relevant facts where requested.
- BIS · Oct 23 2024
- Space-related export-control revisions live
License Exception CSA (Commercial Space Activities) and Wassenaar-40 carve-out reduce friction for cooperative trade. H100 / Blackwell in-orbit jurisdiction remains unsettled — Wavestar applies the most restrictive applicable regime per match.
- ITAR · §121.16
- Non-cooperative grappling controls apply downstream
Compute payloads are not directly defense articles, but their service vehicles (refueling, deorbit, replacement) may be. Wavestar segregates compute settlement from any servicing leg that triggers ITAR.
- EPFD · April 30 2026
- Spectrum modernisation reinforces compute backhaul
FCC's EPFD private-bargaining regime applies to the same Ku/Ka allocations that ODC backhaul will consume. The compute book and the spectrum book share regulatory substrate.
- Sustainability claim parity
- Starcloud (10× lower) vs Saarland (10× higher)
Conflicting carbon-intensity claims in public literature. Wavestar's posture: signed environmental telemetry is part of the result manifest where the buyer requires it; we do not adjudicate the headline claim.
- SEC ATS-N
- v2 long-dated compute forwards activate post-clearance
Filing in flight covers spectrum first; compute forwards layer in once the contract spec is ratified by the working group.
Vertical-integrated orbital compute is the single best argument for neutral clearing the orbital economy will ever produce. We would not have written the page this way if SpaceX had not merged with xAI.
Compute working group · 2026 H2
Compute providers. Hyperscalers. Frontier labs. Ground stations.
Working-group seats for the first cohort drafting the Compute-Hour rulebook. Design-partner reservations clear zero-fee through the first $5M of cleared notional. NVIDIA-class GPU normalisation, observer methodology, and export-control jurisdiction settled in public.