Sovereign Orbital AI inference

The hardware is deliberately unexotic. A satellite carrying GPUs and a commercial optical terminal is a configuration a competitor could assemble, and Starcloud already flies something close to it.[Starcloud-2] Nothing in the bill of materials keeps anyone out.

Defensibility comes instead from jurisdiction, regulatory position, and the customer relationships that follow. That makes compliance and procurement work as load-bearing as the engineering, and it means execution speed matters more than technical secrecy.

A single satellite is over a given ground station for minutes a day.[arXiv 2604.27197] Academic work on orbital data centers is blunt about this. Geometry and weather set a hard ceiling on sustained space-to-ground throughput, and high availability forces a geographically spread ground segment whose cost scales with the bandwidth you need.[arXiv 2604.27197]

So availability is a function of how many satellites and how many ground stations we own. Both cost money. A customer who wants a defined turnaround window is buying our constellation size, and pricing has to reflect what a given fleet can actually deliver.

Reusable Falcon 9 runs about $3,600/kg to LEO today.[Google §4.4] Google treats $200/kg as the point where launch stops dominating cost.[Google §2.4] They model a learning curve of 20% per doubling that reaches under $200/kg by roughly 2035, needing about 1,800 cumulative Starship launches.[Google §4.4] Their aggressive reuse scenario reaches $15 to $38/kg on 100× component reuse that nobody has demonstrated.[Google §2.4]

Starcloud's whitepaper assumes $30 to $50/kg is achievable now and falling toward $10/kg.[Starcloud WP] That figure comes from their own marketing. Public sources spread across a factor of ten to a hundred on the single input our cost case rests on. We cannot move that number.

Google's radiation data is the best public evidence available and it is encouraging. Two caveats carry real weight. On-chip memory was the weak point, showing irregularities at only about 2.7× the required dose.[Google §2.3] And that is radiation alone. Nobody has published what five years of thermal cycling and mechanical stress does to a GPU in orbit.

Two disposal models are open. Dead hardware deorbits and burns up with nothing recovered, or modules are built for swap-out and partial salvage, which is the route Starcloud designs around.[Starcloud WP] The choice moves the economics materially and needs settling early in the design.

Small satellites are worse economically, and the effect is severe. Avionics, propulsion, comms and structure do not shrink with power, so fixed overhead dominates a small bus. Google's own comparison puts Starlink v2-mini at about 20.5 kg/kW, Starlink v1 at 37, OneWeb at 188 and Iridium NEXT at 430.[Google, Table 1] That is a factor of twenty between best and worst. Starcloud makes the same argument about shielding, which scales with surface area while compute scales with volume.[Starcloud WP]

The figure to minimise is kilograms per kilowatt. Below roughly 10 kW that diverges sharply from total mass, and fixed overhead starts eating the launch budget. This is what sets the eight to sixteen GPU range, and it is an economic constraint rather than a thermal one.[own estimate]

There is no convection in vacuum. Heat leaves by radiation only. Junction limits of around 85°C[McCalip] cap how much compute a given radiator area can carry, and that cap is what sets our GPUs per satellite.[McCalip] Starcloud names the deployable radiator as one of two unsolved technical hurdles on their public roadmap.[SpaceNews] We face the same physics with a smaller team.

Mbryonics sits in Galway, Ireland. They sell to ESA, DARPA, Airbus Defence and Space, and the European Defence Agency.[Mbryonics] Their long-haul specifications suit relay and ground links directly.

• STARCOM. Optical inter-satellite and Earth-link terminals. Up to 400 Gbps by 2027, link distances beyond 80,000 km. Sized for relay.[STARCOM]
• STARLIGHT. Companion modem, amplifier, and antenna control unit. Full assembly under 14 kg.[STARLIGHT]
• STARGATE. Their optical ground station, claiming roughly a thousand times the throughput of RF ground links. A direct candidate for our ground segment.[STARGATE]
• TeraBIX. Radiation-hardened photonic transceiver for chip-to-chip links, under 5 mW/Gb/s. Useful inside the satellite.[TeraBIX]

One open question for them. Their forward error correction is tuned for 80,000 km range, and at relay distances the FEC delay could exceed the flight time.[STARCOM] That number is absent from the public spec sheet and is the first thing to ask.

• Launch licensing across EU and ESA bodies, plus any non-EU launch partner.
• Ground station siting and licensing across several member states, which is a substantial part of the build.
• Orbital slot coordination. Contested and worsening. Starcloud has filed for 88,000 satellites. SpaceX for up to a million.[GeekWire]
• Insurance for launch and on-orbit hardware, which no public cost model in this field includes.

The Tech Sovereignty Package and CADA stopped being proposals in June 2026.[EC] They are moving through Parliament now. The four-tier sovereignty framework will score vendors in public procurement.[CADA] An orbital node satisfies the hardest tier by construction. It sits beyond every national territory, which settles the question of applicable law in a way even EU-headquartered cloud providers still struggle to answer.

The argument runs into the supply chain too. Buying optical terminals and ground stations from an Irish vendor already trusted by ESA puts the sovereignty claim into the bill of materials, where a procurement officer can see it.

• Defense and ISR. Dual-use framing opens the EU Defence Fund and national innovation arms, which is non-dilutive capital.
• Regulated finance. GDPR-bound inference, where sitting outside any jurisdiction is a novel legal position worth testing with counsel early.
• Healthcare and pharma. Clinical and genomic inference. These sectors already pay a premium to OVHcloud and Scaleway for sovereign cloud.
• Earth observation operators. They generate terabytes a day and downlink is their bottleneck.[Starcloud-2] Processing in orbit serves them at today's launch prices. Starcloud targets this market explicitly.[Starcloud-2]

All of them treat compute as the product and are building toward scale and eventual cost parity. Jurisdiction remains unclaimed, and the lane is narrow.

This market probably narrows to three to five players. Orbital slots, spectrum coordination, and launch manifest access are scarce and sequential. The scarcest resource is regulatory trust. CADA's sovereignty framework is not expected to be final before the end of 2027.[CADA] Vendors in the room now, with flying hardware, help define what sovereign assurance means for orbital infrastructure. Later entrants cannot buy that position with capital alone.

The arithmetic.[own estimate] A Starlink v2-mini masses 575 kg for 28 kW.[Google, Table 1] We need about 12 kW, so a scaled bus plus GPU chassis and radiator plausibly lands between 350 and 450 kg. Take 400 kg. At $3,600/kg that is roughly $1.4M to launch. Call hardware and integration another $3M. Eight GPUs at 60% utilisation deliver roughly 42,000 GPU-hours a year. Five years against $4.4M of capex needs about $21 per GPU-hour before operating costs.

Mass is the most sensitive input in this calculation. It is scaled from a Starlink bus rather than costed from a detailed design, and every 50 kg moves the break-even by roughly a dollar. A bottoms-up mass budget is an early priority.

Now compare that to the market. Cross-provider median for a B200 sits around $6.25 per GPU-hour, but the spread is enormous, running from $3.75 at the cheapest neocloud to $16.11 on Google Cloud[GPU price index][28 providers] and past $27 on Azure managed instances. Hyperscaler rates already carry a large premium for SLAs, support, and integration.

That reframes the ask. The price sits at roughly 1.3 times what AWS charges a European buyer today, in exchange for a jurisdictional guarantee AWS cannot offer at any price. If launch reaches $200/kg our break-even lands at about $13, which is inside today's hyperscaler range. That is the single figure to validate with a customer before raising anything.

Earth observation operators generate terabytes of raw imagery a day and cannot get it down fast enough. Processing in orbit and downlinking only the result removes their bottleneck. Their reference point is the downlink capacity they lack, which makes this a capability sale rather than a price comparison.[Starcloud-2]

This market is live at today's launch prices and independent of CADA's timeline. It gives the first satellite paying work while the sovereignty motion matures, and it builds ground segment experience on a customer's schedule.

As CADA's assurance tiers become procurement criteria, EU institutions and regulated sectors need vendors who can satisfy the top tier. We sell scheduled and batch inference under a jurisdictional guarantee nobody terrestrial can match.

Pricing here follows compliance rather than compute. A bank barred from running a workload on Azure measures our price against leaving the workload undone, which is where the premium comes from.

• Scheduled inference windows. Guaranteed GPU-hours in defined windows, priced per hour. Suits batch analysis, model serving on a cadence, periodic scoring.
• Sovereign job execution. Data uplinked, processed, results downlinked, nothing retained. Sold on the compliance artefact as much as the compute.
• On-orbit edge processing. Compute sold to other spacecraft operators who need results rather than raw data on the ground.

Every product above is scheduled or batch. Latency depends on when a satellite is overhead, which sets the boundary of what can be sold with a credible SLA.

Dual-use framing opens the EU Defence Fund and national defense innovation arms. Horizon Europe and the EIC fund sovereignty-aligned infrastructure. ESA runs technology development contracts. Every euro from those sources buys hardware without touching the cap table, and a first satellite funded partly by grant capital changes the shape of a seed round considerably.