Can you actually deliver the compute we need?

Workload-first. Numbers, not vibes. ~3 minutes. Deep-dive at /explore (the parameter playground).

Who this page is for

Three buyer archetypes who would pay for orbital compute. Fictional names — typology, not named prospects.

Maria

CTO at a European hyperscaler

Goal: Hit 2027 EU AI Act compliance with EU-resident inference for regulated workloads.
Pain: Existing terrestrial DCs depend on a non-EU power grid + non-EU cooling water — the 4-component sovereignty test fails on the supply-chain axis.
Objection: But you don't have a constellation yet — show me the architecture lock and the build path.

“I need sovereign inference for EU customers by 2027, and ground-only doesn't cut it.”

Thomas

AI infrastructure lead at a European sovereign-AI programme

Goal: Deploy sovereign AI capacity that scales beyond terrestrial GPU supply without grid expansion.
Pain: GPU supply is constrained and cooling is a bottleneck on the ground; orbital compute is the only path to 100K+ TFLOPS of sovereign capacity.
Objection: We need 100 MW equivalent, not 120 satellites — show me the cost-per-TFLOP math.

“The 100 TFLOPS/sat at 600 km density is the right envelope — we just need it in 5 years, not 10.”

Sophie

Head of research compute at a European AI lab

Goal: Burst-compute for large training runs (10K-GPU days per month) on a flexible schedule.
Pain: Hyperscaler GPU queues are 6+ months out and pricing doesn't favour burst workloads; we lose 6 months of training time waiting.
Objection: Orbital latency is too high for synchronous distributed training — we need sub-ms all-reduce.

“If you can give me 50K TFLOPS for 30 days at a time, my training schedule accelerates by 6 months.”

Where we are

Pre-architecture. The roadmap, the parameters, the simulation. Not the constellation yet.

We are at the post-research, pre-architecture stage. The deep-dive is done. The competitive landscape is mapped. The regulatory path is sketched. The engineering trade-offs are understood. What we can deliver today:

A workload-fit study — for your training run, inference workload, or batch-job shape, with concrete numbers (compute, latency, bandwidth, $/hour equivalent) under a realistic orbital-compute scenario.
An SLA-target document — what the constellation can realistically commit to (latency, uptime, sovereignty, data-residency) for a defined anchor workload, with the trade-offs explicit.
A build-path memo — first flight → commercial service, with the milestones, dependencies, and decision points. You see the same roadmap we see.

What we can't deliver today: 100 MW of orbital compute next quarter. That comes after the architecture lock and the anchor customer. /story has the full timeline.

The SLA we are designing toward

Headline numbers. Every one of them is a slider on /explore.

Latency (LEO): 4–12 ms round-trip for LEO-ground-station hops. That's an order of magnitude better than GEO (~600 ms) and competitive with regional terrestrial fibre for inter-continental workloads.
Compute density: ~50–200 TFLOPS per satellite (radiation-hardened, inference-optimised; training runs use multi-sat sharding). Per-constellation compute scales linearly with the satellite count.
Coverage: LEO is non-stationary; a constellation of ~120 sats at 600 km gives continuous coverage above 50° latitude for mid-latitude customer sites. The /explore playground shows the coverage curve.
Bandwidth: downlink 0.1–10 Gbps per satellite (optical inter-sat link + RF ground link); uplink 0.01–1 Gbps. Per-day data volume per satellite: hundreds of GB at the high end of the curve.
EU sovereignty: ground stations in EU jurisdictions, EU-controlled operations, EU-compliant data residency. For European-AI and EU-public-sector customers this is the headline requirement, not a nice-to-have.

Why orbital beats ground for the right workload

The binding constraint on the ground is not the binding constraint in space.

Solar is roughly 8× more productive above the atmosphere — no clouds, no night, no atmospheric attenuation. The energy budget is set by the solar array area, not by the local grid.
Heat sheds by radiation — no fans, no chillers, no water. The thermal design is passive; reliability scales.
Vacuum is free insulation — no convective losses, no corrosion. Hardware MTBF in vacuum is decades, not years.
The workload matters. Orbital compute is not a replacement for ground data centres — it's a complement. Latency-sensitive regional inference stays on the ground. Inter-continental AI-training sharding, sovereign inference at the edge of EU jurisdiction, and burst-compute for time-sensitive research workloads move up. The workload-fit conversation is the whole game.

Play with the parameters

/explore — 4 sliders (satellites, altitude, compute per sat, ground stations) wired to a live latency chart and a scenario summary. Numbers, not slides.

See who else is in this space

/competitors — 11 verified profiles of the actual players (Axiom, Starcloud, Kepler, Google Project Suncatcher, OHB, etc.). Where we fit, where we don't, who we integrate with.

Book a 30-min call (workload) Open the playground

Or read the long-form at /story, see the competitive landscape at /competitors, and the technical deep-dive at /tech.