Zoom into any city on Earth, as close to live as possible. Here's the full engineering blueprint — sensors, orbits, AI, downlink, and cost.
A single satellite can't watch one place continuously. Physics won't allow it.
How often a satellite flies over the same point. This determines how "live" your view is.
| Altitude | Orbital Period | Ground Speed | Passes/Day over One City | Resolution (50cm lens) |
|---|---|---|---|---|
| 300 km (VLEO) | ~90 min | 28,000 km/h | 16× | ~15cm |
| 500 km (LEO standard) | ~95 min | 27,400 km/h | ~14× | ~25cm |
| 600 km | ~97 min | 27,000 km/h | ~13× | ~30cm |
| 1,200 km (MEO) | ~110 min | 26,000 km/h | ~12× | ~60cm |
| 35,786 km (GEO) | 24 hrs | Stationary | Continuous BUT | 5–50m only |
Five engineering layers, each building on the last. Skip one and the system breaks.
The physical satellite: structure, power (solar), propulsion, attitude control + the imaging payload. Electro-optical for daytime, SAR radar for night/cloud, or both. Off-the-shelf bus from ISISPACE, Surrey Satellite, or build in-house for cost at scale.
Getting data off the satellite fast. X-band radio gives ~300 Mbps. Optical laser terminals give 10–100 Gbps but require precise pointing. Global ground station network (or use AWS Ground Station / KSAT as-a-service) to minimise wait time between passes and downlink.
Raw video is too large to send down every pass. AI runs on the satellite itself — detects objects, flags changes, compresses to event clips. NVIDIA Jetson-class GPUs are now radiation-hardened for space. Only send what matters: "3 new vehicles at this location since last pass."
Imagery lands at ground station, ingested to cloud (AWS/Azure/GCP), orthorectified (corrected for satellite angle and Earth curvature), indexed by location and time, stored in a geospatial database (PostGIS, Google Earth Engine). Automated pipeline — zero manual steps.
The dashboard where users log in, draw a box on a map, task a satellite, receive imagery or alerts. Map tiles served via MapboxGL or Cesium.js. AI change detection overlay. Tasking API for programmatic access. This layer is the fastest to build — it's just a web app talking to a geospatial API.
Every component, what it does, and what to use.
Use a Cassegrain telescope design — compact, large aperture, ideal for high-res imaging from a small sat. 50cm mirror → 25cm ground resolution from 500km. 70cm mirror → 17cm resolution. Paired with a TDI (Time Delay Integration) sensor — the satellite moves, the sensor rows shift in sync, effectively giving a longer exposure without blur. Used by WorldView, Pleiades, SPOT.
Synthetic Aperture Radar emits microwave pulses and measures the reflection. Penetrates clouds, fog, smoke, works at night. Resolution down to 25cm (Umbra achieves this). Trade-off: active system uses more power, data is harder to interpret visually — AI needed to make SAR imagery human-readable. Best approach: mix optical + SAR sats in the same constellation.
Install radiation-hardened GPU on each satellite. Run YOLO or RT-DETR object detection on each frame. Classify: vehicles, ships, aircraft, buildings, military equipment. Generate change maps vs. previous pass. Compress output to 1–5% of raw data size. Transmit only the change events and tagged clips. This alone makes the system 50x more efficient.
Like Starlink v2 — satellites pass data to each other via laser instead of waiting for a ground station pass. A satellite over the ocean can relay data through the constellation to one near a ground station, cutting latency from hours to minutes. Required for a truly global near-real-time system. Tesat and Mynaric make the laser terminal hardware.
You need a downlink station within range every orbit (every ~95 min). Strategy: own 3–4 polar ground stations (Svalbard, Alaska, Antarctica, Tierra del Fuego) which see every polar-orbit satellite twice per orbit, plus supplement with AWS Ground Station or KSAT as-a-service for lower latitudes. Target: no satellite goes more than 20 minutes without a downlink opportunity.
Imagery arrives raw — it needs orthorectification (correcting for satellite angle, Earth curvature, terrain). Stack: GDAL for raster processing, PostGIS for spatial indexing, STAC (SpatioTemporal Asset Catalog) for querying imagery by location+time, Titiler for serving map tiles on demand. All runs on Kubernetes — fully automated from ingestion to serving in under 10 minutes.
Frontend: React + Mapbox GL JS or Cesium.js for 3D globe. User draws AOI (Area of Interest) on map, system schedules next satellite pass, delivers result. AI change detection layer highlights what's new vs last pass. WebSocket push notifications: "New activity detected at your watched location." REST + GraphQL API for programmatic access — integrate with Sovereign Foundry or Maven.
Commercial players you can partner with, license data from, or compete against.
Three levels — from data reseller to full sovereign constellation owner.
| Approach | Build Cost | Timeline | Revisit | Resolution | Who Owns Hardware |
|---|---|---|---|---|---|
| Tier 0 — Data Reseller + AI Layer | $5M–$20M | 3–6 months | 90 min–daily | 25cm–1m | Planet / BlackSky / Umbra |
| Tier 1 — 10-Sat Starter Constellation | $80M–$200M | 18–24 months | 2–3 hours | 50cm | You own + augment with commercial |
| Tier 2 — 50-Sat Regional Coverage | $300M–$800M | 2–3 years | 15–30 min | 25–50cm | You own |
| Tier 3 — 200-Sat Global System | $1.5B–$3B | 3–5 years | 3–8 min | 25cm optical + SAR | You own |
| Tier 4 — 500+ Near Real-Time | $4B–$10B | 5–7 years | <2 min | 15–25cm optical + SAR + IR | You own — near-NRO capability |
You don't need to own satellites to start. License data from existing players, build the sovereign AI intelligence layer on top, and sell access to governments who can't trust US providers with their national security imagery. That's a $5M–$20M build that generates $100M+ revenue. Use that cash to build your own constellation progressively.