--- title: "Edge compute & AI, explained | Nexus Observatory" description: "Interactive explainer: edge compute & ai in 3D, chapter by chapter: raw video outruns any backhaul, ship the answer, not the pixels…" url: https://observatory.global/edge-ai/explainer/ publisher: Nexus Observatory Inc. updated: 2026-10-28 --- 8 × **2.99 Gb/s** **Cloud region** Results · **41.0 kb/s** 0 km 1,000 km 2,000 km 3,000 km 4,000 km Region · **1,200 km** Deadline reach · **1,224 km** **Edge** Queue empty Firmware 27d091eb Bootloader 96640d72 Kernel c2bb900b Initrd 0db7606b Runtime 97d6a78c Model c8c2e6e5 Verifier · **quote matches** **Signed plan** Wave 1 · **5 nodes** 122 m AGL 01 02 03 04 05 06 07 8 × 2.99 Gb/s Edge · Explainer Interactive # Compute where the data is born. Instruments flood a site with raw data. A data centre in a 20 ft container runs the models beside them and sends only results upstream. Seven decisions, each worked out live. **Container**: 20 ft · 6 racks **Cryptography**: SHA-256 · ECDSA P-256 **Chapters**: 7 **01** Raw is heavy **02** Inference at the edge **03** Latency budget **04** Store and forward **05** Attestation **06** Signed plan **07** Physical AI **01** · Raw is heavy Bitrate ## Raw video outruns any backhaul. An uncompressed frame is width × height × bits per pixel; 8-bit 4:2:0 colour costs 12 bits a pixel. Multiply by the frame rate and the cameras on site and the stream dwarfs a satellite or microwave link long before it reaches a region. Resolution Frame rate **Per camera, raw**: 2.99 Gb/s **Site total, raw**: 23.89 Gb/s **Per day**: 257.99 TB **Frames to analyse each second**: 240 **02** · Inference at the edge Results ## Ship the answer, not the pixels. The GPU servers in the container's racks turn frames into detections. A binary record of class (2 bytes), confidence (4), box (4 × 4), timestamp (8) and camera (2) is 32 bytes; that, not the frame, is what leaves the site. Record encoding **Record size**: 32 bytes **Results upstream**: 41.0 kb/s **Reduction against raw**: 583,200 : 1 **Results per day**: 442.4 MB **03** · Latency budget Round trip ## Light in fibre sets the floor. Glass slows light to about c/1.47, close to 4.9 µs a kilometre. A control loop that must answer within its deadline pays that twice, out and back, before a single switch, queue or model runs. Distance alone can rule a region out. Round trip against the deadline Edge, 0.5 km 8.0 ms Fibre round trip 11.8 ms Cloud, total 19.8 ms Deadline 20 ms **Farthest region that still meets it**: 1,224 km **04** · Store and forward Outage ## When the link drops, the site keeps working. Inference never stops for the backhaul. Results queue on local disk in order, and when the link returns the backlog drains at whatever capacity the live stream leaves free: drain time = backlog ÷ (link − live rate). Backhaul Disk for the queue **Live rate upstream**: 41.0 kb/s **Backlog at the end of the outage**: 663.6 MB **Drain time after restore**: 4.4 min **Outage the disk can hold**: 12.4 years **05** · Attestation SHA-256 ## Prove what booted before trusting what it says. Each boot stage hashes the next before handing over control and extends a platform configuration register: PCR = SHA-256(PCR ‖ digest). A register can only be extended, never set, so its final value proves the whole chain. A remote verifier compares a signed quote of it with the golden value. Tamper with a stage and watch every later value change. Tamper with a stage - **Firmware** matches digest 32d4c441c39f15b1… · PCR **27d091eb198d5e9d…** - **Bootloader** matches digest e63b965dca79e24e… · PCR **96640d7272ff2502…** - **Kernel** matches digest 707e9ff01a3bba3e… · PCR **c2bb900b5b91a33c…** - **Initrd** matches digest 9ab8295753874ac4… · PCR **0db7606be3601b24…** - **Runtime** matches digest b5575b146d0497ee… · PCR **97d6a78c7314e81d…** - **Model** matches digest 22a2043d13a5e837… · PCR **c8c2e6e596843a93…** **Verifier**: matches the golden value **06** · Signed plan ECDSA P-256 ## A model update is a signed plan, rolled out in waves. The plan names the model by its digest and the waves of the fleet that take it. It is signed with ECDSA over P-256 and SHA-256; every node verifies the signature against a key it already trusts before it fetches a byte. Change one character after signing and no node moves. Plan in transit {"model":"detector","digest":"sha256:80b1542df0ab…","waves":\[{"share":10},{"share":100}\]} **Signature**: 8002960c2e96fc76… · 64 bytes **Verification on each node**: valid **Nodes updating now**: 5 of 50 **07** · Physical AI Survey ## A drone that flies its own survey. The same box plans and checks a survey flight. Ground sample distance is sensor width × altitude ÷ (focal length × image width); overlap sets the spacing of the lines. In Canada, Part IX of the Canadian Aviation Regulations caps routine flight at 122 m above ground. Camera **Ground sample distance**: 2.74 cm/px **Footprint per image**: 150 × 100 m **Lines over 600 × 400 m**: 10 · 44 m apart **Images at 75 % forward overlap**: 257 **Flight at 10 m/s**: 6.40 km · 10.7 min **Altitude against 122 m**: 22 m under [Nexus Observatory](https://observatory.global/)**Edge compute & AI** . Inference beside the instrument.[Talk to us](https://observatory.global/contact/) Edge compute & AI - [AI tower](https://observatory.global/edge-ai/ai-tower/) - [MEC](https://observatory.global/edge-ai/mec/) - [Edge AI](https://observatory.global/edge-ai/edge-ai/) - [Scientific compute](https://observatory.global/edge-ai/scientific-compute/) - [Physical AI](https://observatory.global/edge-ai/physical-ai/) Technologies - [Edge tiers and multi-access edge computing](https://observatory.global/technologies/edge/edge-tiers-and-multi-access-edge-computing/) - [Modular edge data centres and the AI tower in a box](https://observatory.global/technologies/edge/modular-edge-data-centres-and-the-ai-tower-in-a-box/) - [Edge inference, model serving and optimisation](https://observatory.global/technologies/ai/edge-inference-model-serving-and-optimisation/) - [Foundation models for Earth observation and geospatial](https://observatory.global/technologies/ai/foundation-models-for-earth-observation-and-geospatial/) Industries - [Public safety](https://observatory.global/industries/public-safety/) - [Civil resilience](https://observatory.global/industries/resilience/) - [Nature](https://observatory.global/industries/nature/) - [Oceans](https://observatory.global/industries/oceans/) - [Energy](https://observatory.global/industries/energy/) - [Telecom](https://observatory.global/industries/telecom/) - [Transport](https://observatory.global/industries/transport/) Service platforms - [Observatory Protocol](https://observatory.global/protocol/) - [Protocol & Grid](https://observatory.global/grid-services/) - [Software](https://observatory.global/software/) - [Hardware & power](https://observatory.global/hardware/) - [Sensors & instruments](https://observatory.global/sensors/) - [Connectivity](https://observatory.global/connectivity/) - [Earth observation & GeoAI](https://observatory.global/earth-observation/) © 2026 Nexus Observatory Inc. [Privacy](https://observatory.global/legal/privacy/) [Terms](https://observatory.global/legal/terms/) [Accessibility](https://observatory.global/legal/accessibility/) [Cookie settings](https://observatory.global/privacy-choices/) [Data credits](https://observatory.global/credits/)