The unwatched world
The stranded asset: the installed base nobody watches, why labor can’t close the gap, and what “watched” would even mean at fleet scale.
1.5Benterprise cameras
7T hrs/yrfootage produced
<1%ever watched
Worldprint · Private working papers · 2026
There are 1.5 billion enterprise cameras in the world producing roughly 7 trillion hours of video a year, and less than 1% of it is ever watched. That footage is a stranded asset — not because the cameras are missing, but because understanding has been too expensive: in a real-time video pipeline, embedding generation has been ~90% of the cost. That cost is collapsing on every axis at once — inference roughly 10× cheaper per year at constant capability, Blackwell 10–35× cheaper per token, video understanding now cents per hour, vector storage down 4–32× via quantization and disk indexes. Meanwhile the labor that used to compensate — guards, reviewers, operators — has hit a wall: zero projected growth in the US guard workforce, 100–200% turnover, 65% of hotels understaffed. Where the falling cost curve crosses the rising labor wall — 2027–2030 — every industry with cameras gets a new primitive: a searchable, permissioned, encrypted record of everything that happened in a physical space, queryable across video, audio, and business data in one question. NVIDIA’s VSS blueprint deliberately leaves the retrieval layer open for partners; Miriel + CyborgDB is the proven fill. This site is that go-to-market artifact.
Every number above is sourced or labeled in the documents below. Statistics on this hub are previews of what each paper argues; the papers carry the citations.
The record is a data problem before it is an AI problem: video-derived perception, badge and sensor events, and business tables have to live in one queryable layer. Each partner owns a load-bearing piece of it.
The perception pipelines (detection, OCR, faces, embeddings — each stage configurable, doc 03 §3), the episode model of events, one identity per object across cameras, microphones, and sensor tables, and the unified data layer that manages object storage, vector storage, time-series and SQL data together — queryable with one read-only SQL statement across all of them first-party.
The vector database inside that data layer. Every embedding is encrypted at rest and in use — search happens without decrypting — and it is where the stack’s strongest proof point lives: 100 billion encrypted vectors at 199 ms p50, flat from 5B to 100B (doc 03 §4). Cyborg is a separate company from Miriel — this partnership needs both.
The vertical operator for the anchor vertical: guard operations, monitoring-center workflow, and SmartSentinel as the emerging reference implementation (security brief).
The shared argument, told once; verticals hang off it as siblings. Four ways in: from NVIDIA, start with 03 then 04 · running an industry with cameras, go straight to your vertical brief · want the numbers, the explorer in 02 and the calculator in 03 expose every assumption as a slider · hunting weak points, every document ends with its counter-evidence.
The stranded asset: the installed base nobody watches, why labor can’t close the gap, and what “watched” would even mean at fleet scale.
Why now: the four simultaneous cost collapses, the honest counter-evidence, and the crossover analysis. Anchor interactive: the cost-curve explorer.
The Miriel × Metropolis architecture: configurable perception pipelines, event-driven writes, sensor fusion in one query layer, CyborgDB encryption-in-use, and the fleet-economics calculator.
The Metropolis partner motion — LaunchPad demo → POC → Fleet Command — the trilateral (Miriel + Cyborg + 9Squared + SIs), deal shapes, and the joint proof plan.
One template, stamped per industry: killer stats, a day in the footage, the economics with every assumption exposed, and a reference deployment.
The anchor vertical: false-alarm economics, the guard-labor arbitrage, the fusion timeline, and SmartSentinel as the emerging reference implementation.
The strongest third-party evidence base: real-time crime centers, case-solvability lifts, and encryption + audit trail as the deployment unlock.
Aligned with NVIDIA’s warehouse blueprint, with visceral numbers on injuries, cargo theft, and shrink — and a deliberately conservative loss-avoidance model.
NVIDIA’s thinnest public vertical: unattended sites, portable fleets, equipment theft that mostly never comes back, and the Fatal Four as the safety case.
The cameras are already wall-to-wall — casinos are the original surveillance-dense environment. What’s missing is the record: internal theft found in days, not never.
Security, municipal, and warehouse are the launch trio; construction and hospitality are the fast-follows, chosen because they are NVIDIA’s two thinnest public verticals. Hub stats are previews; each brief carries inline sources for every figure.
Three artifacts that demonstrate instead of arguing.
24 slides, keyboard-driven: the unwatched world → the cost curve → the record → vertical proof points → the joint proof ask. Arrow keys or space; O for the overview grid.
A live-feeling console over a fictional mixed-zone campus: camera wall with detection overlays, perception stream, “ask the footage” search with evidence packages, and per-event drill-downs. Every vertical brief points at a zone of it.
Miriel’s live demo, captured for the site: the real product UI over real cameras — continuous perception, episodes with dwell, persistent object memory, evidence-backed answers. The atom the operator console scales up to a campus.
The corpus is a dependency graph, not a list. These are the joins that carry weight — the table fills in as the documents ship.
| From | To | What travels across the join |
|---|---|---|
| 02 · Cost curve | Every vertical brief | The viability crossover dates in each vertical are computed from doc 02’s decline rates — change the slider there and the vertical ROI moves here. |
| 03 · The record | 02 · Cost curve | The fleet-economics calculator prices the architecture doc 03 specifies, using the unit costs doc 02 establishes. |
| 01 · Unwatched world | 03 · The record | “What happened, not frames”: doc 01’s event-driven model of watching is the design goal doc 03’s write policy implements. |
| 04 · Partnership | 03 · The record | The joint proof plan’s metrics (cameras/GPU, recall@10, p50/p99, GB/day) are the architecture’s published-vs-claimed boundary made testable. |
| 03 · The record | Every vertical brief | Each brief’s record-cost default is the fleet calculator’s output for that
vertical’s scene-mix preset (?preset=security|municipal|warehouse) —
verified by scripts/qa-math.js; change one side and the tripwire
fails. |
| 02 · Cost curve | 03 · The record | “Queries without an LLM”: doc 02 §4’s tailgate join is priced there and mechanized in doc 03 §5’s sensor-fusion table — SQL for structure, vectors for similarity, LLM for language. |
| Vertical briefs | Operator console | The console’s scripted storylines are the briefs’ scenarios made clickable — the near-miss and vanished pallet (warehouse §2), the impossible badge swipe (security §2, doc 03 §5), the graduated perimeter response (security §4). |
The floating assistant holds the full text of every published document in context. It knows which page and section you are reading, cites the passages it draws on as live links, and says when the documents don’t settle a question. If no model is configured behind the site, it degrades to a keyword search over the same corpus.