Est. 2026
ZipF - The Physical AI Data Intelligence Company
Language models ate the internet. World models need the street. ZipF licenses a queryable event graph of the physical world: labeled actors, surfaces, weather, time, and place, ranked by how rare they are. The product is not raw footage. It is ground truth a planner can ask for. Spoken Ziff.
Every market figure in this document is verified against published 2025-2026 sources, cited inline and in the appendix. Scenario models are planning constructs, not forecasts.
The boxes are the physical long tail. A handful of scene types fill every public dataset. The rest is where models fail, maps go stale, and insurance is blind. ZipF ranks that tail and licenses it.
Executive Summary
Physical AI is the next foundation-model race, and it is starved in the only place that matters: the messy, long-tail world. Public dumps are full of clear weather, marked lanes, and the same few objects. Planners fail on mixed traffic, two-wheelers cutting across, pedestrians in the carriageway, animals, unmarked geometry, monsoon, night, informal cities, flooded surfaces, and work zones. A lab does not want ten thousand hours of anonymous footage. It wants a query: pedestrians entering the carriageway, rain, night, more than two tracks, this corridor, last thirty days - with time, place, speed, heading, boxes, masks, and a rarity rank.
Whoever owns the labeled event stream of the physical world sits upstream of every model that has to move in it.
ZipF is the company that writes that stream. We produce a Physical Event Graph: every detection, track, surface defect, weather join, and scene event as a structured record with its own coordinates and a telemetry sidecar. Faces and plates are gone before a cut is licensed. The sellable unit is anonymized physical state - not people, not a camera brand. Intelligence is the product. ADAS is the first body that already has a budget. It is the smallest reading of the same record. Every machine that leaves a screen, and every desk that prices the world without moving, will query the same coordinates.
India is where we prove the tail is not a rounding error. Not the highway. The inner road: the galli, the unmarked lane, the society approach, the market street, the flooded service cut, the last meter to a gate. That is the fabric public dumps under-sample. Cities across Asia that share this road fabric are the second geography. The schema is taught here first. Time is a first-class axis: the same coordinate next week is often worth more than a new kilometer today. Change is how maps die and how we stay valuable.
The platform in one view
- Layer 1 · Physical stateContinuous observation of public physical space. Identity is stripped. What remains is motion, surface, weather, and place.
- Layer 2 · Perception and labelsDetect, track, and segment. Vehicles including mixed-traffic classes, pedestrians, riders, animals, signals, barricades, lane and crossing marks, potholes, speed breakers, cracks, manholes, waterlogging. Events on top of objects: mixed traffic, pedestrian in carriageway, animal in scene, impact candidates.
- Layer 3 · The Physical Event GraphEvery label becomes a record with lat/lon, time, speed, heading, inertial state, box, mask, and a Zipfian rarity score. Corridors over weeks become edges. Change is a first-class event.
- Layer 4 · Intelligence productsScene Query API, licensed packs, corridor time machine, long-tail index, world-model eval, map freshness, parametric risk, municipal ops, simulation seeds, embodied-stack packs, spatial-computing anchors, sovereign physical-state feeds.
- Layer 5 · MonetizationSeventeen revenue layers on one graph. Later layers do not need a second graph. They need packaging, governance, and a seat.
The Opportunity
Physical AI already has a data supply chain. It is assembled from the wrong unit. Survey fleets produce beautiful kilometers of the same highway. Public academic sets freeze a few cities in a few seasons. Simulation invents the tail instead of measuring it. Dashcam archives are hours without a query language. Label vendors will annotate whatever you send them. Almost nobody owns a continuously refreshed, rarity-ranked, privacy-redacted event graph of the hard world - the world where two-wheelers, pedestrians, animals, and weather share the same unmarked meter.
The five-layer stack between a scene and a planner
| Layer | Who operates it | What it does |
|---|---|---|
| 5 · Planners and world models | OEM, ADAS, AV, and foundation-model labs | Turn labeled state into prediction, planning, and control |
| 4 · Evaluation and simulation | In-house benches, closed courses, synthetic worlds | Test whether a model survives a scene it has barely seen |
| 3 · Maps and change | HD-map and localization teams | Hold geometry that goes stale the week after it is surveyed |
| 2 · Labeling | Human and model annotators | Turn pixels into classes if someone already captured the right hour |
| 1 · Physical ground truth | Whoever wrote the scene as an event | The record: what happened, where, when, how rare |
ZipF sits at layer 1 and publishes into 2 through 5. We do not sell another unlabeled archive. We sell the record a planner can ask for.
Two structural facts that define the strategy
1. The richest record in physical AI is born at the scene, not in the lab. A world model that has never seen a pedestrian step into mixed two-wheeler flow at night, on a wet unmarked road, will invent that scene or fail it. The valuable record is first-party, geolocated, time-stamped, and rare. We write it as an event with a schema labs already know how to train on: class, track, box, mask, GNSS, inertial state, weather join.
2. Licensed physical intelligence is a proven institutional market. ADAS, HD maps, automotive training data, and DaaS already have buyers and price bands. The open position is owning a source those buyers cannot reconstruct from Western highway footage or from a one-time survey pass.
The macro information loop
- 1 · Physical sceneActors, surface, weather, and geometry coincide
- 2 · Event writeDetect, track, segment, redact, locate, timestamp
- 3 · Rarity rankZipf score: how scarce is this join in the graph
- 4 · QueryA lab asks for a cut, not a dump
- 5 · Train / eval / map / insureThe same record serves four products
- 6 · The world movesNext week the corridor changes. We write it again.
The defining design principle follows directly: ZipF is built on its own rails. The proprietary asset comes from physical observation we control, a label stack we own, and a graph we license. Enrichment (weather, public map tiles, administrative boundaries) stays enrichment. The core dataset is owned, redacted, and free of an upstream platform that can turn the tap off.
The Product
The buyer-facing proposition is one sentence in every market: ask the physical world a question and get a licensed, labeled, privacy-redacted answer. We do not lead with cameras. We lead with the query. A perception team wants 50,000 night-rain two-wheeler conflicts. A humanoid team wants stairs, thresholds, and overhead wires at a coordinate. A drone stack wants the last fifty meters of an urban canyon. A parametric desk wants a signed flood event at a lat/lon. A world-model lab wants an eval suite the public internet cannot leak. Those are the same graph, packaged for different bodies and desks.
The in-house product suite
What an event holds
An event is observed physical activity at a place and a second. We record it with a stable schema. The visual, when it ships, is redacted. The sidecar is the truth a model trains on even when the pixels are gone.
| Inside the record | Fact it writes | Who that fact serves |
|---|---|---|
| Actors and tracks | Class, box, mask, identity-within-clip, speed, heading. Vehicles including mixed-traffic classes, pedestrians, riders, animals, signals, barricades. | Perception, prediction, world models, sidewalk robotics. |
| Surface and geometry | Pothole, speed breaker, crack, manhole, waterlogging, lane and zebra marks, unmarked edge. | HD maps, municipal ops, ride comfort, insurance. |
| Scene events | Mixed traffic, dense two-wheelers, pedestrian in carriageway, animal in scene, impact candidates, work-zone conflict. | Long-tail training, eval, safety cases. |
| Event coordinates | Exact lat/lon of the event itself, not only the path that observed it. A heading, a speed, a timestamp on that meter. | Any planner that must stand on a point, not on a clip. |
| State sidecar | GNSS (lat/lon, speed, heading), inertial / shake, timestamp, weather join. | Localization, reconstruction, simulation seeds, robot odometry. |
| Passable width | How much free geometry is left: alley, parked two-wheeler, curb, gate. The fact a car-sized body cannot occupy this meter. | Any body smaller or larger than a car. |
| Vertical clearance | Wires, awnings, balconies, trees, flyover undersides. Height a humanoid, a tall AGV, or an aerial last-fifty-meters stack must not occupy. | Humanoids, aerial last-meter, inspection, construction. |
| Threshold and handoff | Stairs, ramps, curb cuts, doorways, indoor-outdoor edges. The meter where a body changes surface. | Humanoids, wheelchairs, campus and hospital fleets. |
| Temporal occupancy | The same meter at 2pm (school gate), at night (bazaar), after rain (flood), on a festival day. Clock as geometry. | Routing, cities, parametric desks, any planner with a calendar. |
| Privacy cut | Faces and plates gone. No identity product. Anonymized physical state only. | Every licensed buyer. The legal reason the pack can leave the building. |
| Rarity and change | Zipf rank of the join. Diff against the last pass of the same corridor. | Who pays a premium. Who needs a refresh. |
Line drawings of detect, redact, query, and change sit after the Physical Event Graph.
A coordinate, not a clip
Hours of video are a path. An event is a point. We write the label and the telemetry on the same clock: class, box, mask, plus GNSS, inertial state, speed, heading, and the exact coordinates of what happened. A lab can ask for this meter, this second, this weather, this rarity - and get a record a planner can stand on even if they never open a frame.
That join is what any machine that leaves a screen actually needs. A car, a robot, a humanoid, a drone, a wheelchair, a glasses stack. They do not want a beautiful highway hour. They want a fact at a coordinate: free width, surface, weather, what is blocking the meter, what changed since last week, how rare the join is. ADAS is the first body that already pays. It is the smallest reading of the same record.
The last meter - where survey cars do not go
A roof-mounted survey vehicle needs a legal travel lane and a body that fits it. Most of the physical world a future robot has to finish does not look like that. Narrow lanes, gallis, service cuts, bazaar paths, school-gate crush, temple approach, village tracks, flyover undersides, the gap between parked two-wheelers, the curb a wheelchair or a delivery bot must climb. Those meters are where people already walk and ride. They are also where most public street-level maps are thin or absent.
The last 80 meters of an Indian inner road are not a leftover of the arterial. They are a different geometry: passable width, gate, curb, blocked meter, flood. A planner that only saw a highway cannot finish them. We write events in that geometry - the same weather and surface join we already write for ADAS - and license the cut to the stacks that will live there.
The last meter is one door. Delivery robots are one body that will walk through it. The company is every body that leaves a screen, and every desk that prices the same meter without a vehicle. That story is every body, every desk.
Token architecture - the query carries a pointer, the scene stays server-side
A buyer query does not download the world. It receives a token for a licensed cut. The pixels, if any, resolve server-side. The graph stays ours. The pack is what they train on.
What the query already named
What the token returns
Four data-product layers
One scene, written in full
| Field | Value |
|---|---|
| Event | pedestrian_in_carriageway |
| Actors | 1 pedestrian · 3 two-wheelers · 1 auto |
| Surface | unmarked edge · wet |
| Weather | rain · night |
| Speed | 22 km/h at write |
| Event GPS | 12.9716, 77.5946 |
| IMU | shake 0.41 · heading 214° |
| Tracks | 4 · 1.8 s overlap |
| Rarity | band 4 of 5 |
| Privacy | faces and plates redacted |
| Change | marking gone vs last Tuesday |
| License | train + eval |
One tiny scene on a Bengaluru wet night. The same schema writes Jakarta rain, a Hanoi unmarked edge, a Manila flood. Multiply: density plus time plus labels is how we stop answering questions about a clip and start answering questions about a city, then a country, then a model that claims it can drive, walk, or deliver here.
Each event carries: class · track · box · mask · event lat/lon · heading · speed · IMU · timestamp · weather · passable width · vertical clearance · threshold · temporal occupancy · rarity · change · license class. One ontology.
The Physical Event Graph
The dataset is the Physical Event Graph: every event answers nine questions, and the joins between events answer the questions labs pay for.
The data spine
Inputs (per scene)
Product pillars
Downstream monetization
The canonical physical ontology - the quiet technical moat
Cross-lab intelligence lives or dies on normalization. One stack calls it person, another pedestrian, a third VRU. The same surface appears as pothole, distress, and defect_04. ZipF's normalizer resolves them to one canonical class, then normalizes place, time, weather, units, and rarity across every cut. The result is a canonical physical ontology - a cross-buyer event graph that grows harder to reproduce with every corridor-month of history.
At maturity the benchmark statement reads: ZipF observes N geolocated long-tail events across K corridors and T weeks - a view no single survey pass and no single public dump can compute.
We are not collecting hours. We are collecting the tail.
The moat is rarity plus time, not extra pixels
Anyone can store video. That is not a moat. Extra classes on a label file are a feature a vendor can copy. The durable edge is seeing which scenes almost never happen, then seeing the same place again after the world has moved: rarity → query → license → refresh. Archives see a dump. Survey companies see a pass. ZipF sits on the layer that connects both.
We do not win by collecting more columns on a frame. We win by ranking the tail and watching it change.
Language already proved the mechanism. A few words carry most of the count. The rare words carry the meaning. Physical AI is the same curve. Cars-in-lanes in daylight are the head. The tail is the pedestrian in the carriageway, the animal at dusk, the unmarked breaker under water, the eight-actor conflict. We take Zipf's law literally: rank scenes, price the tail, refresh the corridor.
Illustrative Zipf curve of physical scene types (not a measured ZipF panel). The head is cheap and public. The tail is where models fail. The product is the missing mass, not the first bar.
| What labs already have | What ZipF adds |
|---|---|
| Highway and marked-lane hours | Mixed traffic, two-wheelers, informal geometry |
| A one-time survey pass | The same corridor next week |
| Object lists without weather | Actor × surface × weather as one join |
| Unlabeled or weakly labeled dumps | Events a query can return |
| Identity in the frame | Redacted physical state |
Comparisons describe dataset shape, not a named rival. Used as a mechanism, not as ZipF measured share.
In plain English
- A world-model team asks for night rain and a pedestrian in the lane. We already wrote those events. They train on a cut, not a scavenger hunt.
- A map still says the marking is there. We saw it gone on Tuesday. The next localization error never leaves the lab as a surprise.
- An insurer wants wet-surface defect density in a district. We return a score from events, not from a claim that arrived too late.
- A sidewalk robot needs the curb that cars treat as a lane. The same ontology writes the sidewalk that the road already taught us.
- A hundred hours look like progress. Fifty-one rare events are the dataset. We see the rest of that curve. That is what we own.
Capture (the event) plus representation (what the query asked) plus outcome (whether the model, map, or desk used it). Either side alone is copyable. Both sides together are not.
Uncaptured tail - what a public dump never sees
A public set reports classes it already has. The graph can estimate classes the world keeps producing and the dump keeps missing. If a corridor yields 10,000 vehicle tracks and 80 animal-in-scene events, the 80 are the product. That is a lab product ("you trained on the head") and a market product (safety, insurance, and cities buy the missing mass).
Live physical state, not a static map
Today a map says a lane exists. That is a field someone surveyed last season. On a Tuesday after rain the marking is gone, the breaker is under water, and a work zone has stolen a meter. A planner still trusts the file. The file did not lie about last season. It just was not true for this second.
In two or three years the default interface for a robot, an ADAS stack, or an agent that moves a body through a city is not a downloaded map. It is a query: is this corridor wet, is the crossing marked, did an animal appear, has the work zone moved. Something has to answer with structured state. If that answer is stale, the agent plans the wrong meter. If it is live, the agent only gets it from the network that writes the event.
The person or the model talks to their stack in features. The stack turns that into a structured request. We answer with structured state. The parameters are not generic "road" and "car." They are the fields that decide whether this planner, this minute, should trust the world.
| What the stack asks | Structured query we receive | Live field we return |
|---|---|---|
| Is this corridor still the one we mapped | geo, time, last pass id | Change events, marking loss, work zone |
| Hard VRU cases at night in rain | event, weather, daypart, rarity | Matched events, tracks, sidecar |
| Surface risk for this district | defect class, weather join | Density, severity, last seen |
| Did our VLM actually see this | eval suite id | Grounding score on held-out tail |
| Seed this conflict in sim | event id, license | Tracks, timing, reconstruction pack |
| Can a 1.8m body stand here | width, vertical clearance, threshold | Free geometry, stairs, overhead wires |
| Did this lat/lon flood after the rain | event class, event GPS, weather join | Signed standing-water event, last seen |
| Is the last fifty meters of this canyon clear | geo, vertical, temporal occupancy | Wires, trees, crush, landing surface |
Illustrative request/response contract. The model or map stack is the client. ZipF is the physical-world infrastructure. One event stream, many products on top.
Illustrative Scene Query API payloads. Same graph. Four readings: a car stack, a last-meter robot, a humanoid, a parametric desk. Live fields come from events, not from a scraped video site.
| Buyer | What they pay for | Verified pricing analogue |
|---|---|---|
| World-model / foundation labs | Long-tail events, eval suites, grounding scores | Automotive AI training datasets $953M in 2026 (Grand Research Store) |
| ADAS / OEM | VRU, night, weather, mixed-traffic packs | ADAS systems $57.25B in 2026 (TBRC); components $23.57B (MarketsandMarkets) |
| HD maps | Change, markings, surface, work zones | HD maps $4.90B in 2026 (GII / 360iResearch); $1.34B on a narrower perimeter (MarketsandMarkets) |
| Insurance / telematics | Surface risk, weather join, conflict density | Same DaaS pool: Grand View $17.38B (2024) → $76.80B (2030) |
| Cities / road authorities | Defects, flooding, signal and barricade state | Ops budget, not a model license - sold as tickets and layers |
| Last-mile / delivery robots | Passable width, alley, curb, gate, doorstep, event GPS | Same event schema. No separate published slice yet - sold as robotics packs |
| Two-wheeler / micromobility AV | Informal cuts, mixed flow, unmarked edges | Sits under ADAS / training-data perimeters |
| Campus / yard / port / hospital | Closed-loop people + weather + change | Same ontology the street already taught |
| Logistics / e-commerce last 200m | What is actually passable to a door today | Ops and routing, not a model license |
| Accessibility / last-meter mobility | Curb, broken surface, width, flood | Cities and mobility operators |
| Emergency / after-rain routing | Blocked meters, flood, work-zone diffs | Cities, fleets, insurers |
| Digital twins / informal city | The unofficial path graph, refreshed | Sim, planning, municipal twins |
| Simulation | Reconstruction seeds from real rare events | Sits under the same training-data perimeter |
| Humanoid / embodied AI | Stairs, thresholds, overhead wires, indoor-outdoor handoff | Same event schema. No separate published slice - sold as embodied packs |
| Aerial last-fifty-meters | Wires, trees, urban canyon, landing surface, crush | Same vertical-clearance and occupancy fields |
| Construction / site autonomy | Work-zone change, materials, people, unofficial paths | Ops and safety. Same change layer maps already buy |
| Agriculture / rural tracks | Seasonal surface, animals, unmarked geometry | Same ontology the village corridor already taught |
| Wearable / glasses / spatial compute | Persistent anchors, change, passable geometry for a human body | Same coordinates a robot stands on. Different client |
| Parametric insurance | A signed flood, hail, or blockage event at a lat/lon | Same DaaS / risk pool. Event GPS is the settle field |
| Infrastructure finance | Pavement and flood as a corridor-year risk input | Cities and contractors already buy the tickets. Finance reads the series |
| Climate / public health | Standing water, heat, dust, after-rain occupancy | Agencies and research desks. No identity product |
| Telecom / utilities | Street-canyon geometry, poles, wires, manholes | Same vertical and surface records |
| Real estate / site selection | Approach quality, flood history, access width | The corridor time machine as a site file |
| Accessibility compliance | Curb cuts, ramps, broken surface, width as a civic layer | Cities and mobility operators |
| Curb / charging / last-meter markets | What is actually free at this meter, this hour | Ops, not a model license |
| Legal / reconstruction | A versioned, signed event at a coordinate and a second | Safety cases already buy this shape. Law reads the same log |
| Sovereign physical-state API | A country's unofficial + official path graph, refreshed | National digital infrastructure. Ontology license |
| Agentic physical assistants | Can this body, this hour, finish this meter | The query is the product. The stack is whoever books the walk |
Pricing analogues are public category anchors, not ZipF list prices. Strategic seats remain in the $50K-$2M+/year band already used in §9.
Five modules, one licensed event
The same pipeline writes the scene, the privacy cut, and the query. Each line drawing is the module. The paragraph is what it gives the graph.
01 · Detect, track, segment
Pixels become objects, then tracks, then masks. The grid on the right is a frame of classes: vehicle, two-wheeler, pedestrian, rider, animal, surface. Time binds them into tracks. A mask is the shape a planner can occupy or avoid. Identity is not the product. The class is.
02 · State sidecar
GNSS writes place, speed, heading. IMU writes shake and the body's own motion. A clock binds them to the frame. The field of view is the corridor: where we were, how fast, whether the surface punched the chassis. That is localization on the graph, reconstruction for sim, and a weather join that is not a guess from the pixels alone.
03 · Privacy cut
A face or a plate is identity. Identity is not a physical-AI feature. The cut removes it before a pack can leave. What remains is geometry, class, and motion. Cities, labs, and insurers can buy the same record because the person is already gone.
04 · Rarity rank
Every join gets a Zipf score. Head events are abundant. Tail events are the product. A desk can ask for band 4 and get the scarce mass, not another thousand cars-in-lanes. The rank is how we price, how we sample, and how we refuse to drown a buyer in the average.
05 · Query + change join
Events, rarity, and last-pass diffs land on one clock. The join is class × weather × place × time × change. Train, eval, map, risk. That licensed event is what §7 through §9 sell.
Market Opportunity
Buyers sit in global dollar pools. The graph is written first on the Indian roads a highway archive never sees, then on Asian cities that share that fabric. Those are not the same number. The pools below are what labs and OEMs already pay. The roads below that are why we will be hard to copy.
5.1 · The base economy: ADAS, maps, and training data
Global ADAS systems market, $B. Source: The Business Research Company, ADAS Global Market Report 2026: $49.56B in 2025 → $57.25B in 2026 (15.5% CAGR) → $101.68B in 2030.
HD maps for autonomous vehicles, $B. Source: 360iResearch via GII: $3.85B in 2025 → $4.90B in 2026 → $23.35B in 2032 (29.3% CAGR). MarketsandMarkets publishes a narrower $1.34B (2026) → $2.43B (2033) series. Different perimeters; we do not add them.
| Worldwide metric | Value |
|---|---|
| ADAS systems 2026 (TBRC) | $57.25B → $101.68B by 2030 |
| ADAS components 2026 (MnM) | $23.57B → $58.84B by 2033 |
| ADAS units 2026 (MnM) | 361.4M → 582.6M by 2033 |
| HD maps 2026 (GII) | $4.90B → $23.35B by 2032 |
| HD maps 2026 (MnM, narrower) | $1.34B → $2.43B by 2033 |
| Automotive AI training data 2026 | $953M → $1.645B by 2034 |
| Expected US AVs by 2030 (IIHS) | 4.5 million |
| Global DaaS 2024 → 2030 | $17.38B → $76.80B |
These are pools. ZipF's take is the labeled long-tail layer those pools query. We do not treat ADAS revenue as ZipF revenue.
The addressable economy is every model that must see, predict, or price the physical world. Regulation is pulling ADAS from option to mandate: EU GSR already requires AEB, lane keeping, and driver monitoring on new vehicles; NHTSA FMVSS 127 requires pedestrian AEB on new light vehicles from September 2029. China NCAP 2024 raised the weight of VRU protection. That is not a ZipF number. It is why perception data has a buyer that cannot wait for a perfect AV future.
5.2 · Why the training-data number is the floor, not the company
Automotive AI training datasets at $953M in 2026 are the published slice closest to what we sell. It is also the slice most likely to be under-counted once world models, embodied stacks, parametric desks, spatial computing, and municipal ops start buying the same events. We treat $953M as proof that labeled physical data is already a market, and $57B ADAS / $4.9B maps / $76.8B DaaS as the larger rooms that same graph can walk into. We do not invent a TAM for a body that does not yet have a published perimeter.
Automotive AI training datasets, $M. Source: Grand Research Store: $869M in 2025 → $953M in 2026 → $1,645M in 2034 (9.7% CAGR).
Global Data-as-a-Service, $B. Source: Grand View Research: $17.38B in 2024 → $76.80B in 2030 (28.1% CAGR). The licensed-intelligence pool ZipF also sells into.
5.3 · India's roads - and what the events become
As of 31 March 2025, India has more than 63 lakh km of roads. National Highways are 1,46,204 km. State Highways 1,79,535 km. Other roads 60,19,723 km (PIB / MoRTH). A survey vehicle is built for the first two. The graph is built for the third: the inner road, the unmarked lane, the last meter a highway file does not contain.
That is the uniqueness. Not a vehicle-sales story. A road-data story. The meters between an arterial and a door: gallis, society approaches, compound roads, service cuts, market streets, school-gate crush, night bazaars, flooded inner lanes, speed-breaker sequences, flyover undersides, peri-urban tracks. Labeled, coordinated, refreshed. Cities across Asia that share this fabric - Jakarta, Hanoi, Manila, Bangkok - are the second geography. The schema does not restart.
| The road | What we write on it | What that becomes |
|---|---|---|
| Inner galli / unmarked lane | Passable width, parked bodies, unmarked edge, mixed flow | A path graph maps do not have. Last-meter packs. Two-wheeler and sidewalk autonomy. |
| Society / apartment approach | Gate, compound, curb, stairs, doorstep, blocked meter | Doorstep ontology. Embodied-stack nav. Access quality for a site file. |
| Market / kirana street | Stalls, crush, informal edge, temporal occupancy | A digital twin of the unofficial city. Routing that knows 11am is not 11pm. |
| School-gate street | Clocked crush, weekday occupancy, VRU density | Temporal occupancy index. City ops. Any planner with a calendar. |
| Service lane / flyover underside | Narrow cut, overhead clearance, dark, wet | Vertical clearance. Aerial last-fifty-meters. Inspection. |
| Flooded inner road | Standing water at a lat/lon, after-rain passability | Signed settle events. City tickets. After-rain routing. |
| Broken patch / breaker sequence | Pothole, crack, manhole, speed breaker, IMU shake | Surface distress layer. Insurance risk. Ride-comfort and municipal ops. |
| Work zone / stolen meter | Barricade, marking loss, change vs last pass | Map freshness. The invoice that is time. |
| Peri-urban / village track | Unmarked geometry, animals, seasonal surface | The ontology the arterial never taught. Rural and edge-city packs. |
| Night inner city | Lighting change, glare, sparse marks, the same meter remade | Night-rain VRU packs. The join public dumps starve. |
National Highways, State Highways, and other-road lengths: PIB, Transforming India's Transport Infrastructure, figures as of 31 March 2025. Those kilometers describe the fabric. They are not ZipF revenue.
Go-to-Market: Query First, Density as Flagship
Day-one motion is a query a lab can run and a pack they can train on. Density is how the graph gets good: Indian inner roads that are hard, refreshed, and ranked, then the same schema on Asian cities that share that fabric. We lead with the event types they cannot buy from a highway archive.
- Beachhead geographyIndia first. Inner roads, unmarked lanes, society approaches, market streets, monsoon. Then Asian cities that share that road fabric. The schema is taught once.
- Beachhead buyerADAS and world-model teams that already have a data budget and a failed-on-the-tail story. Maps and insurance follow once time exists. Every later body follows once width, height, and threshold exist. Every desk follows once event GPS is a settle field.
- FlywheelMore rare events → better queries → more seats → more corridor budget → more time → better change → more seats.
The Seventeen-Layer Monetization Stack
One graph. Seventeen ways to invoice it. Later layers do not require a second graph. They require packaging, governance, and a sales motion.
| # | Layer | What ships | Who pays |
|---|---|---|---|
| 1 | Scene packs | Licensed, redacted event cuts for a training run | ADAS, world-model labs |
| 2 | Corridor subscriptions | The same geography, refreshed | Maps, cities, labs that stayed |
| 3 | Scene Query API | Credits, filters, rarity bands | Any team with a pipeline |
| 4 | Long-tail index | Ranked scarcity of joins | Labs that want the gradient |
| 5 | World-model eval | Held-out suites, scores | Foundation and AV labs |
| 6 | VLM grounding | Does this model actually see the meter | Multimodal labs |
| 7 | Change / map freshness | Diffs, marking loss, work zones | HD maps, localization |
| 8 | Surface ops | Pothole, flood, crack tickets | Cities, contractors, infrastructure finance |
| 9 | Insurance / parametric | Risk scores, signed events at a lat/lon | Carriers, fleets, parametric desks |
| 10 | Simulation seeds | Reconstruction packs from real rares | Sim, safety, game, film worlds |
| 11 | Embodied stacks | Width, height, threshold, event GPS + IMU | Robots, humanoids, aerial last-fifty-meters, wheelchairs |
| 12 | Weather-joint packs | Actor × surface × weather | Any perception stack |
| 13 | Custom geo | A corridor they name, we already cover or extend | OEM programs |
| 14 | Exclusive windows | Time-boxed uniqueness on a cut | Strategic seats |
| 15 | Latency SKUs | T+week, T+day, T+hour | Maps and live ops |
| 16 | Prediction / signed logs | Versioned events for a claimed scene class | Safety, assurance, legal reconstruction |
| 17 | Ontology / physical-state API | The schema itself, and a country's live path graph | Platforms, cities, sovereign seats |
Layers 1-4 fund the company. Layers 5-12 are the expansion. Layers 13-17 are the lock-in. None of them require a second graph.
Proprietary Indices and the Scene Query API
Named series make the graph purchasable as a feed, not only as a dump. An index is an aggregate a desk can put on a terminal. An API is a question a model can ask. Both sit on the same events.
Point-in-time fields travel with production feeds: event_time, ingest_time, published_time, correction_time. Latency is a SKU. Distribution can be open, limited, or exclusive. A safety case can ask what the graph knew at a given minute and get a versioned answer.
Revenue Model and Scenarios
The invoice is a license, a seat, or a corridor. We do not take a cut of ADAS hardware. We do not take a cut of an insurance premium. We sell the context those businesses already buy from someone. Planning constructs below are not forecasts and not traction.
| SKU | Planning band | What it is |
|---|---|---|
| Eval pack | $15K-$50K | Held-out suite, one program |
| Training pack | $50K-$250K | Named long-tail cut, term license |
| Corridor year | $80K-$400K | Refresh + change on a geography |
| Query seat | $2K-$15K / mo | API credits and rarity bands |
| Strategic seat | $250K-$2M+ / yr | Exclusivity, latency, custom geo |
Illustrative mix at two planning scales (not a forecast). Packs start the book. Seats and corridors take over as the graph ages. Cities and insurance are the late layers.
A simple compounding story: 20 training packs at $100K is $2M. Add 10 corridor years at $150K and the book is $3.5M. Add 40 query seats at $6K/month and the run-rate is another $2.9M. None of those numbers are commitments. They exist to show that the same events invoice three times without a second graph.
Unit Economics
The unit is a licensed event-month: one corridor, one week of rarity-ranked events, one buyer class. Contribution is license revenue minus the cost to keep that corridor fresh and the pack governed. We publish the shape a buyer already understands: payback on a seat, not payback on a device.
Cash discipline is corridor discipline. We do not paint the planet. We pick geographies where the tail is thick, keep them warm, and sell them many times. A new city is a capital decision. A new layer on an old city is a packaging decision.
The Path to a Self-Funding Graph
The question is not "how many cameras." It is: at what seat price and seat count does license revenue cover the cost of keeping the graph alive. The explorer below is a planning toy. It is not a raise model and not a forecast.
Pick a seat price, a pack attach, and a monthly burn. The bars are installed-seat contribution against that burn.
Seat / month
Attach CAC to land the seat (planning)
Central burn / month (planning)
Planning: installed-seat contribution vs a flat burn+refresh line. The crossing is "the graph funds itself." Not a forecast.
Strategic Focus and Sequencing
- Now · Write the tailOntology, redaction, rarity, a first set of hard Indian corridors, a query a lab can run. No second product.
- Next · Sell the cut twiceA training pack and a refresh. Prove that time is the second invoice. Land one ADAS or world-model seat and one map or city conversation on the same geography.
- Then · Seat the APICredits, eval, grounding. The buyer stops taking dumps. Their pipeline points at us.
- Later · Walk the layersInsurance, every embodied stack, parametric settle, spatial anchors, sovereign feeds. Only after the graph is dense enough that a new layer is packaging.
- Always · The graphThe company is the event, the query, and the license.
Every Surface, Every Geography
Roads are the first surface because they concentrate actors, weather, and geometry at a frequency nothing else matches. The ontology does not stop at the curb.
India is the beachhead because the inner road is ordinary there and the highway file is thin. Jakarta, Hanoi, Manila, and Bangkok share that fabric: unmarked edges, last meters, weather stress, unofficial paths. A sidewalk robot in Austin and a mixed-traffic planner in Bengaluru can consume the same event types once the ontology exists. We prove it here first. The world is still the TAM. The uniqueness is this road.
Every body, every desk
A car was the first machine with a budget to see. A delivery robot is the example people reach for next. Both are true. Both are too small. The graph is physical context for any body that has to finish a meter, and for any desk that has to price, insure, plan, or reconstruct that meter without ever leaving a chair.
The company is not a dataset for cars. It is the live state of the world, sold to whoever has to move through it or settle against it.
Bodies that will query us
Every later machine inherits the same record: class, track, event GPS, IMU, weather, width, height, threshold, rarity, change. The body size changes. The schema does not.
Desks that never move
A desk does not need a chassis. It needs a signed fact at a lat/lon. The same event that trains a planner can settle a contract, score a corridor-year, or tell a city the marking is gone.
The same record, four readings
One labeled event, with telemetry and exact coordinates, is four companies depending on who asks.
| The fact we wrote | A body reads | A desk reads |
|---|---|---|
| Standing water at 12.9716, 77.5946 after rain | A robot or wheelchair cannot finish the meter | A parametric contract settles. A city opens a ticket. A site file ages. |
| Passable width 1.3m, parked two-wheeler, gate closed | A car cannot occupy it. A smaller body still might. | A last-meter market, a logistics router, a digital twin of the unofficial path. |
| Overhead wire at 1.9m on a galli approach | A humanoid or a tall AGV stops. An aerial last-fifty-meters stack reroutes. | A utility inspection. A telecom canyon. A construction-site constraint. |
| School-gate crush, weekdays 14:00, this coordinate | Any planner with a calendar treats the meter as occupied. | A city, an insurer, a temporal twin. Clock as geometry. |
| Stairs, no ramp, curb broken, IMU shake | A wheelchair fails. A humanoid changes gait. A delivery bot needs a different door. | Accessibility compliance. Real-estate access. A civic layer. |
| Marking gone, work zone stole a meter, last pass 9 days ago | ADAS and maps are already wrong. | Change is the invoice. HD maps, cities, reconstruction, spatial anchors. |
We do not need a second graph for a second body. We need to keep writing events in the geometry a car-sized mapper cannot occupy, keep the clock on the same coordinate, and license the cut to whoever has to stand on it or settle against it. Cars pay first. Everything that moves after a car, and everything that prices the world without moving, is the larger company.
ADAS is how the graph gets paid first. Delivery robots were the example that made the last meter obvious. Humanoids, aerial last-fifty-meters, parametric desks, spatial computing, and sovereign physical-state feeds are why the same coordinates keep getting paid.
Appendix
Sources (verified Oct 2026)
1. The Business Research Company, Advanced Driver Assistance Systems Global Market Report 2026 - $49.56B in 2025 → $57.25B in 2026 (15.5% CAGR) → $101.68B in 2030 (15.4% CAGR). Growth attributed to safety mandates, sensor cost decline, and the path toward autonomous driving (GII / TBRC).
2. MarketsandMarkets, ADAS Market 2026-2033 - components $23.57B in 2026 → $58.84B by 2033 (14.0% CAGR); units 361.4M → 582.6M (7.1% CAGR). EU GSR mandates AEB, lane keeping, and DMS; NHTSA FMVSS No. 127 (April 2024) requires pedestrian AEB on new light vehicles from September 2029; China NCAP 2024 raises VRU weighting (MarketsandMarkets).
3. 360iResearch / GII, HD Map for Autonomous Vehicles - $3.85B in 2025 → $4.90B in 2026 → $23.35B by 2032 (29.32% CAGR). Shift from survey-only production toward hybrid refresh and AI-assisted change detection (GII / 360iResearch).
4. MarketsandMarkets, HD Maps 2026-2033 - narrower perimeter $1.34B in 2026 → $2.43B by 2033 (8.9% CAGR). Cited as a second published series, not added to (3) (MarketsandMarkets HD maps).
5. Grand Research Store, Automotive AI Training Datasets - $869M in 2025 → $953M in 2026 → $1,645M by 2034 (9.7% CAGR) (Grand Research Store).
6. National Association of Insurance Commissioners, December 2024 - Insurance Institute for Highway Safety expects 4.5 million self-driving vehicles on US roads by 2030 (cited in TBRC ADAS report).
7. Grand View Research, Data-as-a-Service - $17.38B in 2024 → $76.80B in 2030 (28.1% CAGR) (Grand View DaaS).
8. George Kingsley Zipf, rank-frequency law - frequency of rank k is proportional to k to the power of minus s. The company name is spoken Ziff.
9. Press Information Bureau, Transforming India's Transport Infrastructure (2014-2025) - as of 31 March 2025, India has more than 63 lakh km of roads: National Highways 1,46,204 km, State Highways 1,79,535 km, other roads 60,19,723 km. Second-largest road network in the world (PIB).
Every rare scene on earth is a data point. Every body that leaves a screen will ask for it. ZipF is the machine that ranks them.