On the landing page of useproximity.ai, a small readout blinks with coordinates: LAT 34.0522 N, LON 118.2437 W. The label reads SAT-01. There is no pricing page, no customer logo wall, no founder bio. What there is, instead, is a sentence that describes one of the more ambitious bets in deeptech right now: a company building "the ontology and world models to connect thousands of physical data sources simultaneously, creating a real-time digital twin of planetary activity" [useproximity.ai].
That is Proximity AI. The company is operating in what looks like deliberate stealth, and the public surface is thin by design. But the technical premise it has staked out, if it works, would sit underneath a great deal of what people currently mean when they talk about agents and robots acting in the real world.
The bet
Most AI systems shipped in the last three years operate on text and pixels. They are good at language and increasingly good at images, but they have no persistent, structured understanding of where things are, what is moving, what just changed, or what is permitted. Proximity is trying to build that missing layer. According to a third-party listing, the platform is "designed to bring real-time awareness and understanding of the physical world by connecting diverse data streams like permits, sensors, supply chains, public data, cameras, and satellites" [AgentsPointee].
The key word in the company's own description is ontology. An ontology is the schema that tells a machine that a permit number, a truck's GPS ping, a satellite tile, and a port camera feed are all describing the same warehouse on the same afternoon. Without that schema, you have a data lake. With it, you have something an agent can query and a robot can act on. The company frames the output as a "real-time digital twin of planetary activity" intended to "empower Humans, Agents and Robots alike" [useproximity.ai]. The wedge, in other words, is not a model. It is the connective tissue between models and the physical world.
Why it could be big
The timing argument writes itself. Foundation-model vendors have largely conceded that the next frontier is agents that take actions, and robotics labs are shipping general-purpose policies that need grounded world state to be useful. Both camps need a substrate that knows what is actually happening outside the data center. Today that substrate is stitched together per customer by systems integrators, geospatial vendors, and IoT platforms that rarely speak to each other.
If Proximity can ship an ontology that ingests permits, supply chain events, sensor telemetry, public records, camera feeds, and satellite imagery into a single queryable model, the addressable surface is unusually wide: logistics operators trying to predict port congestion, insurers pricing physical risk, defense and intelligence buyers who already pay for fused situational awareness, and the agent and robotics developers who have nowhere else to source coherent ground truth. The reference to SAT-01 on the company's site [useproximity.ai] suggests the team is willing to own at least part of the sensing stack rather than depend entirely on third-party feeds, which would be consistent with a deeptech rather than pure-software posture.
The product surface today
What the company has publicly committed to is narrow and specific. Two claims anchor the work:
| Claim | Source |
|---|---|
| Ontology and world models connecting thousands of physical data sources into a real-time digital twin | useproximity.ai |
| Platform ingests permits, sensors, supply chains, public data, cameras, satellites | AgentsPointee |
A satellite component, SAT-01, is referenced in the company's own materials [useproximity.ai]. Beyond that, Proximity has not disclosed customers, pricing, or deployment scale in the captured public record.
The honest counterfactual
What skeptics will say is straightforward: digital twin and sensor-fusion platforms have a long history, and the hard part has never been the pitch deck. It has been the per-customer integration tax. Every new data source brings its own schema, latency profile, and access agreement, and the ontology that elegantly unifies ten of them tends to bend awkwardly around the eleventh. Competing efforts inside hyperscalers and geospatial incumbents already have distribution into the buyers Proximity would need to win.
The bull answer, supported by the company's own framing, is that the prior generation of digital twin tools were built before large models could do the semantic lifting required to reconcile messy real-world schemas on the fly. An ontology designed from day one to be consumed by agents, rather than by human GIS analysts, is a different product even if the input feeds look similar. The reference to empowering "Agents and Robots" alongside humans [useproximity.ai] suggests the team is building for that consumer, not for the dashboard era.
What to watch
The next twelve months should answer the questions that matter. First, a named design partner or pilot, particularly in logistics, insurance, or defense, would validate that the ontology survives contact with a real operational dataset. Second, any disclosure around SAT-01, whether it is an owned asset, a partnership, or a demonstration payload, will tell observers how vertically integrated the company intends to be. Third, a funding announcement would clarify which investors are underwriting the world-model thesis at the infrastructure layer rather than the application layer. Until then, the most informative artifact remains the landing page itself, and the choice to lead with coordinates rather than copy.
Technical breakdown
The stack as described implies three distinct engineering problems stacked on top of each other. The first is ingestion: normalizing heterogeneous feeds (orbital imagery measured in tiles and revisit windows, IoT telemetry measured in events per second, permits measured in PDFs and municipal APIs) into a common temporal and spatial frame. The second is the ontology itself: a schema expressive enough to represent that a permit, a truck, and a satellite pixel can all refer to the same physical event, and stable enough that downstream agents can rely on it. The third is the serving layer: exposing this fused state at latencies that are useful to an agent making a decision or a robot executing a plan, which is a meaningfully harder bar than a nightly ETL into a warehouse.
What could go wrong at scale
The failure mode to watch is not technical novelty, it is integration entropy. Ontologies that look clean at ten sources tend to fork at a hundred, and every fork is a maintenance liability that compounds with customer count. If Proximity wins early defense or logistics deployments, the temptation will be to bend the schema to each buyer's data, at which point the "single ontology" pitch quietly becomes a collection of bespoke pipelines wearing a shared logo. The companies that have survived this category did so by being ruthless about what they refused to model. Whether Proximity has that discipline is the question the next two years will settle.