Introducing Arcflow: The Spatial Graph Database That Powers the World Model
Every world model needs a spatial data layer. Not a general-purpose database with a spatial extension bolted on. Not a graph database that treats coordinates as metadata. A database built from first principles for the question: what is near what, right now, and how is that changing?
That database is Arcflow. Today we are opening pre-release access to development partners and builders in the spatial intelligence and world model community.
Why we built Arcflow#
OZ builds continuous world models at physical venues. Every second, our edge infrastructure tracks 22+ entities across six 4K camera feeds, fuses detections into a unified spatial representation, resolves entity identities across overlapping fields of view, maps spatial relationships, and detects events, all within a 120ms budget.
The Venue Graph, the spatial-temporal representation of everything happening at a venue, needed a query engine that could answer spatial questions at the speed of perception. Existing graph databases are designed for social networks, knowledge graphs, and recommendation engines. They treat space as an afterthought. We needed space to be the primary axis.
So we built one.
What Arcflow is#
Arcflow is a C++ graph database with three design commitments:
Spatial predicates are native, not extensions. Every query primitive understands physical space. WITHIN, NEAREST, INTERSECTS, TRAJECTORY_CROSSES, these are first-class operations, not UDFs layered on top of a B-tree. The query planner optimizes for spatial locality the way traditional databases optimize for key locality.
GPU acceleration is built in. Spatial queries are embarrassingly parallel; testing thousands of entities against geometric predicates is exactly the workload GPUs excel at. Arcflow ships with CUDA and Metal backends that accelerate spatial joins, proximity searches, and trajectory intersection queries. CUDA targets the NVIDIA data center GPUs that power production workloads at scale (the A100s, H100s, and GB200s running in state-of-the-art inference clusters). Metal support gives developers on the Mac ecosystem the same development experience locally that scales to those data center GPU powerhouses in production. Write and validate spatial queries on your MacBook, deploy to CUDA clusters without changing a line. On venue-scale workloads, GPU-accelerated spatial joins run 40–100x faster than CPU-only execution.
Temporal is a dimension, not a log. Entities exist in time. Arcflow maintains full temporal state: not just current position, but trajectory history, velocity vectors, and temporal relationships. Querying "where was entity X when entity Y entered zone Z" is a single graph traversal, not a join across time-series tables.
Performance characteristics#
These numbers are from production Venue Graph workloads (22 tracked entities, six camera feeds, continuous operation):
| Operation | Latency | Notes |
|---|---|---|
| Spatial proximity query (k-nearest) | ≤0.3ms | GPU-accelerated, 1000+ entities |
| Trajectory intersection | ≤0.8ms | Full match duration, all entities |
| Entity relationship traversal | ≤0.1ms | 3-hop graph traversal |
| Temporal range query | ≤0.5ms | Arbitrary time window |
| Continuous spatial join (streaming) | ≤1ms per tick | Real-time zone activation |
These are p99 latencies measured at the venue edge under production load. The world model cannot wait for its data layer.
The road to open source#
Arcflow has been powering the Venue Graph in production for over a year. Every match OZ operates (every entity tracked, every spatial relationship mapped, every zone activation detected) runs through Arcflow.
We intend to open-source Arcflow fully. But we are not going to rush a public release and hope it sticks. Open-source infrastructure that matters, the projects that become the default, gets shaped by real workloads before it ships. SQLite was battle-tested inside Airbus and Bloomberg before it became the most deployed database on earth. We want Arcflow's public release to arrive with that same level of production hardening.
That is why we are starting with a closed early adopter programme. Before the open-source release, we are working directly with a small cohort of teams who are building at the frontier of spatial intelligence. Their workloads, their edge cases, their scale requirements will shape the APIs, the query planner, and the GPU acceleration paths that ship in the public release. When Arcflow goes open source, it will not be a research prototype; it will be infrastructure that early adopters are already running in production.
Early adopter programme#
We are selecting partners across three categories:
World model builders. Teams constructing spatial representations of physical environments: autonomous driving, drone operations, warehouse robotics, smart infrastructure. If your system needs to answer spatial queries in real time, Arcflow was built for your workload.
Spatial computing platforms. Teams building AR/VR/XR applications, digital scene understanding, or indoor positioning systems. Arcflow's native spatial predicates and temporal dimensions map directly to spatial computing requirements.
Research groups. Academic and industrial research teams working on spatial reasoning, multi-agent coordination, or physical AI. We are especially interested in collaborations that push the boundaries of what spatial graph queries can express.
Early adopters get the core Arcflow engine, CUDA and Metal GPU acceleration backends, documentation, and direct engineering access from the team that built it. In return, we get the workload diversity that makes the open-source release bulletproof.
What stays proprietary#
To be explicit: Arcflow is the spatial graph engine. It is not the world model. OZ's perception models, multi-camera fusion pipeline, robotic control algorithms, and production orchestration remain proprietary. Arcflow is the data layer, the tool for storing, querying, and reasoning about spatial-temporal data. What you put into it and what you build on top of it is yours.
The thesis#
The world model era requires new infrastructure primitives. Foundation model companies are building world models from internet-scale data. OZ builds world models from ground-truth sensor data at physical venues. Both approaches need spatial-first data infrastructure that does not exist in the current database landscape.
Arcflow is that infrastructure. And we believe the teams that adopt spatial-first data primitives earliest will compound their advantage the fastest, just as the teams that adopted columnar storage earliest won the analytics era, and the teams that adopted graph databases earliest won the social and knowledge graph era.
The spatial graph era is starting. Arcflow is how you build for it.
Apply for the Arcflow early adopter programme at oz.com/docs/get-started. Include your use case; we are selecting partners based on workload fit and willingness to push the boundaries of spatial querying. The cohort is small by design.