Spatial Intelligence: Infrastructure & Application Layer
Physical AI is a multi-$T market. Frontier labs, robotics companies, and world model labs are spending billions on actuators, batteries, training compute, and data. But there's a whole software infrastructure layer that needs to be rebuilt from scratch for video, sensor data, hardware logs, depth, and 3D trajectories. When I was working with data at my last company we hit this wall: AV2 chips existed for encoding/decoding but none of the hardware supported it, uploading TBs of multi-camera data to the cloud was painful, and after extraction most of it sat raw in AWS Glacier bleeding money.
The smartest teams are building their infra: Tesla patented a custom file format (.smol) that replaces MP4/CSV with a header-indexed, tensor architecture: claiming 4x IOPS reduction and 11% smaller files. Standard Intelligence built a 30PB storage cluster from scratch because S3 is 40x more expensive and structurally unusable for high-throughput video workloads. Hugging Face shipped optimized-parquet for structured ML datasets but it stores video as URL pointers, not queryable bytes. Snowflake & Databricks offer 20k+ enterprises micro-partitioning (50-500MB/1GB), which revolutionized structured analytics, but zero equivalent exists for raw video streams, 3D scenes, depth data, or gaussian splats.
On the application side: World Labs just published "3D as code", arguing 3D is to spatial computing what code is to software. Gaussian splats are becoming the native format but today they're still dumb files: millions of primitives you can render but can't query, segment, or edit through code. Nobody has built the database and API layer underneath. Semantic querying of splats is technically proven but zero production APIs exist.