For the last few years, the conversation around artificial intelligence has been dominated by the digital ether. We have marveled at large language models that can draft legal briefs in seconds and generative tools that create photorealistic art from a few keystrokes. But as a former software engineer, I’ve always noticed a glaring gap in this progress: AI is incredibly proficient at manipulating bits, but it remains largely blind to atoms.
Most of the AI we interact with today is trained on the “online” world—the curated archives of the internet, digitized books, and social media feeds. This represents a massive amount of data, but it only captures a sliver of human existence. The vast majority of the global economy happens in physical spaces—warehouses, cornfields, construction sites, and shipping ports—where the primary challenges aren’t about predicting the next word in a sentence, but about understanding the precise geometry of a physical environment.
Niantic Spatial is betting that the next great leap in AI won’t come from better chatbots, but from spatial intelligence AI. By developing “large geospatial models” (LGMs), the company aims to create a machine-readable map of the physical world that allows AI agents and robots to navigate and reason over real-world spaces with the same fluidity that an LLM reasons over text.
At the heart of this ambition is a provocative thesis from Niantic Spatial Executive Chairman John Hanke: the digital economy is only a fraction of the whole. According to Hanke, even as the world has spent decades digitizing information, it has largely ignored the physical infrastructure that sustains life.
Just 20% of the world economy is online but the 80% is not […] the acts of extracting, refining, growing, assembling, combining and shipping the atoms that warm us, shelter us, feed us, and generally make life possible for human beings.
Beyond the Digital Screen: Mapping the ‘Undigitized 80%’
The distinction Hanke makes is critical for the future of automation. In industries like agriculture, energy, and logistics, the “data” isn’t a PDF or a database entry; It’s the physical layout of a facility or the topography of a landscape. If an AI agent is tasked with managing a warehouse or directing an autonomous tractor, a 2D image or a text description is insufficient. The agent needs to realize exactly where it is, the distance to the nearest obstacle, and the geometric constraints of its environment.
What we have is where the concept of the Large Geospatial Model (LGM) comes in. Unlike standard mapping, which often relies on satellite imagery or simple GPS coordinates, an LGM integrates 3D scans, LIDAR, and high-fidelity GPS data to create a geometrically accurate representation of space. The goal is to move beyond “capturing” a space to “encoding” it in a way that a machine can actually use for operational decision-making.
To facilitate this, Niantic Spatial recently launched Scaniverse for businesses. The software is designed to be a “front door” for companies to begin digitizing their physical assets, turning smartphones and 360-degree cameras into high-precision mapping tools.
The Tech Stack: From Gaussian Splats to VPS 2.0
From a technical perspective, the challenge of spatial intelligence is twofold: you have to create a photorealistic model of the world, and you have to ensure a machine can locate itself within that model with centimeter-level precision.
To solve the first problem, Scaniverse utilizes Gaussian splatting. This is a point-based 3D rendering technique that allows for the creation of fully photorealistic digital scenes. Unlike traditional meshes, which can look blocky or artificial, Gaussian splatting captures the nuance of light and texture, making the digital twin almost indistinguishable from the physical original.
The second problem—positioning—is addressed via the Visual Positioning System (VPS) 2.0. While GPS is sufficient for finding a city block, it often fails in “GPS-degraded” environments, such as inside a steel-framed warehouse or deep within an industrial complex. VPS 2.0 provides near centimeter-level localization by comparing the machine’s current visual feed against the mapped environment, allowing for reliable 6DoF (six degrees of freedom) movement.
These tools allow businesses to map environments ranging from single rooms to vast industrial sites, including underwater facilities, without requiring expensive, proprietary hardware.
The Broader Shift Toward Operational AI
Niantic Spatial isn’t the only player recognizing the limitations of text-based AI. Dean Summers, founder and director of engineering at the geospatial AI firm Lampata, suggests that the industry is hitting a ceiling with language models. He argues that the “next real leap” will come from systems that understand the world dynamically and spatially.
According to Summers, foundational models built on Earth Observation (EO) and GIS (Geographic Information Systems) data allow machines to reason over assets and observe changes over time. This shifts AI from a conversational tool into an operational tool. The primary hurdle now is data integration—merging fundamentally different types of geospatial data, such as vectors, rasters, and point clouds, into a single, reliable stream of intelligence.
To help navigate this complexity, the industry is beginning to standardize how these models are built and deployed. A key part of this ecosystem is the development of unified toolkits that allow developers to build on top of these spatial maps regardless of their hardware platform.
Spatial Intelligence Deployment Capabilities
| Technology | Primary Function | Key Benefit | Typical Use Case |
|---|---|---|---|
| Gaussian Splatting | Photorealistic 3D Rendering | Visual fidelity | Digital twins, virtual tours |
| VPS 2.0 | Visual Positioning | Centimeter precision | Robot navigation, indoor GPS |
| LGM | Foundation Geospatial Model | Machine reasoning | City-scale AI planning |
| NSDK 4.0 | Unified Dev Kit | Cross-platform compatibility | App development (Unity, ROS 2) |
What Comes Next for the Physical Web
The transition from a “digital-first” AI to a “spatial-first” AI will likely be measured by how effectively these tools are adopted by the industries that make up that missing 80% of the economy. For a logistics company, this means AI agents that can optimize a warehouse layout in real-time; for construction, it means drones that can detect a structural deviation of a few centimeters against a 3D blueprint.
The next major milestone for Niantic Spatial is the scheduled release of the Niantic Spatial Development Kit (NSDK) 4.0 in April 2026. This unified SDK is expected to connect directly to Scaniverse and VPS 2.0, providing a standardized bridge for developers working across Unity, Swift, Android, and ROS 2 (the Robot Operating System).
As we move toward this date, the focus will likely shift from the novelty of 3D scanning to the practical utility of these maps. The goal is no longer just to see the world in 3D, but to give AI the eyes and the spatial reasoning required to finally step out of the screen and into the physical world.
Do you think spatial intelligence is the missing link for robotics, or is the “undigitized 80%” too complex to ever fully map? Let us know your thoughts in the comments.
