A Mercedes-Benz sedan equipped with Level 3 autonomous driving doesn't just "see" the road. It queries a constantly refreshed high-definition map from HERE Technologies, cross-referencing lane geometry, road topology, and regulatory boundaries dozens of times per second. NVIDIA's GR00T humanoid robot platform—now shipping to partners like Agility Robotics and Figure AI—requires similar grounding: terrain data, obstacle detection, building layouts. Without those coordinates, a warehouse robot becomes an expensive lawn ornament.
And when an insurance underwriter's AI agent evaluates flood risk for a Houston property? Ideally, it cites FEMA's National Flood Hazard Layer, complete with vintage and confidence interval. In practice, it often hallucinates.
The gap between those two realities is where Ansh Chokshi sees an opportunity—or perhaps an imperative. His San Francisco-based startup, Mireye, is part of Y Combinator's Summer 2026 cohort with a straightforward pitch: index every inch of the earth and make it as queryable as the web. Starting, for now, with just the United States and federal data sources.
It's an audacious claim in a market that's already crowded with geospatial players. But Chokshi argues the existing landscape misses the point. Language models trained on the open internet have no reliable way to answer "What's the wildfire risk at 34.0522°N, 118.2437°W?" They guess, or worse, they confabulate. Physical AI agents—machines moving beyond text and pixels into loading docks and construction sites—need authoritative, queryable context. And right now, that context is scattered across dozens of federal custodians, inconsistent formats, and irregular update schedules.
A Market Hungry for Coordinates
The numbers suggest Mireye isn't chasing a niche. Grand View Research pegs the geospatial analytics market at $117.2 billion in 2026, projecting it will reach $234 billion by 2033, though other analysts like Fortune Business Insights suggest it could climb even higher, to $309.84 billion by 2034. Location intelligence, a subset focused on extracting business insights from spatial data, reached $27.9 billion this year and is projected to hit $76.4 billion by 2033, according to Grand View Research. Those figures reflect surging demand from logistics, insurance, energy—and now robotics, where coordinates carry actual consequence.
Physical AI, once speculative, has become a Gartner-certified trend. The consultancy named it a top strategic technology for 2026. BCG's April report framed it as "reshaping robotics today," with 2050 scenarios placing embodied agents as structural economic drivers. McKinsey's May analysis of European labor markets estimated that AI agents—both software and physical—could account for 82 percent of automation value. Goldman Sachs, in a forecast originating in 2024 that remains widely cited, projects the humanoid robot market alone will hit $38 billion by 2035.
Yet the infrastructure lags. Three academic surveys published in early 2026—"From Perception to Action: Spatial AI Agents and World Models," "Agentic AI for Remote Sensing," and "Physical AI Agents: Integrating Cognitive Intelligence with Real-World Action"—converge on a common bottleneck: agents that perceive, reason, and act in physical space require world models and spatial grounding that obey geospatial constraints. Training data scraped from Reddit and Wikipedia doesn't help when a robot needs to know if a parcel sits in a FEMA flood zone or if a solar farm site overlays a protected wetland.
The Mess Behind the Maps
The gap isn't exactly new. Geospatial data has always been messy, fragmented, sometimes maddeningly inconsistent.
The U.S. Geological Survey maintains elevation through its 3D Elevation Program. NOAA oversees coastal flood hazards. The USDA publishes the Cropland Data Layer. FEMA updates the National Flood Hazard Layer on irregular, county-by-county schedules. The Department of Energy tracks transmission infrastructure. The Bureau of Land Management announced a modernization of its General Land Office records system in July 2026. Each dataset arrives in a different format—GeoTIFF, Shapefile, GeoJSON, NetCDF—with different projections, different vintages, different access patterns.
For developers building AI agents, this creates a nightmare. An agent evaluating whether to site a data center in rural Nevada should pull terrain slope, wildfire fuel models from LANDFIRE, solar irradiance from NREL's National Solar Radiation Database (refreshed July 1, 2026), transmission proximity from EIA datasets, and parcel boundaries from county records. And ideally, it should attach timestamps and source URLs to every field.
That's a tall order. But standardization efforts have accelerated. The SpatioTemporal Asset Catalog (STAC) became an OGC Community Standard in 2025, and the National Geospatial Advisory Committee labeled it "Highly AI-Ready" in March 2026. Cloud-Optimized GeoTIFF is now baseline across NASA, USGS, and commercial platforms. Overture Maps Foundation—the open mapping consortium backed by Meta, Microsoft, Amazon, and TomTom—ships monthly data releases. Its June 17, 2026 drop (version 2026-06-17.0, schema v1.17.0) improved building and point-of-interest coverage globally. In May, Microsoft announced it now powers Microsoft Maps with Overture data, citing faster release cadence and better POI density.
Then there's the Model Context Protocol (MCP), Anthropic's specification for letting language models invoke external tools. In the first half of 2026, Mapbox, CARTO, and Microsoft all released MCP servers that expose geospatial APIs to AI agents. Mapbox's MCP server wraps its Location AI components. CARTO's "CARTO for Agents," launched May 14, lets agents run spatial SQL queries, generate analyses, and publish maps within enterprise governance frameworks. Microsoft Planetary Computer Pro went generally available June 2, with MCP capabilities for AI agents to execute geospatial workflows on satellite imagery and environmental datasets.
These infrastructure pieces—open standards, cloud-native formats, agent-native protocols—converged in 2026 to make sourced, queryable geospatial data technically feasible at scale. Whether the market actually needs another layer on top of that stack is the open question.
Citations Included
Mireye's bet is that it does.
The company's product is an API (and MCP server) that turns a latitude-longitude pair into structured, sourced geospatial facts. Two endpoints anchor the service. /v1/ask uses an LLM planner with inline citations to synthesize answers: "Is this location in a flood zone? What's the wildfire risk?" /v1/fetch provides deterministic field retrieval with per-field provenance—source name, dataset vintage, fetched timestamp, confidence score. The MCP server mirrors both modes, so agents running in Claude Desktop, Cursor, or custom workflows can call Mireye tools without custom integration code.
Coverage is U.S.-only at launch, pulling exclusively from federal sources: USGS, NOAA, USDA, USFS, USFWS, EPA, EIA, FAA, FHWA, BTS, and the U.S. Census Bureau. According to the company's documentation, the API exposes 296 fields across seven layers—terrain, land cover, buildings, roads, water, risk, utilities—plus 15 preset queries. Every value returns not just the data but its lineage: which agency published it, when, and how confident the system is in the result.
Chokshi, who previously led AI and machine learning engineering at Seismic and studied computer science, math, and economics at the University of Toronto, frames the problem plainly on Mireye's Y Combinator profile. "Models guess on specific questions about specific places," he writes, because geospatial data is fragmented and unstructured. The goal is an index of the earth "as queryable as the web."
His prior startup, Poker Pit, reached $250,000 in annual recurring revenue before being acquired—a solid, if modest, exit. Mireye represents a bigger swing.
Early targeting focuses on underwriting, lending, energy siting, and general agent skills requiring terrain, risk, utilities, or parcel data. An insurance AI evaluating a homeowner policy can fetch FEMA flood zone designation, LANDFIRE wildfire fuel models, and USGS elevation—all with citations an auditor can verify. A solar developer's agent screening thousands of potential sites can pull NREL solar irradiance, EIA transmission proximity, USDA land-cover constraints, and local parcel boundaries, then surface only parcels that pass all filters.
The pricing model is "free for now" during early access. Commercial pricing will be announced after the pilot period ends. (Translation: they're still figuring it out.)
A Crowded Field, Divergent Approaches
Mireye isn't alone in chasing this space, though the competition comes at the problem from different angles.
Mapbox's MCP server provides geocoding, routing, and isochrone tools—closer to traditional location services than raw geospatial data layers. CARTO emphasizes enterprise governance: its "CARTO for Agents" product lets agents query proprietary datasets within a company's spatial data warehouse, respecting access controls and lineage rules. Microsoft Planetary Computer Pro focuses on Earth observation—satellite imagery, climate models, environmental catalogs accessible via MCP for scientific and impact-monitoring workflows.
What distinguishes Mireye, at least in theory, is the focus on federal authoritative sources with per-field provenance and the explicit U.S.-only, U.S.-government-data posture. That's both a strength and a constraint. Federal data is open, well-documented, and legally unencumbered. But international expansion would require navigating non-federal sources, commercial licenses, and varying data-quality regimes. For a two-person startup—Chokshi hasn't disclosed team size, but YC batches tend to start lean—that's a long road.
The Compliance Tailwind
Three forces will likely shape how this market evolves, assuming Mireye survives long enough to see them play out.
First: compliance and auditability requirements are tightening. The EU AI Act's general provisions became applicable August 2, 2026. Transparency requirements for AI-generated content kick in by December 2. Staged high-risk obligations roll out through 2027-2028. In the U.S., the FTC's proposed settlement with Kochava, filed May 4, 2026, bars selling or sharing sensitive location data without affirmative consent. State privacy laws already treat precise geolocation as sensitive personal information, and Washington's "My Health My Data Act" bans geofencing around health facilities.
Systems that surface per-field provenance and provide auditors with re-fetchable citations—exactly what Mireye's architecture delivers—may have an advantage in regulated industries. Or they may just be table stakes.
Second: the open geospatial infrastructure will continue to mature. The Open Geospatial Consortium opened public comment on 3D Tiles 2.0 on June 30, 2026, targeting efficient streaming of large, semantically rich 3D geodata—critical for robotics navigation and augmented reality applications. Overture Maps ships monthly releases, and adoption is accelerating. Microsoft Maps, TomTom's Orbis lane-level maps for autonomous vehicles, and research projects like OpenEarthAgent (a unified framework for tool-augmented geospatial agents published in February 2026) all build on these foundations.
The trajectory is toward a world where open base maps, algorithmically refreshed, coexist with proprietary high-definition maps for specialized domains like autonomous driving. Whether there's room for a third category—authoritative federal data layers, normalized and citation-ready—remains to be seen.
Third, and perhaps most importantly: physical AI deployments are no longer speculative. NVIDIA announced in March 2026 that its Isaac robotics platform, GR00T humanoid models, and Cosmos world models now integrate with partners including ABB, Agility Robotics, Figure AI, and Universal Robots. Mercedes-Benz's Level 3 DRIVE PILOT system uses HERE HD Live Map in production vehicles today. Volkswagen Group's CARIAD selected TomTom Orbis Maps for its autonomous driving systems in January 2026.
These aren't prototypes. They're shipping products that need real-time, accurate spatial context. And they need it now, not after a lengthy procurement process.
The Next Industrial Revolution (Assuming the Data Holds Up)
The next phase won't just be about collecting more data. It'll be about making existing authoritative data agent-accessible with verifiable provenance.
Take energy siting. Lawrence Berkeley National Laboratory's "Queued Up: 2026 Edition," published in May, found that interconnection queues for new generation projects remain backlogged despite some easing in 2025. An agent that can screen thousands of potential sites against solar resource data (NREL NSRDB), grid proximity (EIA), land-use constraints (USDA Cropland Data Layer, USFWS protected areas), and local zoning (county parcel data) in seconds—and cite every constraint it surfaces—could accelerate a process that currently takes months of manual GIS work.
Could. The word carries weight.
Because the question for founders building on this layer isn't just technical. It's epistemological. Do you trust a model's parametric memory to know what's at 34.0522°N, 118.2437°W—a seemingly arbitrary coordinate in Los Angeles—or do you query an index that returns USGS elevation, FEMA flood zone, LANDFIRE fuel model, and EPA air quality monitoring station proximity with source URLs and vintages attached?
For applications where wrong answers carry financial or safety consequences, the answer seems clear. An insurance underwriter can't shrug off a miscalculated flood risk as a hallucination. A solar developer can't afford to break ground on a site that violates wetland protections.
The infrastructure to make that query as natural as a web search arrived, more or less, in 2026. What gets built on top of it will define the next industrial revolution—assuming, of course, that the data layer underneath is as reliable as the agents are ambitious. And assuming Mireye, or someone like it, can build a business around something the federal government already publishes for free.
That's the bet Chokshi is making. Whether the market rewards him for it is another question entirely.
