Vexcel opened Vexcel Intelligence to customers worldwide on Tuesday. You type a question about the physical world in plain language, and it returns matching aerial images and their locations, ranked by relevance. Behind it sit more than 3 billion vector embeddings built from 15.1 million km² of the company’s aerial photography across 47 countries. Customers can search in Vexcel’s own Studio app or wire the library into their own AI agents through APIs and an MCP server.
That’s about a tenth of the planet’s land surface, shot with Vexcel’s own UltraCam cameras from aircraft instead of satellites: 7.5 cm per pixel in cities and 15 cm across wider areas. Vexcel calls its imagery 16 times the resolution of the best commercial satellite pictures over the US. That figure works if you count pixels per square metre against a 30 cm satellite; measured along a line, 7.5 cm is four times sharper. Averaged across the library, there’s one embedding for roughly every half hectare of ground.
It took most of the year to get here. Vexcel showed a handful of searchable cities in January and opened a preview in February. In July it launched an MCP server, the connector that lets assistants such as ChatGPT, Claude, Copilot and Gemini query its imagery and property data. August brought Custom Elements: name an object (propane tanks, putting greens, telephone poles, docks) and the system finds it, with no model training on the customer’s side. This week the search went live across the whole library.
The mechanics are simple. Each patch of imagery is converted into a list of numbers that describes what’s in it, and a typed question goes through the same conversion; the search returns the patches whose numbers sit closest to the question’s. Nobody has to build a “cooling tower detector” first. In cities each place is searchable from five views, straight down plus 45-degree obliques from the four compass directions, so the search can see facades and the sides of structures that a top-down photo misses. Rural areas get the top-down view only.
Vexcel’s own sample queries show who it’s built for. Insurers get “find pools with safety fences,” road departments “find frost heaves on roadways,” utilities “find trees growing near overhead power lines.” The one an OSINT analyst should notice is pitched at real estate: “find data centers with large cooling systems and backup generators.” Hyperscalers often build under project code names and shell LLCs and say little until a site is running. The cooling plant and the generator yard show up from the air anyway, and a question typed in one sentence now returns a ranked list of candidate sites to check by eye.
The oblique views matter for verification too. Placing a video frame on a map usually comes down to matching facades, and satellites mostly show roofs. In principle, a description of what’s in the frame can now be run as a search, as long as the frame was shot somewhere Vexcel flies.
The Map Ends Where the Planes Can’t Fly
That condition is the whole story for anyone who watches conflicts. Aerial survey needs a plane overhead, and a plane needs permission to use the airspace. Vexcel’s fleet covers the US, the UK, Canada, Australia, New Zealand, Japan, India and Western Europe. Brazil, South Africa, Poland and the three Baltic states were added in 2024, and Mexico this year, starting with 40 cities that hold about half its population. Of the more than 6.6 million km² on this year’s flight plan, 5.2 million are in the US, about a million in Europe, and roughly half a million everywhere else combined.
So more than three-quarters of this year’s flying is over the US. What the product offers is a very sharp picture of countries that let survey aircraft in. None of the places that dominate OSINT work right now can be flown: there’s no survey plane over Ukraine’s front line, Gaza, Sudan or Iran’s side of the Strait of Hormuz. The newest European additions, Poland and the Baltic states, sit on NATO’s eastern flank. The front line itself is off the map.
Time is the other limit. A survey is a snapshot from the last time a plane went over. Vexcel’s biggest US cities get up to three collections a year, which is plenty for asking what changed since spring and no use at all for what happened this morning. For that you still need satellites, which can revisit daily and go where no plane can.
The trade is sharpness. Semantic search over satellite pixels already exists: Google publishes annual embeddings for all of the planet’s land and coastal waters in Earth Engine. Its squares are 10 metres across, about the footprint of a small house, where Vexcel’s pixel is 7.5 cm. At 10 metres you can tell a forest from a car park. You can’t find a propane tank.
Who Gets to Ask
Access is the second filter. You sign in with a Vexcel API key, you can query only the products your licence covers, and newcomers start with a demo and a sandbox account. There’s no public price list. The industries Vexcel sells to are insurance, government, utilities, telecom, energy, finance and construction. Most volunteer OSINT researchers won’t get a seat.
That gate carries more weight than it looks. A query that finds trees brushing power lines also finds the substations, and an agent connected through MCP will run the same question across a whole state without getting bored. The control Vexcel describes is the licence: who gets a key, and what it unlocks. The privacy question is closer to home. “Pools with safety fences” is an underwriting question, and American homeowners have spent the past few years complaining that insurers dropped them over aerial photos, some of them out of date. Plain-language search makes that kind of check cheaper to run, on every house at once.
Inside those 47 countries, the physical world can now be searched by sentence. For the places that make the news, it’s still satellites.