Gotta catch them all (for the military)

A version of this article originally appeared in Quartz’s members-only Weekend Brief newsletter. Quartz members get access to exclusive newsletters and more. Sign up here.

Within days of Pokémon Go launching ten years ago on July 7, millions of people were wandering parks and sidewalks, phones raised, hunting digital creatures overlaid on the physical world. Nintendo's stock doubled. Memes proliferated (remember Pokémon Go to the polls?). Pedestrians fell off cliffs playing the game.

Then, as suddenly as it arrived, the frenzy subsided. By mid-September 2016, the game had already lost roughly four-fifths of its American players. The media deemed Pokémon Go a flash in the pan, like so many apps that came before it.

But the dedicated users never gave up. Those Pokémon fans helped build out a dataset that has turned out to be exactly what the AI industry is starving for in 2026.

Gotta catch them all, then catch them again

Starting in 2021, players could opt in to scan real-world locations with their phone cameras, uploading tagged images of storefronts, public squares, parks and train stations in exchange for in-game perks. By 2024, Niantic claimed to be ingesting roughly a million new scans per week.

What they accumulated would have cost a fortune to build any other way. Satellite imagery looks straight down and misses what is happening at street level. Street-level photography captures a single moment from a moving car. But these billions of images were of specific locations shot from multiple angles, across different times of day, seasons and weather conditions, by people who were standing still and carefully capturing data.

Last May, Niantic sold its games, including Pokémon Go, to Scopely, a mobile publisher majority-owned by Saudi Arabia's Public Investment Fund, for $3.5 billion. Niantic kept the data and the AI models trained on it, spinning them out into a new company called Niantic Spatial.

That new company’s first public product is a visual positioning system that can pinpoint a location to within a few centimeters using only a handful of images of nearby buildings. The timing is auspicious. The AI world has grown obsessed with world models, the technology that some believe will be the next leap in machine intelligence. World models require systems that understand physical space, not just language.

The company announced a partnership with Coco Robotics, which operates roughly a thousand last-mile delivery robots across American and European cities, to help the machines navigate urban environments where GPS routinely fails. Dense city blocks bounce satellite signals off buildings, creating positioning errors of fifty meters or more. 

Niantic Spatial's model, trained on billions of images taken by Pokémon Go players at specific, well-documented locations over years, gets around those issues and can locate a robot precisely enough to ensure it stops at exactly the right doorstep.

What players built without knowing

A dataset that precise and that vast was never going to stay in the delivery robot business for long. And the American military has never had much trouble finding uses for technology that works.

Cute Pokémon are no different. In December, Niantic Spatial announced a partnership with Vantor, a company that builds spatial-detection software for drones, including military ones focused on navigation in GPS-denied environments like active conflict zones. 

The same capability that gets a robot to stop at exactly the right doorstep translates pretty easily to a drone locating exactly the right target. Vantor announced a U.S. Army contract worth up to $217 million earlier this year, which suggests the military finds the possibility fairly compelling.

There are other questions worth asking too. When Niantic sold the games to Scopely last year, that sale included the ongoing apparatus for collecting location and movement data from more than 100 million monthly players, now under the majority ownership of Saudi Arabia's Public Investment Fund. The kingdom has a well-documented history of using surveillance infrastructure against journalists, activists and dissidents.

It is a pattern that the last decade of consumer tech has made familiar, if not exactly comfortable. When people interact with free software, they are participating in a data-collection exercise whose downstream uses they cannot anticipate or control. The terms of service said the scans could be used to improve gameplay and mapping. They did not say anything about drones. If it can happen to Pokémon, it can happen to anything.