We plugged AI into Japan's 3D city model and X-rayed 1,022 buildings near Tokyo Station

PLATEAU, the 3D city model built by Japan's Ministry of Land, Infrastructure, Transport and Tourism (MLIT), now talks directly to AI.

Plenty of people in Japan know PLATEAU as the government's 3D map of its cities. Fewer know that you can now ask an AI, "Where's the building data for the Tokyo Station area?" and get the file back. From there, it's a short step to counting all 1,022 buildings in a roughly 1 km square and seeing how deep each one is projected to flood. It has come a lot further than most people realize.

The government has also developed technology that builds the 3D city itself from satellite imagery and generative AI, cutting the time to model 10 km² of city from about 3 to 6 months to about 1 to 2 hours (MLIT).

Our editorial team put PLATEAU through its paces with AI. Here's what we found, and how AI is starting to change the project.

PLATEAU started in fiscal 2020 (Japan's fiscal year runs April to March) with 3D models of 56 cities. By the end of fiscal 2025, it covered 329. The data is free, even for commercial use, as long as you credit the source (PLATEAU site policy).

In 2026, AI began to reshape how PLATEAU is both made and used. MLIT's "PLATEAU Vision 2026" materials, released in July, compared automated production from satellite data and generative AI with the old manual process.

How people use it is changing too. In December 2025, a gateway that lets AI search PLATEAU directly (the PLATEAU MCP server) opened to the public, no login required (Re:Earth Engineering). PLATEAU used to be a map for people to look at. Now it's turning into data that AI both reads and builds.

WHAT IS

PLATEAUby MLIT

Japan's national 3D city model, built by MLIT and free for anyone to use

Aiming to become a “digital public good”── PLATEAU Vision 2026

1. 2020Project launch3D models of 56 cities released as open data in year one
2. 2023Into game enginesOfficial tools for loading the data into Unity and Unreal Engine
3. 2025.12An AI gateway opensAn experimental server lets AI search the data directly
4. 2026.07PLATEAU Vision 2026329 cities covered. Next: automated production with satellites and AI
5. NOWToward 500 citiesThe target for the end of fiscal 2027, with a 100% update rate by fiscal 2032

PLATEAU starts from local governments' urban planning maps, adds building heights from aerial surveys, building uses and more, and assembles the city in 3D. Every building and road is a separate object carrying its shape plus details such as use, floor count and disaster projections (PLATEAU FAQ).

If an ordinary 3D map is a scale model of a city, PLATEAU is one where every building comes with a résumé. That opens the door to all sorts of uses. One event recreated Saitama Shintoshin, an urban center just north of Tokyo, inside Fortnite (PLATEAU). There's even an official MLIT tool that turns the data into a Minecraft world (GitHub).

We connected our AI to the PLATEAU MCP server, the AI gateway that opened in December 2025. It's experimental, and no login is needed.

Once connected, the AI gets 14 tools: search by region and dataset, find the file for each roughly 1 km grid square, pull up a single building, and search PLATEAU's specifications.

First question: how much is in there? Answer: 9,157 registered datasets. Next, we searched the specs for "flood depth," and it pulled up the chapter that defines the disaster-risk data. That used to mean an expert digging through thick manuals. Now AI looks it up for you.

Next we asked, "Where's the building data for the Tokyo Station area?" It returned a file location for each roughly 1 km grid square (a "third-order mesh" on Japanese maps). We downloaded the file (about 166 MB) and tallied every building in it.

Even in central Tokyo, nearly 80% of buildings came with a flood projection for heavy rain or storm surges. For 130 of them, the projection was 0.5 m or more (about knee height on an adult), and 26 of those reached 1 m or more.

And the deepest? Looking it up through the AI gateway, we landed on a six-story medical facility in Hatchobori 3-chome, Chuo Ward, about 1 km from Tokyo Station. If the Kanda River basin floods under the worst-case rainfall used in official projections, the site is projected to flood up to 4.1 m deep, nearly submerging the ground floor.

You'd never spot this building by staring at a map. It surfaced only because we ranked all 1,022 buildings by depth.

The same building data is visible in PLATEAU VIEW, the browser app. We switched to the selection tool and clicked a building on the Yaesu (east) side of Tokyo Station.

Up came GranTokyo South Tower. Use: business. Measured height: 209.5 m. Floors: 42 above ground, 2 below. Scroll down for the site's floor area ratio, 900% (total floor space allowed relative to lot size). It even shows that if the Kanda or Sumida River overflowed under worst-case rainfall, the site is projected to flood up to 0.32 m.

Details you once had to piece together from real estate documents and disaster maps now come up with a single click. That's a lot of legwork gone.

Type "江東区" (Koto Ward, eastern Tokyo) into the search box and more than 20 datasets appear: buildings, roads, land use, river floods, storm surges, evacuation facilities and more. One is an "Arakawa River time-series flood simulation."

It recreates, step by step, how water would spread if the levee on the Arakawa River's right bank (the right side looking downstream, 10 km from the mouth) broke under worst-case rainfall. Hit play and blue water pours into the city. (Left: about 8 hours in. Right: about 2 days and 9 hours in.)

Water that first covers only the north reaches southern Koto Ward in just over two days. A hazard map gives you one picture: "this area floods." Here, you can see the order it happens in and how long it takes. For thinking about how much time you'd have to evacuate, it's hard to beat.

The on-screen notes say plainly that the simulation is a guide based on government projections; a real flood may not hit the same places or spread the same way.

The biggest win was pulling data through the AI gateway. Ask where the Tokyo Station building data is, get the file, and you can sort flood risk for 1,022 buildings without ever opening the map.

I think we're watching the groundwork being laid for AI that understands real cities. PLATEAU has already been used in trials that locate a self-driving car by matching its camera images to the 3D model (PLATEAU), and in trials that plan routes for robots and drones (PLATEAU). As research on AI that predicts how the world behaves (so-called world models) advances, accurate, clearly licensed data on real cities keeps gaining value. Japan built that with public money, made it free, and has it ready for 329 cities. That's a big advantage.

We hit a few snags while tallying. AI takes the data at its word. The tallest building in our square (235.2 m) had its name field set to "Joto Elementary School," a ward-run public school. The school occupies the lower floors of Tokyo Midtown Yaesu, and its name had been attached to the whole tower (Wikipedia). Ask an AI for the tallest building and it could well say: an elementary school.

PLATEAU combines data from different origins, such as local government registries and surveys. Update timing varies by area, typically every 1 to 5 years (PLATEAU FAQ). Vision 2026 also plans to mix AI-built models of varying accuracy (Vision 2026 overview).

So when you query it with AI, have it also check when the data is from and how accurate it is. If a value looks off, go back to the source. With that one extra step, it's an exceptional dataset, and a free one.

A city that took 3 to 6 months to build now takes 1 to 2 hours with satellites and AI. MLIT has also developed technology that applies building exteriors automatically from smartphone photos, and plans to start involving residents in updates in fiscal 2026 (MLIT).

Stale data has been PLATEAU's biggest weakness. Now the job of keeping it current is shifting from specialist contractors to AI and citizens. AI builds it, AI reads it, citizens fix it. If that works, PLATEAU becomes a "latest version of the city" companies can use daily, for planning stores, choosing logistics hubs or drafting disaster-resilient business plans.

Start by opening a place you know in PLATEAU VIEW, such as your office or home if you're in Japan. If you use AI, connect it to the MCP server too. A familiar neighborhood looks different once you see it as data.